electromyogram signal enhancement and upper-limb...

Sara

Ab

ba

spo

ur ELEC

TRO

MYO

GR

AM

SIGN

AL EN

HA

NC

EMEN

T AN

D U

PP

ER-LIM

B MYO

ELECTR

IC PA

TTERN

REC

OG

NITIO

N2019

ISBN 978-91-7485-418-3ISSN 1651-4238

Address: P.O. Box 883, SE-721 23 Västerås. SwedenAddress: P.O. Box 325, SE-631 05 Eskilstuna. SwedenE-mail: [email protected] Web: www.mdh.se

Mälardalen University Doctoral Dissertation 283

Electromyogram Signal Enhancement and Upper-Limb Myoelectric Pattern RecognitionSara Abbaspour

Sara Abbaspour received the MSc degree in biomedical engineering from Amirkabir University of Technology, Iran in 2011. Thereafter, as a signal processor and Java programmer, she was part of a group to design “Computerized Screening ofChildren Congenital Heart Disease”. She has been a Doctoral Candidate at Mälardalen University, Sweden since 2014. Her research interest include biomedical signal processing in thearea of filter designing and pattern recognition

Losing a limb causes difficulties in our daily life. To regain the ability to live an in-dependent life, artificial limbs have been developed. Hand prostheses belong toa group of artificial limbs that can be controlled by the user through the activity ofthe remnant muscles above the amputation. Electromyogram (EMG) is one of the sources that can be used for control methods for hand prostheses. Surface EMGs are powerful, non-invasive tools that provide information about neuromuscular activity of the subjected muscle, which has been essential to its use as a source of control for prosthetic limbs. However, the complexity of this signal introduces a big challenge to its applications. EMG pattern recognition to decode different limb movements is an important advancement regarding the control of powered prostheses. It has the potential to enable the control of powered prostheses using the generated EMG by muscular contractions as an input. However, its use has yet to be transitioned into wide clinical use. Different algorithms have been developed in state of the art to decode different movements; however, the challenge still lies in different stages of a successful hand gesture recognition and improvements in these areas could potentially increase the functionality of powered prostheses. This thesis firstly focuses on improving the EMG signal’s quality by proposing novel and advanced filtering techniques. Four efficient approaches (adaptive neuro-fuzzy inference system-wavelet, artificial neural network-wavelet, adaptive subtraction and automated independent component analysis-wavelet) are proposed to im-prove the filtering process of surface EMG signals and effectively eliminate ECGinterferences. Then, the offline performance of different EMG-based recognition algorithms for classifying different hand movements are evaluated with the aim of obtaining new myoelectric control configurations that improves the recognition stage. Afterwards, to gain proper insight on the implementation of myoelectricpattern recognition, a wide range of myoelectric pattern recognition algorithms are investigated in real time on 15 healthy volunteers.

Sammanfattning

Att förlora en extremitet orsakar svårigheter i vår vardag. För att återfå förmågan till ett självständigt liv har artificiella händer och ben utvecklats. Handproteser kan kontrolleras av användaren genom aktiviteten hos återstående muskler ovanför amputationen. Elektromyogram (EMG) är en av de källor som kan användas till kontrollmetoder för handproteser. Yt-EMG är kraftfulla icke-invasiva verktyg som ger information om neuromuskulär aktivitet hos en specifik muskel, vilket är avgörande för dess användning att styra proteser. Komplexiteten hos signalen utgör dock en stor utmaning. EMG-mönsterigenkänning för att avkoda olika handrörelser är ett viktigt framsteg när det gäller kontroll av motoriserade proteser. Denna metod har potential att möjliggöra styrning av proteser genom att använda EMG-signalerna från muskelkontraktioner som insignal. Denna metod har dock ännu inte fått någon stor klinisk spridning. Olika algoritmer har utvecklats inom området för att avkoda olika rörelser; men utmaningen att identifiera olika handrörelser i olika faser kvarstår, och förbättringar inom dessa områden kan komma att öka funktionaliteten hos motoriserade proteser. Denna avhandling undersöker flera aspekter kring detta, först hur kvaliteten hos EMG-signaler kan förbättras genom att nya och avancerade filtreringstekniker. Fyra effektiva tillvägagångssätt (adaptivt neuro-fuzzy inference system-wavelet, artificiellt neuralt nätverk-wavelet, adaptiv subtraktion och automatiserad oberoende komponentanalys-wavelet) presenteras för att förbättra filtreringsprocessen för yt-EMG-signaler och effektivt eliminera EKG-störningar. Även offline-prestanda för olika EMG-baserade igenkänningsalgoritmer undersöks, däribland förmågan att klassificera olika handrörelser med sikte på att erhålla nya myoelektriska kontrollkonfigurationer som förbättrar igenkänningen. För att undersöka hur väl de myoelektriska mönsterigenkänningssalgoritmerna fungerar i verkliga situationer, har ett brett spektrum av myoelektriska algoritmer undersökts i realtid. 15 friska frivilliga försökspersoner har använt systemet och resultaten tyder på att linjär diskriminantanalys (LDA) och maximum likelihood-metoden (ML-metoden) är bättre än de andra klassificeringsmetoderna. Realtidsundersökningen

Sammanfattning

Att förlora en extremitet orsakar svårigheter i vår vardag. För att återfå förmågan till ett självständigt liv har artificiella händer och ben utvecklats. Handproteser kan kontrolleras av användaren genom aktiviteten hos återstående muskler ovanför amputationen. Elektromyogram (EMG) är en av de källor som kan användas till kontrollmetoder för handproteser. Yt-EMG är kraftfulla icke-invasiva verktyg som ger information om neuromuskulär aktivitet hos en specifik muskel, vilket är avgörande för dess användning att styra proteser. Komplexiteten hos signalen utgör dock en stor utmaning. EMG-mönsterigenkänning för att avkoda olika handrörelser är ett viktigt framsteg när det gäller kontroll av motoriserade proteser. Denna metod har potential att möjliggöra styrning av proteser genom att använda EMG-signalerna från muskelkontraktioner som insignal. Denna metod har dock ännu inte fått någon stor klinisk spridning. Olika algoritmer har utvecklats inom området för att avkoda olika rörelser; men utmaningen att identifiera olika handrörelser i olika faser kvarstår, och förbättringar inom dessa områden kan komma att öka funktionaliteten hos motoriserade proteser. Denna avhandling undersöker flera aspekter kring detta, först hur kvaliteten hos EMG-signaler kan förbättras genom att nya och avancerade filtreringstekniker. Fyra effektiva tillvägagångssätt (adaptivt neuro-fuzzy inference system-wavelet, artificiellt neuralt nätverk-wavelet, adaptiv subtraktion och automatiserad oberoende komponentanalys-wavelet) presenteras för att förbättra filtreringsprocessen för yt-EMG-signaler och effektivt eliminera EKG-störningar. Även offline-prestanda för olika EMG-baserade igenkänningsalgoritmer undersöks, däribland förmågan att klassificera olika handrörelser med sikte på att erhålla nya myoelektriska kontrollkonfigurationer som förbättrar igenkänningen. För att undersöka hur väl de myoelektriska mönsterigenkänningssalgoritmerna fungerar i verkliga situationer, har ett brett spektrum av myoelektriska algoritmer undersökts i realtid. 15 friska frivilliga försökspersoner har använt systemet och resultaten tyder på att linjär diskriminantanalys (LDA) och maximum likelihood-metoden (ML-metoden) är bättre än de andra klassificeringsmetoderna. Realtidsundersökningen

ii

visar också att förutom LDA och ML-metoden, så är algoritmerna med flerlagersperception bättre än de övriga algoritmerna då de jämförs med avseende på klassificeringsnoggrannhet och beräkningshastighet.

ii

visar också att förutom LDA och ML-metoden, så är algoritmerna med flerlagersperception bättre än de övriga algoritmerna då de jämförs med avseende på klassificeringsnoggrannhet och beräkningshastighet.

Abstract

The electromyogram (EMG) provides information about neuromuscular activity in a non-invasive way. Therefore, it has been considered as an important and effective control input for hand prostheses. EMG pattern recognition to decode different limb movements is an important advance for the control of powered prostheses. However, its use has yet to be transitioned into wide clinical use. Different algorithms have been implemented to enhance the functionality and usability of prosthetic hands. However, the challenge still lies in the preprocessing, feature extraction and classification phases, which are important stages of successful hand gesture recognition using surface EMG signals. Various artifacts originated from different sources inevitably contaminate EMG signals; therefore, the preprocessing stage is required to remove unwanted interferences. The electrical activity of the heart muscles is one of the artifacts that interfere with the EMG signals recorded from the upper trunk muscles, thereby making EMG signals unreliable. Different methods have been proposed to minimize the extent to which electrocardiogram (ECG) interferes with surface EMG signals; however, in spite of the numerous attempts to eliminate or reduce this artifact, the problem of the accurate and effective denoising of the EMG remains a challenge. To overcome this challenge, four efficient approaches (combined adaptive neuro-fuzzy inference system and wavelet, combined artificial neural network and wavelet, adaptive subtraction and automated independent component analysis-wavelet) were proposed to improve the filtering process of surface EMG signals and effectively eliminate ECG interferences. Then, this thesis evaluated the offline performance of different EMG-based recognition algorithms for classifying individual hand movements. Various combinations of 44 features and six classifiers were investigated, an efficient feature set (FS) was obtained, and new myoelectric configurations were proposed to improve the motion recognition accuracy. Overall, an FS/maximum likelihood estimation (MLE) combination was found to be the most suitable for higher recognition accuracy, and the mean absolute value/k-nearest neighbor combination for lower processing time.

Abstract

The electromyogram (EMG) provides information about neuromuscular activity in a non-invasive way. Therefore, it has been considered as an important and effective control input for hand prostheses. EMG pattern recognition to decode different limb movements is an important advance for the control of powered prostheses. However, its use has yet to be transitioned into wide clinical use. Different algorithms have been implemented to enhance the functionality and usability of prosthetic hands. However, the challenge still lies in the preprocessing, feature extraction and classification phases, which are important stages of successful hand gesture recognition using surface EMG signals. Various artifacts originated from different sources inevitably contaminate EMG signals; therefore, the preprocessing stage is required to remove unwanted interferences. The electrical activity of the heart muscles is one of the artifacts that interfere with the EMG signals recorded from the upper trunk muscles, thereby making EMG signals unreliable. Different methods have been proposed to minimize the extent to which electrocardiogram (ECG) interferes with surface EMG signals; however, in spite of the numerous attempts to eliminate or reduce this artifact, the problem of the accurate and effective denoising of the EMG remains a challenge. To overcome this challenge, four efficient approaches (combined adaptive neuro-fuzzy inference system and wavelet, combined artificial neural network and wavelet, adaptive subtraction and automated independent component analysis-wavelet) were proposed to improve the filtering process of surface EMG signals and effectively eliminate ECG interferences. Then, this thesis evaluated the offline performance of different EMG-based recognition algorithms for classifying individual hand movements. Various combinations of 44 features and six classifiers were investigated, an efficient feature set (FS) was obtained, and new myoelectric configurations were proposed to improve the motion recognition accuracy. Overall, an FS/maximum likelihood estimation (MLE) combination was found to be the most suitable for higher recognition accuracy, and the mean absolute value/k-nearest neighbor combination for lower processing time.

iv

Then, to gain proper insight into the clinical implementation of myoelectric pattern recognition, this thesis investigated the offline and real-time performance of nine classification algorithms to decode ten individual hand and wrist movements. The results on 15 healthy volunteers suggested that linear discriminant analysis (LDA) and MLE significantly (p<0.05) outperformed the other classifiers. The real-time investigation illustrated that in addition to the LDA and MLE, multilayer perceptron also outperformed the other algorithms in the classification accuracy and completion rate.

iv

Then, to gain proper insight into the clinical implementation of myoelectric pattern recognition, this thesis investigated the offline and real-time performance of nine classification algorithms to decode ten individual hand and wrist movements. The results on 15 healthy volunteers suggested that linear discriminant analysis (LDA) and MLE significantly (p<0.05) outperformed the other classifiers. The real-time investigation illustrated that in addition to the LDA and MLE, multilayer perceptron also outperformed the other algorithms in the classification accuracy and completion rate.

To my wonderful family, particularly to my understanding and patient husband, Mehdi, who has put up with these

many years of research.

To my wonderful family, particularly to my understanding and patient husband, Mehdi, who has put up with these

many years of research.

Acknowledgments

This thesis would not have been possible without the inspiration and support of a number of wonderful individuals - my thanks and appreciation to all of them for being part of this journey and making this thesis possible.

Foremost, I would like to express my sincere gratitude to my supervisors, Professor Maria Lindén, Associate Professor Hamid GholamHosseini, Dr. Maria Ehn, Associate Professor Shahina Begum and Dr. Giacomo Spampinato. I am grateful to them for providing valuable suggestions, comments and feedback throughout my studies. In particular, I thank my main supervisor Professor Maria Lindén for her continuous support of my study and research. Her guidance helped me in all of the time of the research and writing of this thesis. I could not have imagined having a better supervisor and mentor for my study.

I am grateful to Associate Professor Max Ortiz Catalan and Adam Naber for all the discussions and their excellent feedback during our joint work.

I would like to thank all the participants who gave generously of their time and electromyogram data to this thesis. I further thank Daniel Morberg for his grateful help in the modification of the BioPatRec software. I am thankful to Annika Havbrandt and Daniel Flemström for their help in providing me with a nice and quiet place for performing the experimental tests. I would like to acknowledge my friend Zeinab and her husband Komeil which I appreciate their help in preparing the cover photo.

I thank all the lecturers and professors who I learned a lot from during meetings, lectures, seminars and PhD courses.

My sincere thanks also go to my friends and colleagues in the IDT department for their friendship and support and for creating a cordial working environment. I would like to thank the IDT administration staff, in particular, Carola Ryttersson, for their help with practical issues.

Acknowledgments

This thesis would not have been possible without the inspiration and support of a number of wonderful individuals - my thanks and appreciation to all of them for being part of this journey and making this thesis possible.

Foremost, I would like to express my sincere gratitude to my supervisors, Professor Maria Lindén, Associate Professor Hamid GholamHosseini, Dr. Maria Ehn, Associate Professor Shahina Begum and Dr. Giacomo Spampinato. I am grateful to them for providing valuable suggestions, comments and feedback throughout my studies. In particular, I thank my main supervisor Professor Maria Lindén for her continuous support of my study and research. Her guidance helped me in all of the time of the research and writing of this thesis. I could not have imagined having a better supervisor and mentor for my study.

I am grateful to Associate Professor Max Ortiz Catalan and Adam Naber for all the discussions and their excellent feedback during our joint work.

I would like to thank all the participants who gave generously of their time and electromyogram data to this thesis. I further thank Daniel Morberg for his grateful help in the modification of the BioPatRec software. I am thankful to Annika Havbrandt and Daniel Flemström for their help in providing me with a nice and quiet place for performing the experimental tests. I would like to acknowledge my friend Zeinab and her husband Komeil which I appreciate their help in preparing the cover photo.

I thank all the lecturers and professors who I learned a lot from during meetings, lectures, seminars and PhD courses.

My sincere thanks also go to my friends and colleagues in the IDT department for their friendship and support and for creating a cordial working environment. I would like to thank the IDT administration staff, in particular, Carola Ryttersson, for their help with practical issues.

viii

It is a pleasure to thank my friends for the wonderful times we shared, specially the Friday night dinners. You gave me the necessary distractions from my research and made my time memorable.

Finally, I would like to express my deep and sincere gratitude to my family for their continuous and unparalleled love, help and support; my parents for supporting me spiritually throughout my life; my sister Sima for always being there for me as a friend; and my brothers Pouya and Sina who offered invaluable support over the years. I would like to acknowledge my husband and best friend, Mehdi. A big thank must go to him for his love, support and encouragement. Without you, it would not have been done! Maya, my dearest daughter, you have made me stronger, better and more fulfilled than I could have ever imagined. I love you to the moon and back.

The work was financed by the Knowledge Foundation’s research profile Embedded Sensor System for Health (ESS- H).

Sara Abbaspour February 2019

Västerås, Sweden

viii

It is a pleasure to thank my friends for the wonderful times we shared, specially the Friday night dinners. You gave me the necessary distractions from my research and made my time memorable.

Finally, I would like to express my deep and sincere gratitude to my family for their continuous and unparalleled love, help and support; my parents for supporting me spiritually throughout my life; my sister Sima for always being there for me as a friend; and my brothers Pouya and Sina who offered invaluable support over the years. I would like to acknowledge my husband and best friend, Mehdi. A big thank must go to him for his love, support and encouragement. Without you, it would not have been done! Maya, my dearest daughter, you have made me stronger, better and more fulfilled than I could have ever imagined. I love you to the moon and back.

The work was financed by the Knowledge Foundation’s research profile Embedded Sensor System for Health (ESS- H).

Sara Abbaspour February 2019

Västerås, Sweden

List of Publications

Papers included in the doctoral thesis1

Paper A: A Novel Approach for Removing ECG Interferences from Surface EMG signals Using a Combined ANFIS and Wavelet, Sara Abbaspour, Ali Fallah, Maria Lindén and Hamid GholamHosseini. Journal of Electromyography and Kinesiology, 2015.

Paper B: A Combination Method for Electrocardiogram Rejection from Surface Electromyogram. Sara Abbaspour, Ali Fallah, the Open Biomedical Engineering Journal, 2014.

Paper C: Removing ECG Artifact from Surface EMG Signal Using Adaptive Subtraction Technique. Sara Abbaspour, Ali Fallah, Journal of Biomedical Physics and Engineering, 2014.

Paper D: ECG Artifact Removal from Surface EMG Signal Using an Automated Method Based on Wavelet-ICA. Sara Abbaspour, Maria Lindén, Hamid GholamHosseini, 12th International Conference on Wearable Micro and Nano Technologies for Personalized Health, 2015.

Paper E: Evaluation of surface EMG-based recognition algorithms for decoding hand movements. Sara Abbaspour, Maria Lindén, Hamid GholamHosseini and Max Ortiz-Catalan, Conditionally accepted to the Medical & Biological Engineering & Computing Journal, 2018.

The included articles have been reformatted to comply with the doctoral thesis layout. The original paper C, available on the web, has a correction “Removing ECG Artifact from the Surface EMG Signal Using Adaptive Subtraction Technique, published in J Biomed Phys Eng 2014; 4 (1): 31-38” where Figure 10.6 has been updated.

List of Publications

Papers included in the doctoral thesis1

Paper A: A Novel Approach for Removing ECG Interferences from Surface EMG signals Using a Combined ANFIS and Wavelet, Sara Abbaspour, Ali Fallah, Maria Lindén and Hamid GholamHosseini. Journal of Electromyography and Kinesiology, 2015.

Paper B: A Combination Method for Electrocardiogram Rejection from Surface Electromyogram. Sara Abbaspour, Ali Fallah, the Open Biomedical Engineering Journal, 2014.

Paper C: Removing ECG Artifact from Surface EMG Signal Using Adaptive Subtraction Technique. Sara Abbaspour, Ali Fallah, Journal of Biomedical Physics and Engineering, 2014.

Paper D: ECG Artifact Removal from Surface EMG Signal Using an Automated Method Based on Wavelet-ICA. Sara Abbaspour, Maria Lindén, Hamid GholamHosseini, 12th International Conference on Wearable Micro and Nano Technologies for Personalized Health, 2015.

Paper E: Evaluation of surface EMG-based recognition algorithms for decoding hand movements. Sara Abbaspour, Maria Lindén, Hamid GholamHosseini and Max Ortiz-Catalan, Conditionally accepted to the Medical & Biological Engineering & Computing Journal, 2018.

The included articles have been reformatted to comply with the doctoral thesis layout. The original paper C,available on the web, has a correction “Removing ECG Artifact from the Surface EMG Signal Using Adaptive Subtraction Technique, published in J Biomed Phys Eng 2014; 4 (1): 31-38” where Figure 10.6 has been updated.

x

Paper F: Real-time and offline evaluation of myoelectric pattern recognition for upper limb prosthesis control. Sara Abbaspour, Adam Naber, Max Ortiz-Catalan, Hamid GholamHosseini and Maria Lindén, in submission to a journal, 2018.

The contributions made by Sara Abbaspour to papers A, B, C, D, E and F; 1 = Main responsibility, 2 = Contributed to a high extent.

Steps in scientific research A B C D E F

Idea and study formulation 1 1 2 1 1 1

Experimental design 1 2 1 1 -* 1

Performance of experiments 1 2 1 1 -* 1

Programming 1 1 1 1 1 1

Data analysis 1 1 1 1 1 1

Writing of manuscript 1 1 1 1 1 1

*In paper E, the data were obtained from a database; therefore, Sara Abbaspour did not contribute to the experimental design or performance of the experiments.

x

Paper F: Real-time and offline evaluation of myoelectric pattern recognition for upper limb prosthesis control. Sara Abbaspour, Adam Naber, Max Ortiz-Catalan, Hamid GholamHosseini and Maria Lindén, in submission to a journal, 2018.

The contributions made by Sara Abbaspour to papers A, B, C, D, E and F; 1 = Main responsibility, 2 = Contributed to a high extent.

Steps in scientific research A B C D E F

Idea and study formulation 1 1 2 1 1 1

Experimental design 1 2 1 1 -* 1

Performance of experiments 1 2 1 1 -* 1

Programming 1 1 1 1 1 1

Data analysis 1 1 1 1 1 1

Writing of manuscript 1 1 1 1 1 1

*In paper E, the data were obtained from a database; therefore, Sara Abbaspour didnot contribute to the experimental design or performance of the experiments.

Additional peer-reviewed publications not included in the doctoral thesis

1. Evaluation of Wavelet Based Methods in Removing Motion Artifact from ECG Signal, Sara Abbaspour, Hamid GholamHosseini and Maria Lindén, 16th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics, Gothenburg, 2015.

2. A Comparison of Adaptive Filter and Artificial Neural Network Results in Removing Electrocardiogram Contamination from Surface EMG, Sara Abbaspour, Ali Fallah, Ali Maleki, IEEE Proceeding, 20th National Conference on Electrical Engineering (ICEE), Tehran, 2012.

3. A Comparison of Adaptive Neuro-fuzzy Inference System and Real-time Filtering in Cancellation of ECG Artifact from Surface EMG, Sara Abbaspour, Ali Fallah, Ali Maleki, IEEE Proceeding, 20th National Conference on Electrical Engineering (ICEE), Tehran, 2012.

Additional peer-reviewed publications not included in the doctoral thesis

1. Evaluation of Wavelet Based Methods in Removing Motion Artifact from ECGSignal, Sara Abbaspour, Hamid GholamHosseini and Maria Lindén, 16th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics, Gothenburg,2015.

2. A Comparison of Adaptive Filter and Artificial Neural Network Results inRemoving Electrocardiogram Contamination from Surface EMG, SaraAbbaspour, Ali Fallah, Ali Maleki, IEEE Proceeding, 20th National Conferenceon Electrical Engineering (ICEE), Tehran, 2012.

3. A Comparison of Adaptive Neuro-fuzzy Inference System and Real-time Filteringin Cancellation of ECG Artifact from Surface EMG, Sara Abbaspour, Ali Fallah,Ali Maleki, IEEE Proceeding, 20th National Conference on ElectricalEngineering (ICEE), Tehran, 2012.

List of Abbreviations

ANFIS Adaptive Neuro-Fuzzy Inference System ANC Adaptive Noise Canceller ANN Artificial Neural Network AR Autoregressive ARV Averaged Rectified Value BPN Back Propagation Network CC Correlation Coefficient DAMV Difference Absolute Mean Value DASDV Difference Absolute Standard Deviation Value DC Direct Current ET Elapsed Time ECG Electrocardiogram EMG Electromyogram EMD Empirical Mode Decomposition FIR Finite Impulse Response FD Frequency Domain FE Frequency Energy HPF High-Pass Filter ICA Independent Component Analysis IAV Integrated Absolute Value KNN K-Nearest Neighborhood LDA Linear Discriminant Analysis LogRMS Logarithmic Root Mean Square MFL Maximum Fractal Length MLE Maximum Likelihood Estimation MAV Mean Absolute Value MPV Mean of Peek Value MAVS Mean Absolute Value Slope

List of Abbreviations

ANFIS Adaptive Neuro-Fuzzy Inference System ANC Adaptive Noise Canceller ANN Artificial Neural Network AR Autoregressive ARV Averaged Rectified Value BPN Back Propagation Network CC Correlation Coefficient DAMV Difference Absolute Mean Value DASDV Difference Absolute Standard Deviation Value DC Direct Current ET Elapsed Time ECG Electrocardiogram EMG Electromyogram EMD Empirical Mode Decomposition FIR Finite Impulse Response FD Frequency Domain FE Frequency Energy HPF High-Pass Filter ICA Independent Component Analysis IAV Integrated Absolute Value KNN K-Nearest NeighborhoodLDA Linear Discriminant AnalysisLogRMS Logarithmic Root Mean SquareMFL Maximum Fractal LengthMLE Maximum Likelihood EstimationMAV Mean Absolute ValueMPV Mean of Peek ValueMAVS Mean Absolute Value Slope

xiv

ms Milliseconds MLP Multilayer Perceptron NMF Non-Negative Matrix Factorization NLE Normalized Logarithmic Energy Perc Percentile PSD Power Spectrum Density PCA Principal Component Analysis FS Proposed Feature Set RLS Recursive Least Squares RegTree Regression Tree RFN Regulatory Feedback Networks RE Relative ErrorR Gs Research Goals RQs Research Questions RMS Root Mean Square SNR Signal-to-Noise Ratio SSC Slope Sign Changes SD Standard Deviation SSOM Supervised Self-Organized Map SVM Support Vector Machine TD Time Domain TDAR Time Domain and Autoregressive TDF Time-Frequency Domain WL Waveform Length WP Wavelet Packet WT Wavelet Transforms ZC Zero Crossings

xiv

ms Milliseconds MLP Multilayer Perceptron NMF Non-Negative Matrix Factorization NLE Normalized Logarithmic Energy Perc Percentile PSD Power Spectrum Density PCA Principal Component Analysis FS Proposed Feature Set RLS Recursive Least Squares RegTree Regression Tree RFN Regulatory Feedback Networks RE Relative ErrorR Gs Research Goals RQs Research Questions RMS Root Mean Square SNR Signal-to-Noise Ratio SSC Slope Sign Changes SD Standard Deviation SSOM Supervised Self-Organized Map SVM Support Vector Machine TD Time Domain TDAR Time Domain and Autoregressive TDF Time-Frequency Domain WL Waveform Length WP Wavelet Packet WT Wavelet Transforms ZC Zero Crossings

Contents

I Thesis 1

1 Introduction 1 1.1 Preprocessing ..................................................................................... 2 1.2 Feature Extraction and Classification ................................................ 3 1.3 Thesis overview ................................................................................. 4

2 Related Work 5 2.1 Removing ECG Interference ............................................................. 5 2.2 Offline Pattern Recognition ............................................................... 6 2.3 Real-Time Pattern Recognition ......................................................... 7

3 Research Overview 9 3.1 Research Goals and Research Questions ........................................... 9 3.2 Scientific Contributions ................................................................... 10

4 Methodology 15 4.1 Filtering Techniques for Removing ECG Interference ................... 15

4.1.1 Data Acquisition .............................................................. 15 4.1.2 Conventional Filtering Techniques ................................. 16

4.1.2.1 High-pass Filter ...................................................... 17 4.1.2.2 Spike Clipping ....................................................... 17 4.1.2.3 Gating Method ....................................................... 17 4.1.2.4 Hybrid Technique .................................................. 17 4.1.2.5 Template Subtraction ............................................. 18 4.1.2.6 Independent Component Analysis ......................... 18 4.1.2.7 Wavelet Transform ................................................ 19 4.1.2.8 Combined Wavelet and ICA .................................. 19 4.1.2.9 Adaptive Noise Canceller ...................................... 20 4.1.2.10 Artificial Neural Network .................................... 21

Contents

I Thesis 1

1 Introduction 1 1.1 Preprocessing ..................................................................................... 2 1.2 Feature Extraction and Classification ................................................ 3 1.3 Thesis overview ................................................................................. 4

2 Related Work 5 2.1 Removing ECG Interference ............................................................. 5 2.2 Offline Pattern Recognition ............................................................... 6 2.3 Real-Time Pattern Recognition ......................................................... 7

3 Research Overview 9 3.1 Research Goals and Research Questions ........................................... 9 3.2 Scientific Contributions ................................................................... 10

4 Methodology 15 4.1 Filtering Techniques for Removing ECG Interference ................... 15

4.1.1 Data Acquisition .............................................................. 15 4.1.2 Conventional Filtering Techniques ................................. 16

4.1.2.1 High-pass Filter ...................................................... 17 4.1.2.2 Spike Clipping ....................................................... 17 4.1.2.3 Gating Method ....................................................... 17 4.1.2.4 Hybrid Technique .................................................. 17 4.1.2.5 Template Subtraction ............................................. 18 4.1.2.6 Independent Component Analysis ......................... 18 4.1.2.7 Wavelet Transform ................................................ 19 4.1.2.8 Combined Wavelet and ICA .................................. 19 4.1.2.9 Adaptive Noise Canceller ...................................... 20 4.1.2.10 Artificial Neural Network .................................... 21

4.1.2.11 Adaptive Neuro-Fuzzy Inference System ............ 22 4.1.3 Proposed Novel Filtering Techniques ............................. 22

4.1.3.1 ANFIS-Wavelet (Paper A) ..................................... 22 4.1.3.2 ANN-Wavelet (Paper B) ........................................ 23 4.1.3.3 Adaptive Subtraction (Paper C) ............................. 23 4.1.3.4 Automated Wavelet-ICA (Paper D) ....................... 23

4.2 Offline Pattern Recognition (Paper E) ............................................. 24 4.2.1 Data Acquisition .............................................................. 24 4.2.2 Data Analysis ................................................................... 24

4.3 Real-Time Pattern Recognition (Paper F) ....................................... 28 4.3.1 Data Acquisition .............................................................. 28 4.3.2 Data Analysis ................................................................... 28

5 Results and Discussion 31 5.1 Filtering Techniques for Removing ECG Interference ................... 31

5.1.1 Conventional Filtering Techniques ................................. 31 5.1.1.1 High-Pass Filter ..................................................... 31 5.1.1.2 Spike Clipping ....................................................... 32 5.1.1.3 Gating Method ....................................................... 32 5.1.1.4 Hybrid Technique .................................................. 33 5.1.1.5 Template Subtraction ............................................. 35 5.1.1.6 Independent Component Analysis ......................... 36 5.1.1.7 Wavelet Transform ................................................ 37 5.1.1.8 Combined Wavelet and ICA .................................. 39 5.1.1.9 Adaptive Noise Canceller ...................................... 39 5.1.1.10 Artificial Neural Network .................................... 40 5.1.1.11 Adaptive Neuro-Fuzzy Inference System ............ 41

5.1.2 Proposed Novel Filtering Techniques ............................. 42 5.1.2.1 ANFIS-Wavelet (Paper A) ..................................... 42 5.1.2.2 ANN-Wavelet (Paper B) ........................................ 43 5.1.2.3 Adaptive Subtraction (Paper C) ............................. 44 5.1.2.4 Automated Wavelet-ICA (Paper D) ....................... 44

5.2 Offline Pattern Recognition (Paper E) ............................................. 49 5.3 Real-Time Pattern Recognition (Paper F) ....................................... 52

6 Conclusion 55

7 Future Work 57

4.1.2.11 Adaptive Neuro-Fuzzy Inference System ............ 22 4.1.3 Proposed Novel Filtering Techniques ............................. 22

4.1.3.1 ANFIS-Wavelet (Paper A) ..................................... 22 4.1.3.2 ANN-Wavelet (Paper B) ........................................ 23 4.1.3.3 Adaptive Subtraction (Paper C) ............................. 23 4.1.3.4 Automated Wavelet-ICA (Paper D) ....................... 23

4.2 Offline Pattern Recognition (Paper E) ............................................. 24 4.2.1 Data Acquisition .............................................................. 24 4.2.2 Data Analysis ................................................................... 24

4.3 Real-Time Pattern Recognition (Paper F) ....................................... 28 4.3.1 Data Acquisition .............................................................. 28 4.3.2 Data Analysis ................................................................... 28

5 Results and Discussion 31 5.1 Filtering Techniques for Removing ECG Interference ................... 31

5.1.1 Conventional Filtering Techniques ................................. 31 5.1.1.1 High-Pass Filter ..................................................... 31 5.1.1.2 Spike Clipping ....................................................... 32 5.1.1.3 Gating Method ....................................................... 32 5.1.1.4 Hybrid Technique .................................................. 33 5.1.1.5 Template Subtraction ............................................. 35 5.1.1.6 Independent Component Analysis ......................... 36 5.1.1.7 Wavelet Transform ................................................ 37 5.1.1.8 Combined Wavelet and ICA .................................. 39 5.1.1.9 Adaptive Noise Canceller ...................................... 39 5.1.1.10 Artificial Neural Network .................................... 40 5.1.1.11 Adaptive Neuro-Fuzzy Inference System ............ 41

5.1.2 Proposed Novel Filtering Techniques ............................. 42 5.1.2.1 ANFIS-Wavelet (Paper A) ..................................... 42 5.1.2.2 ANN-Wavelet (Paper B) ........................................ 43 5.1.2.3 Adaptive Subtraction (Paper C) ............................. 44 5.1.2.4 Automated Wavelet-ICA (Paper D) ....................... 44

5.2 Offline Pattern Recognition (Paper E) ............................................. 49 5.3 Real-Time Pattern Recognition (Paper F) ....................................... 52

6 Conclusion 55

7 Future Work 57

Bibliography 59

II Included Papers 69 8 Paper A: A Novel Approach for Removing ECG Interferences from Surface EMG signals Using a Combined ANFIS and Wavelet 71 8.1 Introduction ..................................................................................... 73 8.2 Material and Methods ...................................................................... 74

8.2.1 Signal Recording and Simulation .................................... 74 8.2.2 ANFIS .............................................................................. 76 8.2.3 Wavelet transform ........................................................... 77 8.2.4 ANFIS-Wavelet Transform ............................................. 78 8.2.5 Quantitative Evaluation Criteria ...................................... 78

8.3 Results ............................................................................................. 79 8.3.1 HPF Result ....................................................................... 81 8.3.2 ANN Result ..................................................................... 82 8.3.3 Wavelet Result ................................................................. 82 8.3.4 Template subtraction result ............................................. 82 8.3.5 ANC Result ...................................................................... 83 8.3.6 ANFIS Result .................................................................. 83

8.4 Discussion ........................................................................................ 85 Bibliography ................................................................................................ 86

9 Paper B: A Combination Method for Electrocardiogram Rejection from Surface Electromyogram 89 9.1 Introduction ..................................................................................... 91 9.2 Materials and Methods .................................................................... 92

9.2.1 Signal Recording and Simulation .................................... 92 9.2.2 Artificial Neural Network ................................................ 94 9.2.3 Wavelet Transform Based on Nonlinear Thresholding ... 95 9.2.4 Artificial Neural Network-Wavelet Transform ............... 96 9.2.5 Quantitative Evaluation Criteria ...................................... 97

9.3 Results ............................................................................................. 97 9.4 Discussion ...................................................................................... 100 Bibliography .............................................................................................. 102

Bibliography 59

II Included Papers 69 8 Paper A: A Novel Approach for Removing ECG Interferences from Surface EMG signals Using a Combined ANFIS and Wavelet 71 8.1 Introduction ..................................................................................... 73 8.2 Material and Methods ...................................................................... 74

8.2.1 Signal Recording and Simulation .................................... 74 8.2.2 ANFIS .............................................................................. 76 8.2.3 Wavelet transform ........................................................... 77 8.2.4 ANFIS-Wavelet Transform ............................................. 78 8.2.5 Quantitative Evaluation Criteria ...................................... 78

8.3 Results ............................................................................................. 79 8.3.1 HPF Result ....................................................................... 81 8.3.2 ANN Result ..................................................................... 82 8.3.3 Wavelet Result ................................................................. 82 8.3.4 Template subtraction result ............................................. 82 8.3.5 ANC Result ...................................................................... 83 8.3.6 ANFIS Result .................................................................. 83

8.4 Discussion ........................................................................................ 85 Bibliography ................................................................................................ 86

9 Paper B: A Combination Method for Electrocardiogram Rejection from Surface Electromyogram 89 9.1 Introduction ..................................................................................... 91 9.2 Materials and Methods .................................................................... 92

9.2.1 Signal Recording and Simulation .................................... 92 9.2.2 Artificial Neural Network ................................................ 94 9.2.3 Wavelet Transform Based on Nonlinear Thresholding ... 95 9.2.4 Artificial Neural Network-Wavelet Transform ............... 96 9.2.5 Quantitative Evaluation Criteria ...................................... 97

9.3 Results ............................................................................................. 97 9.4 Discussion ...................................................................................... 100 Bibliography .............................................................................................. 102

10 Paper C: Removing ECG Artifact from Surface EMG Signal Using Adaptive Subtraction Technique 107 10.1 Introduction ................................................................................... 109 10.2 Material and Methods ........................................................................ 109 10.3 Adaptive Subtraction Technique ................................................... 111

10.3.1 ECG Template ............................................................... 111 10.3.2 Low Pass Filter .............................................................. 112 10.3.3 Subtraction ..................................................................... 112 10.3.4 Evaluation Criteria ......................................................... 112

10.4 Results ........................................................................................... 113 10.5 Discussion ...................................................................................... 115 Bibliography .............................................................................................. 116

11 Paper D: ECG Artifact Removal from Surface EMG Signal Using an Automated Method Based on Wavelet-ICA 117 11.1 Introduction ................................................................................... 119 11.2 Material and Methods .................................................................... 119

11.2.1 Signal Recording and Simulating .................................. 119 11.2.2 Automated Wavelet-ICA Technique ............................. 120

11.3 Results ........................................................................................... 122 11.4 Discussion and Conclusion ............................................................ 123 Bibliography .............................................................................................. 124

12 Paper E: Evaluation of Surface EMG-Based Recognition Algorithms for Decoding Hand Movements 127 12.1 Introduction ................................................................................... 129 12.2 Methods ......................................................................................... 130

12.2.1 Data Collection and Preprocessing ................................ 130 12.2.2 Time Domain Features .................................................. 132 12.2.3 Frequency Domain Features .......................................... 137 12.2.4 Time-Frequency Domain Features ................................ 139 12.2.5 Dimensionality Reduction ............................................. 141 12.2.6 Classification ................................................................. 141

12.3 Results ........................................................................................... 144 12.3.1 TD Feature Result .......................................................... 144

10 Paper C: Removing ECG Artifact from Surface EMG Signal Using Adaptive Subtraction Technique 107 10.1 Introduction ................................................................................... 109 10.2 Material and Methods ........................................................................ 109 10.3 Adaptive Subtraction Technique ................................................... 111

10.3.1 ECG Template ............................................................... 111 10.3.2 Low Pass Filter .............................................................. 112 10.3.3 Subtraction ..................................................................... 112 10.3.4 Evaluation Criteria ......................................................... 112

10.4 Results ........................................................................................... 113 10.5 Discussion ...................................................................................... 115 Bibliography .............................................................................................. 116

11 Paper D: ECG Artifact Removal from Surface EMG Signal Using an Automated Method Based on Wavelet-ICA 117 11.1 Introduction ................................................................................... 119 11.2 Material and Methods .................................................................... 119

11.2.1 Signal Recording and Simulating .................................. 119 11.2.2 Automated Wavelet-ICA Technique ............................. 120

11.3 Results ........................................................................................... 122 11.4 Discussion and Conclusion ............................................................ 123 Bibliography .............................................................................................. 124

12 Paper E: Evaluation of Surface EMG-Based Recognition Algorithms for Decoding Hand Movements 127 12.1 Introduction ................................................................................... 129 12.2 Methods ......................................................................................... 130

12.2.1 Data Collection and Preprocessing ................................ 130 12.2.2 Time Domain Features .................................................. 132 12.2.3 Frequency Domain Features .......................................... 137 12.2.4 Time-Frequency Domain Features ................................ 139 12.2.5 Dimensionality Reduction ............................................. 141 12.2.6 Classification ................................................................. 141

12.3 Results ........................................................................................... 144 12.3.1 TD Feature Result .......................................................... 144

12.3.2 FD Feature Result .......................................................... 147 12.3.3 TFD Feature Result ....................................................... 148 12.3.4 Processing Time ............................................................ 150

12.4 Discussion ...................................................................................... 151 12.5 Conclusions ................................................................................... 153 Bibliography .............................................................................................. 154

13 Paper F: Real-Time and Offline Evaluation of Myoelectric Pattern Recognition for Upper Limb Prosthesis Control 159 13.1 Introduction ................................................................................... 161 13.2 Methodology .................................................................................. 162

13.2.1 Data Acquisition ............................................................ 162 13.2.2 Data Analysis ................................................................. 164 13.2.3 Evaluation ...................................................................... 167

13.3 Results ........................................................................................... 168 13.3.1 Accuracy ........................................................................ 168 13.3.2 Selection Time ............................................................... 173 13.3.3 Completion Time ........................................................... 174 13.3.4 Completion Rate ............................................................ 175

13.4 Discussion ...................................................................................... 175 13.4.1 Offline and Real-Time Accuracy .................................. 176 13.4.2 Trends in Offline and Real-Time Performance Metrics 177 13.4.3 Sources for Bias and Variability .................................... 177

13.5 Conclusions ................................................................................... 178 Bibliography .............................................................................................. 178

12.3.2 FD Feature Result .......................................................... 147 12.3.3 TFD Feature Result ....................................................... 148 12.3.4 Processing Time ............................................................ 150

12.4 Discussion ...................................................................................... 151 12.5 Conclusions ................................................................................... 153 Bibliography .............................................................................................. 154

13 Paper F: Real-Time and Offline Evaluation of Myoelectric Pattern Recognition for Upper Limb Prosthesis Control 159 13.1 Introduction ................................................................................... 161 13.2 Methodology .................................................................................. 162

13.2.1 Data Acquisition ............................................................ 162 13.2.2 Data Analysis ................................................................. 164 13.2.3 Evaluation ...................................................................... 167

13.3 Results ........................................................................................... 168 13.3.1 Accuracy ........................................................................ 168 13.3.2 Selection Time ............................................................... 173 13.3.3 Completion Time ........................................................... 174 13.3.4 Completion Rate ............................................................ 175

13.4 Discussion ...................................................................................... 175 13.4.1 Offline and Real-Time Accuracy .................................. 176 13.4.2 Trends in Offline and Real-Time Performance Metrics 177 13.4.3 Sources for Bias and Variability .................................... 177

13.5 Conclusions ................................................................................... 178 Bibliography .............................................................................................. 178

I

Thesis

I

Thesis

Chapter 1

Introduction

Losing a limb causes difficulties in daily life. To regain the ability to live an independent life, artificial limbs have been developed during the past decades. Hand prostheses are one of the artificial limbs that can be controlled by the user through the activity of the remnant muscles above the amputation. The electromyogram (EMG) is one of the sources that can be used as a control method for a hand prosthesis [1]. Surface EMGs are powerful non-invasive tools [2] that provide information about the neuromuscular activity of the studied muscle, which is essential to its use as a source of control for prosthetic limbs. However, the complexity of this signal introduces a great challenge to its applications [3]. EMG pattern recognition to decode different limb movements is an important advance for the control of powered prostheses [4, 5]. This approach could potentially enable the control of powered prosthetics using the EMG generated by muscular contractions as an input [6]. However, its use has yet to be transitioned into wide clinical practice [4, 5]. The processing stages of EMG pattern recognition are data collection, preprocessing (that is, filtering and windowing), feature extraction/selection, classification and evaluation (see Figure 1.1).

Figure 1.1: The flowchart of the processing stages of electromyogram pattern recognition.

EMG Signal Acquisition

Preprocessing Feature Extraction Classification Class Labels

Chapter 1

Introduction

Losing a limb causes difficulties in daily life. To regain the ability to live an independent life, artificial limbs have been developed during the past decades. Hand prostheses are one of the artificial limbs that can be controlled by the user through the activity of the remnant muscles above the amputation. The electromyogram (EMG) is one of the sources that can be used as a control method for a hand prosthesis [1]. Surface EMGs are powerful non-invasive tools [2] that provide information about the neuromuscular activity of the studied muscle, which is essential to its use as a source of control for prosthetic limbs. However, the complexity of this signal introduces a great challenge to its applications [3]. EMG pattern recognition to decode different limb movements is an important advance for the control of powered prostheses [4, 5]. This approach could potentially enable the control of powered prosthetics using the EMG generated by muscular contractions as an input [6]. However, its use has yet to be transitioned into wide clinical practice [4, 5]. The processing stages of EMG pattern recognition are data collection, preprocessing (that is, filtering and windowing), feature extraction/selection, classification and evaluation (see Figure 1.1).

Figure 1.1: The flowchart of the processing stages of electromyogram pattern recognition.


Preprocessing Feature Extraction Classification Class Labels

2 Chapter 1. Introduction

In the first stage (data collection), sensor devices are used to measure the required data. Challenges are inevitable in the data collection stage, especially when they are measured on the human body. First, the experimental design needs to be determined (what type of data, which area or body part to be measured, which sensor device and processing techniques to be used, etc.). After making decisions on all these aspects of the experimental design, it might be necessary to replace some of them with other choices. Then, it is time to write the ethical vetting application to get permission to record data from humans’ bodies. Writing this application requires a solid understanding of the study (the big picture of the whole study). After the application is approved, one is able to proceed with the recording stage. After finding a suitable place/room that can be used for the recording purposes (quiet, less distracting for both the subject and the data) and preparing everything that is necessary for the recording process, it is time to look for volunteers. Then, the experiment is performed. During the performance of the experiments, several unexpected events might happen (the device simply might not work or stop in the middle of recording, an unexpected result may be obtained because of problems with the software, there is noise despite filtering the signal, etc.). After the data are collected, it is subjected to the preprocessing stage.

1.1 Preprocessing The preprocessing stage is required to remove unwanted interferences from the recorded signals [4]. When surface EMG signals are recorded, noise is often captured from different sources such as inherent noise in the electronic components, power line interference, motion artifacts, and the inherent instability of signals and especially biological signals [7]. Electrical interferences produced by the heart (electrocardiogram (ECG)) significantly affect the EMG signal recorded from the upper trunk muscles and make its analysis and quantification unreliable [8] (see Figure 1.2). Where the quality of the EMG signal is of interest, it is essential to remove the ECG interferences from the EMG signals. However, this removal of the ECG interferences from the EMG signals is challenging because their frequency spectra considerably overlap. The frequency range of the surface EMG signals is between 0 Hz and 400 Hz, depending on the amount of fatty tissue and the muscle type [9]. The frequency range of the ECG signals is between 0 Hz and 200 Hz, and the highest power occurs at frequencies less than 45 Hz [9].


In the first stage (data collection), sensor devices are used to measure the required data. Challenges are inevitable in the data collection stage, especially when they are measured on the human body. First, the experimental design needs to be determined (what type of data, which area or body part to be measured, which sensor device and processing techniques to be used, etc.). After making decisions on all these aspects of the experimental design, it might be necessary to replace some of them with other choices. Then, it is time to write the ethical vetting application to get permission to record data from humans’ bodies. Writing this application requires a solid understanding of the study (the big picture of the whole study). After the application is approved, one is able to proceed with the recording stage. After finding a suitable place/room that can be used for the recording purposes (quiet, less distracting for both the subject and the data) and preparing everything that is necessary for the recording process, it is time to look for volunteers. Then, the experiment is performed. During the performance of the experiments, several unexpected events might happen (the device simply might not work or stop in the middle of recording, an unexpected result may be obtained because of problems with the software, there is noise despite filtering the signal, etc.). After the data are collected, it is subjected to the preprocessing stage.

1.1 Preprocessing The preprocessing stage is required to remove unwanted interferences from the recorded signals [4]. When surface EMG signals are recorded, noise is often captured from different sources such as inherent noise in the electronic components, power line interference, motion artifacts, and the inherent instability of signals and especially biological signals [7]. Electrical interferences produced by the heart (electrocardiogram (ECG)) significantly affect the EMG signal recorded from the upper trunk muscles and make its analysis and quantification unreliable [8] (see Figure 1.2). Where the quality of the EMG signal is of interest, it is essential to remove the ECG interferences from the EMG signals. However, this removal of the ECG interferences from the EMG signals is challenging because their frequency spectra considerably overlap. The frequency range of the surface EMG signals is between 0 Hz and 400 Hz, depending on the amount of fatty tissue and the muscle type [9]. The frequency range of the ECG signals is between 0 Hz and 200 Hz, and the highest power occurs at frequencies less than 45 Hz [9].

Chapter 1. Introduction 3

Figure 1.2: a) Surface electromyogram recorded from biceps muscle of the right side during regular muscle contraction and relaxation, b) EMG contaminated by ECG interference, obtained by adding ECG interference to the EMG shown in (a).

Few studies have performed direct comparisons between different methods for a given data set. Difficulties in comparison between studies arise due to the different signals, electrodes and recording techniques. Understanding the impact of filtering methods on the amplitude and frequency parameters of the desired signal is vital given the widespread use of EMGs. To achieve this understanding, the techniques that are employed must be assessed for efficacy and possible outcomes detrimental to the measurement of the EMG data. Hence, our objective in this thesis was to evaluate commonly used methods and propose suitable approaches to improve this process.

1.2 Feature Extraction and Classification After the preprocessing stage is done, pattern recognition takes advantage of the feature extraction to extract maximal information from the recorded EMG about the performed motions [6]. Different features in the time domain (TD) [4], frequency domain (FD) and time-frequency domain (TFD) [10] are used to characterize EMG signals with discrete values. TD features are statistical properties of EMG signals in the time domain. They have a lower processing time and small dimensions and are easy to implement compared to features in other domains. FD features are usually statistical properties of the power spectral density (PSD) of the EMG signal [11]. These features have been used in few studies on movement pattern recognition [12-14]; features in this domain are mostly used to assess muscle fatigue and analyze

0 2 4 6 8 10 12 14 16 18 20-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

Amplitude(mv)

a

0 2 4 6 8 10 12 14 16 18 20Time(s)

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

Amplitude(mv)

b

Chapter 1. Introduction 3

Figure 1.2: a) Surface electromyogram recorded from biceps muscle of the right side during regular muscle contraction and relaxation, b) EMG contaminated by ECG interference, obtained by adding ECG interference to the EMG shown in (a).

Few studies have performed direct comparisons between different methods for a given data set. Difficulties in comparison between studies arise due to the different signals, electrodes and recording techniques. Understanding the impact of filtering methods on the amplitude and frequency parameters of the desired signal is vital given the widespread use of EMGs. To achieve this understanding, the techniques that are employed must be assessed for efficacy and possible outcomes detrimental to the measurement of the EMG data. Hence, our objective in this thesis was to evaluate commonly used methods and propose suitable approaches to improve this process.

1.2 Feature Extraction and Classification After the preprocessing stage is done, pattern recognition takes advantage of the feature extraction to extract maximal information from the recorded EMG about the performed motions [6]. Different features in the time domain (TD) [4], frequency domain (FD) and time-frequency domain (TFD) [10] are used to characterize EMG signals with discrete values. TD features are statistical properties of EMG signals in the time domain. They have a lower processing time and small dimensions and are easy to implement compared to features in other domains. FD features are usually statistical properties of the power spectral density (PSD) of the EMG signal [11]. These features have been used in few studies on movement pattern recognition [12-14]; features in this domain are mostly used to assess muscle fatigue and analyze

0 2 4 6 8 10 12 14 16 18 20-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

Amplitude(mv)

a

0 2 4 6 8 10 12 14 16 18 20Time(s)

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

Amplitude(mv)

b


motor unit recruitment [10]. TD and FD features are suitable for analyzing stationary signals. However, EMG signals are nonstationary and therefore features in the TFD have been introduced [15]. Features in the TFD are obtained after applying wavelet transforms (WT) to EMG signals. If the selected feature set occupies a high-dimensional space, a dimensionality reduction method is used to decrease the dimensionality of the feature space [16]. The extracted features from the EMG signals are fed to a classifier to decode different patterns.

Different algorithms have been developed to decode different movements. Often, these algorithms provide high offline classification accuracies in decoding various motions. Since offline analysis is easier to perform and no active subjects are needed for the trial, it is the first step in evaluating pattern recognition algorithms. Offline analysis removes a great amount of variability from the recordings, which helps maintain consistency between experiments but also reduces how well the results mirror the performance in real life. As the relative performance of the classifiers is not maintained between offline and real-time analyses, optimal feature extraction and robust classification continue to be open challenges, and improvements in these areas could potentially increase the functionality of powered prostheses. To address this challenge, this thesis investigated the offline performance of a wide range of pattern recognition algorithms. New myoelectric control configurations were obtained to improve motion recognition based on several evaluation criteria. Since the majority of investigations on the state of the art have been conducted offline, and minimal experience has been gained with real-time use in clinical settings, a comparative real-time performance analysis between algorithms had yet to be performed. To this purpose, this thesis investigated the offline and real-time performance of classification algorithms in decoding individual hand and wrist movements.

1.3 Thesis overview This thesis consists of two parts:

The thesis begins with an introduction in Chapter 1; Chapter 2 introduces related work in the research area. Chapter 3 presents a research overview consisting of the research goals and research questions followed by corresponding research contributions. Chapter 4 describes the research methodology. Chapter 5 summarizes and discusses the results. Chapter 6 draws conclusions from the results, and finally, Chapter 7 provides suggestions for future work.

The second part of the thesis presents a collection of six research papers.


motor unit recruitment [10]. TD and FD features are suitable for analyzing stationary signals. However, EMG signals are nonstationary and therefore features in the TFD have been introduced [15]. Features in the TFD are obtained after applying wavelet transforms (WT) to EMG signals. If the selected feature set occupies a high-dimensional space, a dimensionality reduction method is used to decrease the dimensionality of the feature space [16]. The extracted features from the EMG signals are fed to a classifier to decode different patterns.

Different algorithms have been developed to decode different movements. Often, these algorithms provide high offline classification accuracies in decoding various motions. Since offline analysis is easier to perform and no active subjects are needed for the trial, it is the first step in evaluating pattern recognition algorithms. Offline analysis removes a great amount of variability from the recordings, which helps maintain consistency between experiments but also reduces how well the results mirror the performance in real life. As the relative performance of the classifiers is not maintained between offline and real-time analyses, optimal feature extraction and robust classification continue to be open challenges, and improvements in these areas could potentially increase the functionality of powered prostheses. To address this challenge, this thesis investigated the offline performance of a wide range of pattern recognition algorithms. New myoelectric control configurations were obtained to improve motion recognition based on several evaluation criteria. Since the majority of investigations on the state of the art have been conducted offline, and minimal experience has been gained with real-time use in clinical settings, a comparative real-time performance analysis between algorithms had yet to be performed. To this purpose, this thesis investigated the offline and real-time performance of classification algorithms in decoding individual hand and wrist movements.

1.3 Thesis overview This thesis consists of two parts:

The thesis begins with an introduction in Chapter 1; Chapter 2 introduces related work in the research area. Chapter 3 presents a research overview consisting of the research goals and research questions followed by corresponding research contributions. Chapter 4 describes the research methodology. Chapter 5 summarizes and discusses the results. Chapter 6 draws conclusions from the results, and finally, Chapter 7 provides suggestions for future work.

The second part of the thesis presents a collection of six research papers.

Chapter 2

Related Work

This chapter provides the required background information and related work, both to point out the many contributions of previous researchers and to place our contributions in the proper context.

2.1 Removing ECG Interference The choice of an efficient method for removing artifacts from biomedical signals is crucial in achieving reliable signal processing. ECG artifact removal from EMG signals has been performed using techniques such as a high-pass filter (HPF) [17], gating [18], spike clipping [19], a hybrid approach [20], the subtraction method [21], independent component analysis (ICA) [7, 22-24], wavelet transforms [25-27], wavelet-ICA [28], artificial neural networks (ANNs) [29, 30], an adaptive noise canceller (ANC) [30-33] and an adaptive neuro-fuzzy inference system (ANFIS) [34-37]. The conventional high-pass filter is a simple and fast method, but it is not suitable for more complex applications such as prosthetic control since it removes a large amount of useful information from the EMG signals [30]. In template subtraction, the efficacy relies on the accuracy of QRS complex detection and the stationarity of the ECG signals [30]. The gating method is perhaps the most frequently used technique for ECG removal, which, although simple, suffers from the loss of portions of the EMG signal overlying the QRS complexes [18]. An ANC has been used to reduce the ECG artifacts [18], but due to the heavy computational cost, it is not suitable for clinical applications [25]. The wavelet transform is an online method with a low computation cost that does not require multiple inputs. However, using this method, some artifacts remain in the original signal, and part of the desired signal is removed [38]. The ICA method is an online method that operates on a

Chapter 2

Related Work

This chapter provides the required background information and related work, both to point out the many contributions of previous researchers and to place our contributions in the proper context.

2.1 Removing ECG Interference The choice of an efficient method for removing artifacts from biomedical signals is crucial in achieving reliable signal processing. ECG artifact removal from EMG signals has been performed using techniques such as a high-pass filter (HPF) [17], gating [18], spike clipping [19], a hybrid approach [20], the subtraction method [21], independent component analysis (ICA) [7, 22-24], wavelet transforms [25-27], wavelet-ICA [28], artificial neural networks (ANNs) [29, 30], an adaptive noise canceller (ANC) [30-33] and an adaptive neuro-fuzzy inference system (ANFIS) [34-37]. The conventional high-pass filter is a simple and fast method, but it is not suitable for more complex applications such as prosthetic control since it removes a large amount of useful information from the EMG signals [30]. In template subtraction, the efficacy relies on the accuracy of QRS complex detection and the stationarity of the ECG signals [30]. The gating method is perhaps the most frequently used technique for ECG removal, which, although simple, suffers from the loss of portions of the EMG signal overlying the QRS complexes [18]. An ANC has been used to reduce the ECG artifacts [18], but due to the heavy computational cost, it is not suitable for clinical applications [25]. The wavelet transform is an online method with a low computation cost that does not require multiple inputs. However, using this method, some artifacts remain in the original signal, and part of the desired signal is removed [38]. The ICA method is an online method that operates on a

6 Chapter 2. Related Work

multichannel signal that adds to the complexity of the hardware [39]. Although different techniques have been proposed in the state of the art to remove ECG artifacts from EMG signals, the problem of the accurate and effective denoising of an EMG remains a challenge. Among all the proposed methods in the literature, ANN and ANFIS have shown the best performance, but the results of these methods could be improved when used in combination with other techniques for specific applications. Table 5.2 provides extensive information on the advantages and disadvantages of the filtering methods.

2.2 Offline Pattern Recognition Various algorithms have been proposed to enhance the functionality and usability of prosthetic hands [40] based on feature extraction and motion classification methods [41]. Englehart et al. [3] extracted features in both the time and time-frequency domains. Linear discriminant analysis (LDA) and multilayer perceptron (MLP) were used as classifiers to decode four classes of movements using two channels of surface EMG signals. Accuracy rates up to 93.7% were achieved. Phinyomark et al. evaluated the performance of an LDA classifier in combination with 37 features in the time and frequency domains to discriminate six hand gestures using five channels of surface EMG signals. They obtained classification accuracies up to 92.1% [14]. Oskoei et al. [42] evaluated the performance of ANN and LDA classifiers in combination with different feature sets including a set of TD features introduced by Hudgins et al. [43] (mean absolute value [MAV], zero crossings [ZC], slope sign changes [SSC], and waveform length [WL]) to discriminate six classes of movements. They obtained the lowest classification error of 1.88% (accuracy of 98.1%). Hargrove et al. [44] compared the classification accuracy of ANN and LDA classifiers in combination with four different feature sets – the Hudgins TD feature set, the autoregressive (AR) model, combined TD and AR (TDAR), and root mean square (RMS) – to discriminate 10 different classes of isometric contraction. The results on 12 healthy subjects revealed that the TDAR/LDA combination had the best performance, with an accuracy up to 97%. Phinyomark and Scheme [45] investigated the performance of a support vector machine (SVM) classifier in combination with several individual features and feature sets to discriminate several hand gestures using EMG signals obtained from different datasets, with classification accuracies up to 95%. Al-Taee and Al-Jumaily [46] proposed a set of features to decode ten individual and combined finger movements in combination with a sequential forward searching classifier. Two-channel surface EMG signals recorded from eight subjects were used, and they achieved an average accuracy rate of 99.1%. They also performed a comparison between their proposed feature set and other well-known feature sets,


multichannel signal that adds to the complexity of the hardware [39]. Although different techniques have been proposed in the state of the art to remove ECG artifacts from EMG signals, the problem of the accurate and effective denoising of an EMG remains a challenge. Among all the proposed methods in the literature, ANN and ANFIS have shown the best performance, but the results of these methods could be improved when used in combination with other techniques for specific applications. Table 5.2 provides extensive information on the advantages and disadvantages of the filtering methods.

2.2 Offline Pattern Recognition Various algorithms have been proposed to enhance the functionality and usability of prosthetic hands [40] based on feature extraction and motion classification methods [41]. Englehart et al. [3] extracted features in both the time and time-frequency domains. Linear discriminant analysis (LDA) and multilayer perceptron (MLP) were used as classifiers to decode four classes of movements using two channels of surface EMG signals. Accuracy rates up to 93.7% were achieved. Phinyomark et al. evaluated the performance of an LDA classifier in combination with 37 features in the time and frequency domains to discriminate six hand gestures using five channels of surface EMG signals. They obtained classification accuracies up to 92.1% [14]. Oskoei et al. [42] evaluated the performance of ANN and LDA classifiers in combination with different feature sets including a set of TD features introduced by Hudgins et al. [43] (mean absolute value [MAV], zero crossings [ZC], slope sign changes [SSC], and waveform length [WL]) to discriminate six classes of movements. They obtained the lowest classification error of 1.88% (accuracy of 98.1%). Hargrove et al. [44] compared the classification accuracy of ANN and LDA classifiers in combination with four different feature sets – the Hudgins TD feature set, the autoregressive (AR) model, combined TD and AR (TDAR), and root mean square (RMS) – to discriminate 10 different classes of isometric contraction. The results on 12 healthy subjects revealed that the TDAR/LDA combination had the best performance, with an accuracy up to 97%. Phinyomark and Scheme [45] investigated the performance of a support vector machine (SVM) classifier in combination with several individual features and feature sets to discriminate several hand gestures using EMG signals obtained from different datasets, with classification accuracies up to 95%. Al-Taee and Al-Jumaily [46] proposed a set of features to decode ten individual and combined finger movements in combination with a sequential forward searching classifier. Two-channel surface EMG signals recorded from eight subjects were used, and they achieved an average accuracy rate of 99.1%. They also performed a comparison between their proposed feature set and other well-known feature sets,

Chapter 2. Related Work 7

including the Hudgins set. Table 2.1 summarizes the obtained offline classification accuracies in the state of the art. Although these offline studies have shown that the accurate decoding of gestures from electrodes placed on the forearm can be achieved, optimal feature extraction and robust classification continue to be open challenges. Furthermore, due to the variations in the methodologies used to evaluate the algorithms, it is difficult to compare their results. Only a few studies have quantitatively evaluated the performance of a wide range of classifiers and features to discriminate hand and finger movements using the same database and methodology [11, 14, 41, 42, 44].

Table 2.1: Selected examples of offline classification accuracies achieved from the state of the art.

Authors Classification accuracy (%) Reference

Englehart et al. 93.7 [3]

Phinyomark et al. 92.1 [14]

Oskoei et al. 98.1 [42]

Hargrove et al. 97.0 [44]

Phinyomark and Scheme 95.0 [45]

Al-Taee and Al-Jumaily 99.1 [46]

2.3 Real-Time Pattern Recognition Since most investigations have been performed offline, the questions of real-time control and its accuracy have been left open [47]. A limited number of studies have been performed to investigate the real-time performance of pattern recognition algorithms. Guo et al. [41] performed a real-time survey on seven healthy subjects to compare the ANN and SVM classifiers in combination with four features. Their experimental results on six channels of surface EMG signals showed that SVM performed better than LDA in real time, with an average accuracy of 85.9%. Wang et al. [48] investigated the offline performance of SVM and naive Bayes classifiers in combination with eight time-domain features to decode eight finger movements using five channels of surface EMG. Based on the results, six features and one classifier (SVM) were chosen for real-time tests on four healthy subjects. Real-time accuracies between 90-100% were achieved when the subjects were presented with visual feedback. Sebelius et al. [49] conducted a six-subject study to investigate the effect of training subjects on the acquired accuracy using a virtual hand for feedback. They employed local approximation using a lazy learning algorithm to classify muscle patterns of ten wrist/finger movements using eight channels of surface EMG. Ortiz-

Chapter 2. Related Work 7

including the Hudgins set. Table 2.1 summarizes the obtained offline classification accuracies in the state of the art. Although these offline studies have shown that the accurate decoding of gestures from electrodes placed on the forearm can be achieved, optimal feature extraction and robust classification continue to be open challenges. Furthermore, due to the variations in the methodologies used to evaluate the algorithms, it is difficult to compare their results. Only a few studies have quantitatively evaluated the performance of a wide range of classifiers and features to discriminate hand and finger movements using the same database and methodology [11, 14, 41, 42, 44].

Table 2.1: Selected examples of offline classification accuracies achieved from the state of the art.


Englehart et al. 93.7 [3]

Phinyomark et al. 92.1 [14]

Oskoei et al. 98.1 [42]

Hargrove et al. 97.0 [44]

Phinyomark and Scheme 95.0 [45]

Al-Taee and Al-Jumaily 99.1 [46]

2.3 Real-Time Pattern Recognition Since most investigations have been performed offline, the questions of real-time control and its accuracy have been left open [47]. A limited number of studies have been performed to investigate the real-time performance of pattern recognition algorithms. Guo et al. [41] performed a real-time survey on seven healthy subjects to compare the ANN and SVM classifiers in combination with four features. Their experimental results on six channels of surface EMG signals showed that SVM performed better than LDA in real time, with an average accuracy of 85.9%. Wang et al. [48] investigated the offline performance of SVM and naive Bayes classifiers in combination with eight time-domain features to decode eight finger movements using five channels of surface EMG. Based on the results, six features and one classifier (SVM) were chosen for real-time tests on four healthy subjects. Real-time accuracies between 90-100% were achieved when the subjects were presented with visual feedback. Sebelius et al. [49] conducted a six-subject study to investigate the effect of training subjects on the acquired accuracy using a virtual hand for feedback. They employed local approximation using a lazy learning algorithm to classify muscle patterns of ten wrist/finger movements using eight channels of surface EMG. Ortiz-


Catalan et al. [50] compared the offline and real-time performances of LDA, MLP and regulatory feedback network (RFN) classifiers in combination with the Hudgins set of time domain features to decode 10 classes of individual motions. The highest offline classification accuracy, with an average of 92.1%, was achieved by LDA, and RFN achieved the highest real-time classification accuracy, with an average of 67.4%. Ortiz-Catalan et al. [51], in another study, evaluated the offline performance of four different classification algorithms (LDA, MLP, self-organized map and RFN) in combination with the Hudgins set of TD features to decode individual and simultaneous movements (a total of 27 movements) using eight channels of forearm surface EMG signals. MLP was chosen for the real-time test on 10 healthy subjects, and average accuracies of 54.9% and 54.8% were obtained for individual and simultaneous movements, respectively. Table 2.2 summarizes the real-time classification accuracies obtained using different methods. These studies illustrate that the offline accuracy can be a misleading metric to evaluate the usability of decoding algorithms for real-time applications [51]. Since a good offline accuracy does not guarantee good online results and the results obtained from offline pattern recognition might only satisfy the researcher and not the clinicians, further investigation is required to corroborate the translation of these findings to real-time performance [52].

Table 2.2: Selected examples of real-time classification accuracies achieved from the state of the art.


Guo et al. 85.9 [41]

Wang et al. 90-100 [48]

Ortiz-Catalan et al. 67.4 [50]




Catalan et al. [50] compared the offline and real-time performances of LDA, MLP and regulatory feedback network (RFN) classifiers in combination with the Hudgins set of time domain features to decode 10 classes of individual motions. The highest offline classification accuracy, with an average of 92.1%, was achieved by LDA, and RFN achieved the highest real-time classification accuracy, with an average of 67.4%. Ortiz-Catalan et al. [51], in another study, evaluated the offline performance of four different classification algorithms (LDA, MLP, self-organized map and RFN) in combination with the Hudgins set of TD features to decode individual and simultaneous movements (a total of 27 movements) using eight channels of forearm surface EMG signals. MLP was chosen for the real-time test on 10 healthy subjects, and average accuracies of 54.9% and 54.8% were obtained for individual and simultaneous movements, respectively. Table 2.2 summarizes the real-time classification accuracies obtained using different methods. These studies illustrate that the offline accuracy can be a misleading metric to evaluate the usability of decoding algorithms for real-time applications [51]. Since a good offline accuracy does not guarantee good online results and the results obtained from offline pattern recognition might only satisfy the researcher and not the clinicians, further investigation is required to corroborate the translation of these findings to real-time performance [52].

Table 2.2: Selected examples of real-time classification accuracies achieved from the state of the art.


Guo et al. 85.9 [41]

Wang et al. 90-100 [48]




Chapter 3

Research Overview

An overview of the doctoral thesis is provided in this chapter. First, the research goals and research questions of this thesis are presented. Then, a summary of the scientific contributions are provided.

3.1 Research Goals and Research Questions Accuracy in recognizing different movements and a low response time are crucial for a successful real-time surface EMG-based control system [53]. In addition to the choice of feature extraction and classification methods, the quality of the surface EMG signal is an important factor that affects the classification accuracy. Therefore, the research goals (RGs) of this thesis are:

RG1: Enhancing the filtering process of surface EMG signals in removing ECG interference by

• Combining some existing methods

• Converting conventional methods to online and automatic approaches

with the aim of improving the EMG signal quality based on several quantitative and qualitative criteria.

RG2: Investigating both the offline and real-time performance of myoelectric pattern recognition algorithms to

• Determine new configurations that improve the accuracy of hand gesture recognition with surface EMG signals

Chapter 3

Research Overview

An overview of the doctoral thesis is provided in this chapter. First, the research goals and research questions of this thesis are presented. Then, a summary of the scientific contributions are provided.

3.1 Research Goals and Research Questions Accuracy in recognizing different movements and a low response time are crucial for a successful real-time surface EMG-based control system [53]. In addition to the choice of feature extraction and classification methods, the quality of the surface EMG signal is an important factor that affects the classification accuracy. Therefore, the research goals (RGs) of this thesis are:

RG1: Enhancing the filtering process of surface EMG signals in removing ECG interference by

• Combining some existing methods

• Converting conventional methods to online and automatic approaches

with the aim of improving the EMG signal quality based on several quantitative and qualitative criteria.

RG2: Investigating both the offline and real-time performance of myoelectric pattern recognition algorithms to

• Determine new configurations that improve the accuracy of hand gesture recognition with surface EMG signals

10 Chapter 3. Research Overview

• Gain proper insight into the clinical implementation of myoelectric pattern recognition

This will lead to the development of an accurate surface EMG-based sensing system and the reduction of recognition error in controlling prosthetic hands.

During the development and investigation process, some research questions (RQs) are expected to be answered.

RQ1: How can different strategies be applied to enhance the filtering process of surface EMG signals in removing ECG interference?

a) What combination of filtering methods could improve the results based on quantitative and qualitative criteria?

b) How can a template subtraction method be converted to an online method?

c) How can a user-dependent wavelet-ICA be converted to an automatic method?

RQ2: Which configurations of myoelectric pattern recognition could improve the offline performance of hand gesture recognition (based on the obtained accuracy rate and processing time)?

RQ3: How accurate could myoelectric pattern recognition algorithms perform in real time compared to offline?

a) Which algorithms are potentially optimal for real-time prosthetic control?

b) Which hand movements could be recognized most accurately in offline and in real time?

3.2 Scientific Contributions The scientific contributions of the thesis, which address the formulated research questions, are presented in this section. The contributions are organized in five parts. In the first part, I propose combined approaches to improve the performance of several filtering methods in removing ECG interference from surface EMG signals. To evaluate and compare the performance of our proposed techniques with that of the existing methods, in addition to the implementation of the proposed techniques, I also implement several conventional filtering methods. Furthermore, I record EMG and ECG signals from several healthy volunteers. In the second part, I introduce an online technique to enhance the filtering process of surface EMG signals in removing ECG interference. In the third part, I develop an automated approach to improve the ECG interference removal process from surface EMG signals. In the fourth part, I identify


• Gain proper insight into the clinical implementation of myoelectric pattern recognition

This will lead to the development of an accurate surface EMG-based sensing system and the reduction of recognition error in controlling prosthetic hands.

During the development and investigation process, some research questions (RQs) are expected to be answered.

RQ1: How can different strategies be applied to enhance the filtering process of surface EMG signals in removing ECG interference?

a) What combination of filtering methods could improve the results based on quantitative and qualitative criteria?

b) How can a template subtraction method be converted to an online method?

c) How can a user-dependent wavelet-ICA be converted to an automatic method?

RQ2: Which configurations of myoelectric pattern recognition could improve the offline performance of hand gesture recognition (based on the obtained accuracy rate and processing time)?

RQ3: How accurate could myoelectric pattern recognition algorithms perform in real time compared to offline?

a) Which algorithms are potentially optimal for real-time prosthetic control?

b) Which hand movements could be recognized most accurately in offline and in real time?

3.2 Scientific Contributions The scientific contributions of the thesis, which address the formulated research questions, are presented in this section. The contributions are organized in five parts. In the first part, I propose combined approaches to improve the performance of several filtering methods in removing ECG interference from surface EMG signals. To evaluate and compare the performance of our proposed techniques with that of the existing methods, in addition to the implementation of the proposed techniques, I also implement several conventional filtering methods. Furthermore, I record EMG and ECG signals from several healthy volunteers. In the second part, I introduce an online technique to enhance the filtering process of surface EMG signals in removing ECG interference. In the third part, I develop an automated approach to improve the ECG interference removal process from surface EMG signals. In the fourth part, I identify

Chapter 3. Research Overview 11

an efficient feature set and new configurations of myoelectric pattern recognition to improve the hand motion recognition accuracy. Finally, in the fifth part, I evaluate the real-time and offline performance of recognition algorithms for classifying individual hand movements by recording EMG signals from 15 healthy subjects (12 trials per subject) for the offline and real-time evaluations. Table 3.1 summarizes the contributions of this doctoral thesis, connecting the research questions to the included papers.

Table 3.1: Summarizing the contributions of this doctoral thesis, connecting the research questions to the included papers.

Papers RQ1 RQ2 RQ3

Paper A Contribution 1: proposing combined approaches

Paper B Contribution 1: proposing combined approaches

Paper C

Contribution 1: proposing combined approaches, Contribution 2: introducing an online technique

Paper D

Contribution 1: proposing combined approaches, Contribution 3: developing an automated approach

Paper E

Contribution 4: identifying an efficient feature set and new configurations of myoelectric pattern recognition

Paper F

Contribution 5: evaluating real-time and offline performance of recognition algorithms

Contributions 1: Combined Approaches

Two combined filtering approaches, ANFIS-wavelet and ANN-wavelet, were proposed to improve the quality of surface EMG signals based on several quantitative and qualitative criteria. In these approaches, ANFIS/ANN was first employed to remove ECG interference from contaminated EMG signals. In this step, a large amount of pollution was removed; however, the result showed that there are still low-frequency noise components in the denoised EMG signal. A wavelet transform with


an efficient feature set and new configurations of myoelectric pattern recognition to improve the hand motion recognition accuracy. Finally, in the fifth part, I evaluate the real-time and offline performance of recognition algorithms for classifying individual hand movements by recording EMG signals from 15 healthy subjects (12 trials per subject) for the offline and real-time evaluations. Table 3.1 summarizes the contributions of this doctoral thesis, connecting the research questions to the included papers.

Table 3.1: Summarizing the contributions of this doctoral thesis, connecting the research questions to the included papers.

Papers RQ1 RQ2 RQ3

Paper A Contribution 1: proposing combined approaches

Paper B Contribution 1: proposing combined approaches

Paper C

Contribution 1: proposing combined approaches, Contribution 2: introducing an online technique

Paper D

Contribution 1: proposing combined approaches, Contribution 3: developing an automated approach

Paper E

Contribution 4: identifying an efficient feature set and new configurations of myoelectric pattern recognition

Paper F

Contribution 5: evaluating real-time and offline performance of recognition algorithms

Contributions 1: Combined Approaches

Two combined filtering approaches, ANFIS-wavelet and ANN-wavelet, were proposed to improve the quality of surface EMG signals based on several quantitative and qualitative criteria. In these approaches, ANFIS/ANN was first employed to remove ECG interference from contaminated EMG signals. In this step, a large amount of pollution was removed; however, the result showed that there are still low-frequency noise components in the denoised EMG signal. A wavelet transform with


nonlinear thresholding was proposed as a postprocessor to remove the residual noise from the output of the ANFIS/ANN method. In addition to the abovementioned techniques, an adaptive subtraction in combination with a low-pass filter was proposed. The low-pass filter was used to remove noise from the obtained ECG templates. An automated wavelet-ICA, which is a combination of three different techniques (wavelet transform, ICA and an HPF), was proposed to remove ECG interference from automatically detected noisy components. To compare the results and show the effectiveness of our proposed technique, in addition to our proposed approaches, I implemented eleven commonly used methods, recorded the required signals (e.g., EMG, ECG) from several healthy volunteers.

Targeting research question: RQ1 (a, b, and c) Included papers: Papers A, B, C and D

Contributions 2: Online Approach

The subtraction method is one of the filtering techniques that is used to remove ECGs from EMGs. This method achieved very good performance in improving the quality of EMGs; however, it is not suitable for online applications. Therefore, an approach named adaptive subtraction was proposed to convert the template subtraction to an online technique. This method contains four steps; (1) QRS detection in the contaminated EMG signals, (2) formation of an ECG template by averaging the detected electrocardiogram complexes, (3) using a low-pass filter to remove noise from the obtained ECG template, and (4) subtracting the ECG template from the contaminated EMG signal where the QRS complexes are detected. The proposed technique is an online method that is somewhat effective in removing ECG artifacts.

Targeting research question: RQ1 b Included papers: Paper C

Contributions 3: Automated Approach

An automated wavelet-ICA was proposed to convert the user-dependent wavelet-ICA to an automatic method with a better result based on its qualitative and quantitative criteria. In conventional ICA-based methods, after creating the ICA components, independent components are classified manually as the EMG signal or ECG interference. In this proposed technique, an automated algorithm was used to detect the ECG interference components automatically. Each channel of independent components was evaluated separately, and the components with ECG interference were selected automatically. After finding the noisy component in ICA-based methods, the noisy component is usually set to zero. This procedure removes useful


nonlinear thresholding was proposed as a postprocessor to remove the residual noise from the output of the ANFIS/ANN method. In addition to the abovementioned techniques, an adaptive subtraction in combination with a low-pass filter was proposed. The low-pass filter was used to remove noise from the obtained ECG templates. An automated wavelet-ICA, which is a combination of three different techniques (wavelet transform, ICA and an HPF), was proposed to remove ECG interference from automatically detected noisy components. To compare the results and show the effectiveness of our proposed technique, in addition to our proposed approaches, I implemented eleven commonly used methods, recorded the required signals (e.g., EMG, ECG) from several healthy volunteers.

Targeting research question: RQ1 (a, b, and c) Included papers: Papers A, B, C and D

Contributions 2: Online Approach

The subtraction method is one of the filtering techniques that is used to remove ECGs from EMGs. This method achieved very good performance in improving the quality of EMGs; however, it is not suitable for online applications. Therefore, an approach named adaptive subtraction was proposed to convert the template subtraction to an online technique. This method contains four steps; (1) QRS detection in the contaminated EMG signals, (2) formation of an ECG template by averaging the detected electrocardiogram complexes, (3) using a low-pass filter to remove noise from the obtained ECG template, and (4) subtracting the ECG template from the contaminated EMG signal where the QRS complexes are detected. The proposed technique is an online method that is somewhat effective in removing ECG artifacts.

Targeting research question: RQ1 b Included papers: Paper C

Contributions 3: Automated Approach

An automated wavelet-ICA was proposed to convert the user-dependent wavelet-ICA to an automatic method with a better result based on its qualitative and quantitative criteria. In conventional ICA-based methods, after creating the ICA components, independent components are classified manually as the EMG signal or ECG interference. In this proposed technique, an automated algorithm was used to detect the ECG interference components automatically. Each channel of independent components was evaluated separately, and the components with ECG interference were selected automatically. After finding the noisy component in ICA-based methods, the noisy component is usually set to zero. This procedure removes useful


information from the reconstructed signal. Therefore, a high-pass filter with a cutoff frequency of 20 Hz was used to remove ECG interference from the selected noisy component. Finally, inverse ICA and then an inverse wavelet transform were applied to reconstruct the denoised EMG signal.

Targeting research question: RQ1 c Included papers: Paper D

Contributions 4: New Configurations of Myoelectric Pattern Recognition

In this thesis, I quantitatively evaluated the performance of a wide range of classifiers and features to discriminate hand and finger movements using the same database and methodology to avoid a lot of variability in the recordings and maintain consistency between experiments so that it is possible to compare the results of different algorithms. Various combinations of 44 different features in different domains (time, frequency, and time-frequency) and six classifiers were investigated offline to determine new myoelectric control configurations that improve the classification accuracy of motion recognition with surface EMG signals. Some of the features are commonly used in this area (and were chosen by extensively reviewing the literature), and some of them are new (and I would like to present their results). Among the classifiers, LDA, MLP, K-nearest neighborhood (KNN) and SVM were chosen by reviewing the state of the art; however, there has not been enough investigation on the performance of maximum likelihood estimation (MLE) and regression trees (RegTree) for hand gesture recognition, although they have been extensively used in different areas such as gait measurement and speech recognition. As a result, an efficient feature set and several myoelectric control configurations were proposed to reduce the recognition error in controlling prosthetic hands.

Targeting research question: RQ2 Included papers: Paper E

Contributions 5: Real-time and Offline Evaluation of Recognition Algorithms

The real-time and offline performance of nine classification algorithms in detecting ten individual hand movements were investigated using four channels of surface EMG signals. Some of the investigated algorithms were commonly used in the literature, and some were based on relatively new paradigms in pattern recognition, which were chosen by intensively reviewing the state of the art. The classification accuracy and processing time were used as the evaluation criteria to compare the experimental results in offline and in real time. To conduct this study, I made some modifications and implemented some additional algorithms (i.e., KNN, RegTree,


information from the reconstructed signal. Therefore, a high-pass filter with a cutoff frequency of 20 Hz was used to remove ECG interference from the selected noisy component. Finally, inverse ICA and then an inverse wavelet transform were applied to reconstruct the denoised EMG signal.

Targeting research question: RQ1 c Included papers: Paper D

Contributions 4: New Configurations of Myoelectric Pattern Recognition

In this thesis, I quantitatively evaluated the performance of a wide range of classifiers and features to discriminate hand and finger movements using the same database and methodology to avoid a lot of variability in the recordings and maintain consistency between experiments so that it is possible to compare the results of different algorithms. Various combinations of 44 different features in different domains (time, frequency, and time-frequency) and six classifiers were investigated offline to determine new myoelectric control configurations that improve the classification accuracy of motion recognition with surface EMG signals. Some of the features are commonly used in this area (and were chosen by extensively reviewing the literature), and some of them are new (and I would like to present their results). Among the classifiers, LDA, MLP, K-nearest neighborhood (KNN) and SVM were chosen by reviewing the state of the art; however, there has not been enough investigation on the performance of maximum likelihood estimation (MLE) and regression trees (RegTree) for hand gesture recognition, although they have been extensively used in different areas such as gait measurement and speech recognition. As a result, an efficient feature set and several myoelectric control configurations were proposed to reduce the recognition error in controlling prosthetic hands.

Targeting research question: RQ2 Included papers: Paper E

Contributions 5: Real-time and Offline Evaluation of Recognition Algorithms

The real-time and offline performance of nine classification algorithms in detecting ten individual hand movements were investigated using four channels of surface EMG signals. Some of the investigated algorithms were commonly used in the literature, and some were based on relatively new paradigms in pattern recognition, which were chosen by intensively reviewing the state of the art. The classification accuracy and processing time were used as the evaluation criteria to compare the experimental results in offline and in real time. To conduct this study, I made some modifications and implemented some additional algorithms (i.e., KNN, RegTree,


MLE) to BioPatRec, which is a modular platform implemented in MATLAB [50]. I recorded EMG signals from 15 healthy volunteers while performing ten individual hand movements. In total, each subject performed 12 trials (two for training the subjects (offline and real time), one for the offline training/testing of the classifiers and nine for the real-time testing) to evaluate the nine classification algorithms.

Targeting research question: RQ3 Included papers: Paper F


MLE) to BioPatRec, which is a modular platform implemented in MATLAB [50]. I recorded EMG signals from 15 healthy volunteers while performing ten individual hand movements. In total, each subject performed 12 trials (two for training the subjects (offline and real time), one for the offline training/testing of the classifiers and nine for the real-time testing) to evaluate the nine classification algorithms.

Targeting research question: RQ3 Included papers: Paper F

Chapter 4

Methodology

The research process during this doctoral thesis started with defining several challenges in the field. To understand the research challenges, required information was collected from the related work. Then, several research questions were formulated. To answer the established research questions, the research work was started by collecting and analyzing data using existing methods in the state of the art. The obtained result was analyzed and evaluated using several quantitative and qualitative criteria. The research process was continued by setting up our research design in collecting and analyzing data and interpreting results. Since research is an iterative process, I went through all these steps several times. The research questions were updated, and new approaches were investigated to address the research goals presented in Chapter 3. This chapter provides detailed information about data collection, the employed and proposed methods in this study and the evaluation criteria.

4.1 Filtering Techniques for Removing ECG Interference

4.1.1 Data Acquisition

In this project, I recorded raw EMG signals (signals without ECG interference) from the biceps and deltoid muscles of the right side of several healthy volunteers. In total, ten channels of surface EMG signals were obtained. The reason for choosing these muscles was the hand prosthesis control applications. While recording the EMG signals, the subjects were asked to activate the subjected muscles. Two channels of ECG interference were also recorded from the pectoralis major muscles of the left side. The ECG interferences were added to the raw EMGs to create the contaminated

Chapter 4

Methodology

The research process during this doctoral thesis started with defining several challenges in the field. To understand the research challenges, required information was collected from the related work. Then, several research questions were formulated. To answer the established research questions, the research work was started by collecting and analyzing data using existing methods in the state of the art. The obtained result was analyzed and evaluated using several quantitative and qualitative criteria. The research process was continued by setting up our research design in collecting and analyzing data and interpreting results. Since research is an iterative process, I went through all these steps several times. The research questions were updated, and new approaches were investigated to address the research goals presented in Chapter 3. This chapter provides detailed information about data collection, the employed and proposed methods in this study and the evaluation criteria.

4.1 Filtering Techniques for Removing ECG Interference


In this project, I recorded raw EMG signals (signals without ECG interference) from the biceps and deltoid muscles of the right side of several healthy volunteers. In total, ten channels of surface EMG signals were obtained. The reason for choosing these muscles was the hand prosthesis control applications. While recording the EMG signals, the subjects were asked to activate the subjected muscles. Two channels of ECG interference were also recorded from the pectoralis major muscles of the left side. The ECG interferences were added to the raw EMGs to create the contaminated

16 Chapter 4. Methodology

EMG signals. To be able to use some of the artifact removal methods, such as ANC, ANN and ANFIS, it was necessary to record a one-channel ECG signal as the reference input to the abovementioned methods. Therefore, the ECG signal was recorded from V5 area. While recording the ECG and ECG interferences, the subjects were asked to lay in a completely relaxed position. To record the ECG and EMG signals, a Powerlab/16SP device was used. The raw EMG signals were bandpass filtered from 0.3 to 500 Hz with an analogue filter to reduce the effects of high-frequency noise and avoid aliasing problems. A notch filter (centered at 50 Hz) was also used to remove power line interference from the EMG signals. The signals were recorded with a sampling frequency of 2000 Hz. To remove undesirable motion artifacts, the raw EMG signal was high-pass filtered with a cutoff frequency of 5 Hz, and the direct current (DC) value was removed from the ECG signal and the ECG artifact.

To obtain a quantitative evaluation of the methods, it is necessary that in addition to the contaminated EMG, a corresponding raw EMG signal is also available. Therefore, as is presented in Figure 4.1, the contaminated EMG signal was achieved by multiplying the ECG artifacts (recorded from the pectoralis major muscle of the left side) by a factor (C=0.65) and adding it to the raw (clean) EMG signals (recorded from the biceps and deltoid muscles of the right side).

Figure 4.1: The model to achieve the contaminated EMG signals.

The signal-to-noise ratio (SNR) value of the contaminated EMG signals was considered zero (dB). Considering this initial SNR value helps us to determine how different methods change the desired signal. This model was applied to all signals recorded from five healthy subjects to achieve ten channels of contaminated EMG signals, and finally a 60-second segment of signals was selected from each channel to be processed.

4.1.2 Conventional Filtering Techniques

To investigate the effectiveness of ECG artifact removal methods, different currently used techniques such as HPF, spike clipping, gating, a hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANN, ANC and ANFIS were

Clean EMG

ECG artifact

Contaminated EMG


EMG signals. To be able to use some of the artifact removal methods, such as ANC, ANN and ANFIS, it was necessary to record a one-channel ECG signal as the reference input to the abovementioned methods. Therefore, the ECG signal was recorded from V5 area. While recording the ECG and ECG interferences, the subjects were asked to lay in a completely relaxed position. To record the ECG and EMG signals, a Powerlab/16SP device was used. The raw EMG signals were bandpass filtered from 0.3 to 500 Hz with an analogue filter to reduce the effects of high-frequency noise and avoid aliasing problems. A notch filter (centered at 50 Hz) was also used to remove power line interference from the EMG signals. The signals were recorded with a sampling frequency of 2000 Hz. To remove undesirable motion artifacts, the raw EMG signal was high-pass filtered with a cutoff frequency of 5 Hz, and the direct current (DC) value was removed from the ECG signal and the ECG artifact.

To obtain a quantitative evaluation of the methods, it is necessary that in addition to the contaminated EMG, a corresponding raw EMG signal is also available. Therefore, as is presented in Figure 4.1, the contaminated EMG signal was achieved by multiplying the ECG artifacts (recorded from the pectoralis major muscle of the left side) by a factor (C=0.65) and adding it to the raw (clean) EMG signals (recorded from the biceps and deltoid muscles of the right side).

Figure 4.1: The model to achieve the contaminated EMG signals.

The signal-to-noise ratio (SNR) value of the contaminated EMG signals was considered zero (dB). Considering this initial SNR value helps us to determine how different methods change the desired signal. This model was applied to all signals recorded from five healthy subjects to achieve ten channels of contaminated EMG signals, and finally a 60-second segment of signals was selected from each channel to be processed.


To investigate the effectiveness of ECG artifact removal methods, different currently used techniques such as HPF, spike clipping, gating, a hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANN, ANC and ANFIS were

Clean EMG

ECG artifact

Contaminated EMG

Chapter 4. Methodology 17

evaluated for ECG removal from EMG signals. For a better understanding of the artifact removal methods, a short description of each method is presented below.

4.1.2.1 High-pass Filter

A finite impulse response (FIR) high-pass filter with a Hamming window (length=100), and different cutoff frequencies of 10, 20, 30 and 60 Hz was implemented to remove ECG artifacts from EMG signals [17, 28, 54]. The result with a cutoff frequency of 30 Hz was better than that with the other cutoff frequencies. A large amount of noise remained in the denoised signal when using cutoff frequencies of 10 and 20 Hz, but when a cutoff frequency of 50 Hz was used, a large amount of useful signal was removed.

4.1.2.2 Spike Clipping

As is presented in equation (1), spike clipping relies on a threshold level ( ). If the amplitude of the recorded signal ( ) exceeds this value, the signal is limited in amplitude at this level. is the denoised signal.

(1)

4.1.2.3 Gating Method

The gating method is very similar to the clipping technique, as it requires a threshold to be determined [19]. In this method, if the amplitude of the recorded signal ( ) exceeds this value, the signal is clamped to zero (equation (2)). The performance of this method depends on the thresholding process.

(2)

4.1.2.4 Hybrid Technique

A hybrid technique was employed to remove ECG artifacts from EMG signals [20]. As is presented in Figure 4.2, it combines a spike-clipping algorithm and an HPF. In this approach, the spike-clipping algorithm is based on an adaptive thresholding. The threshold depends on the amplitude of the previous samples and a gain determined by the user. In each iteration of the spike-clipping algorithm, an averaged rectified value (ARV) of the contaminated signal is calculated in a window of N duration. Then, a threshold is set as a gain of the ARV. If any spike is greater than the threshold, it is clipped and set to the cutoff level. The EMG signal in the output of the spike-clipping technique is sent to a fourth-order Butterworth HPF with a cutoff frequency of 30 H.


evaluated for ECG removal from EMG signals. For a better understanding of the artifact removal methods, a short description of each method is presented below.

4.1.2.1 High-pass Filter

A finite impulse response (FIR) high-pass filter with a Hamming window (length=100), and different cutoff frequencies of 10, 20, 30 and 60 Hz was implemented to remove ECG artifacts from EMG signals [17, 28, 54]. The result with a cutoff frequency of 30 Hz was better than that with the other cutoff frequencies. A large amount of noise remained in the denoised signal when using cutoff frequencies of 10 and 20 Hz, but when a cutoff frequency of 50 Hz was used, a large amount of useful signal was removed.


As is presented in equation (1), spike clipping relies on a threshold level ( ). If the amplitude of the recorded signal ( ) exceeds this value, the signal is limited in amplitude at this level. is the denoised signal.

(1)


The gating method is very similar to the clipping technique, as it requires a threshold to be determined [19]. In this method, if the amplitude of the recorded signal ( ) exceeds this value, the signal is clamped to zero (equation (2)). The performance of this method depends on the thresholding process.

(2)


A hybrid technique was employed to remove ECG artifacts from EMG signals [20]. As is presented in Figure 4.2, it combines a spike-clipping algorithm and an HPF. In this approach, the spike-clipping algorithm is based on an adaptive thresholding. The threshold depends on the amplitude of the previous samples and a gain determined by the user. In each iteration of the spike-clipping algorithm, an averaged rectified value (ARV) of the contaminated signal is calculated in a window of N duration. Then, a threshold is set as a gain of the ARV. If any spike is greater than the threshold, it is clipped and set to the cutoff level. The EMG signal in the output of the spike-clipping technique is sent to a fourth-order Butterworth HPF with a cutoff frequency of 30 H.


The evaluation of the spike-clipping algorithm in each iteration requires a high computation time, which is one of the drawbacks of this method.

Figure 4.2: The structure of the hybrid technique.

4.1.2.5 Template Subtraction

The template subtraction approach consists of two stages: the first stage involves creating an ECG template, and the second stage involves subtracting the created template from the contaminated EMG signal. In the first stage, the accurate detection of the ECG beats from the EMG signal is performed based on the R wave detection algorithm [55-57]. After detecting the R waves, a frame of some samples around this peak is selected for accommodating one QRS complex in each frame. Using the mean of the selected QRS complexes, the ECG template is obtained. In the second stage, a denoised EMG signal is obtained by subtracting the ECG template from the contaminated EMG, where the QRS complex has been manifested by the R peak detection algorithm [21, 28, 54].

4.1.2.6 Independent Component Analysis

Independent component analysis extracts statistically independent components from a set of measured signals [6]. This method finds a linear representation of non-Gaussian data. When n linear mixtures of m independent components

are observed, ICA can be used to estimate the mixing matrix,

based on the information of their independence, which allows us to separate the original source signals from their mixtures. To use the ICA method, the model in equation (3) is considered:

(3)

where is a linear mixture, is an independent source signal and is a full-rank mixing matrix. After finding the noisy component in ICA-based methods, the noisy component is usually set to zero [58]. This procedure removes useful information from the reconstructed signal. Therefore, it is preferred to use an HPF to remove ECG artifacts from the selected component [7]. Finally, inverse ICA is applied to reconstruct the denoised EMG signal.

Contaminated EMG

Adaptive spike clipping

High pass filter Denoised EMG


The evaluation of the spike-clipping algorithm in each iteration requires a high computation time, which is one of the drawbacks of this method.

Figure 4.2: The structure of the hybrid technique.


The template subtraction approach consists of two stages: the first stage involves creating an ECG template, and the second stage involves subtracting the created template from the contaminated EMG signal. In the first stage, the accurate detection of the ECG beats from the EMG signal is performed based on the R wave detection algorithm [55-57]. After detecting the R waves, a frame of some samples around this peak is selected for accommodating one QRS complex in each frame. Using the mean of the selected QRS complexes, the ECG template is obtained. In the second stage, a denoised EMG signal is obtained by subtracting the ECG template from the contaminated EMG, where the QRS complex has been manifested by the R peak detection algorithm [21, 28, 54].


Independent component analysis extracts statistically independent components from a set of measured signals [6]. This method finds a linear representation of non-Gaussian data. When n linear mixtures of m independent components

are observed, ICA can be used to estimate the mixing matrix,

based on the information of their independence, which allows us to separate the original source signals from their mixtures. To use the ICA method, the model in equation (3) is considered:

(3)

where is a linear mixture, is an independent source signal and is a full-rank mixing matrix. After finding the noisy component in ICA-based methods, the noisy component is usually set to zero [58]. This procedure removes useful information from the reconstructed signal. Therefore, it is preferred to use an HPF to remove ECG artifacts from the selected component [7]. Finally, inverse ICA is applied to reconstruct the denoised EMG signal.

Contaminated EMG

Adaptive spike clipping

High pass filter Denoised EMG


Some fundamental assumptions of ICA are (1) the number of mixtures is greater than or equal to the number of independent sources . (2) The mixtures are linear combinations of the sources, and there is no delay or external noise included. (3) The sources are stationary and are not moving during the recording process [22]. ICA has three shortcomings: (1) amplitude ambiguity, (2) source ambiguity and (3) the need for the number of recordings to be equal to or greater than the number of sources [22].

4.1.2.7 Wavelet Transform

Wavelet transform is a time-frequency representation of a signal that has been introduced to overcome the limitations in time and frequency resolutions of the classical Fourier transform. Instead of using a sine wave, which is the basic function for the Fourier transform, a basic waveform ( ) is used, which can be modified to

basic functions ( ) in equation (4), obtained from dilations and shifts of the basic

waveform [38, 59]. is a scaling parameter, and represents a translation parameter [38].

(4)

The wavelet decomposes a signal into different frequency components and studies each component with a resolution matched to its scale [60]. This property can be used for denoising purposes [38]. The performance of the wavelet transform depends on the type of wavelet transform, type of wavelet thresholding rule and number of decomposition levels. In the denoising process, the corrupted EMG signal is first decomposed by a wavelet transform to create a wavelet decomposition of the raw signal representing different frequency components of the signal [25, 38]. The wavelet coefficients in the low-frequency scales then underwent a nonlinear thresholding process, where the absolute value of the coefficients greater than the threshold is set to zero. The inverse wavelet transform is then implemented using the new coefficients to obtain the denoised EMG signal [25, 38].

4.1.2.8 Combined Wavelet and ICA

A new approach for an artifact removal process is to use single-channel ICA [61]. Empirical mode decomposition (EMD) and the wavelet transform are approaches that could enable the ICA method to be used in a single-channel analysis. The combined wavelet-ICA method could successfully remove the ECG artifacts [28]. However, EMD has a high computational cost, which is one of the drawbacks of this method [39]. The idea behind the combined wavelet transform and ICA method is to


Some fundamental assumptions of ICA are (1) the number of mixtures is greater than or equal to the number of independent sources . (2) The mixtures are linear combinations of the sources, and there is no delay or external noise included. (3) The sources are stationary and are not moving during the recording process [22]. ICA has three shortcomings: (1) amplitude ambiguity, (2) source ambiguity and (3) the need for the number of recordings to be equal to or greater than the number of sources [22].


Wavelet transform is a time-frequency representation of a signal that has been introduced to overcome the limitations in time and frequency resolutions of the classical Fourier transform. Instead of using a sine wave, which is the basic function for the Fourier transform, a basic waveform ( ) is used, which can be modified to

basic functions ( ) in equation (4), obtained from dilations and shifts of the basic

waveform [38, 59]. is a scaling parameter, and represents a translation parameter [38].

(4)

The wavelet decomposes a signal into different frequency components and studies each component with a resolution matched to its scale [60]. This property can be used for denoising purposes [38]. The performance of the wavelet transform depends on the type of wavelet transform, type of wavelet thresholding rule and number of decomposition levels. In the denoising process, the corrupted EMG signal is first decomposed by a wavelet transform to create a wavelet decomposition of the raw signal representing different frequency components of the signal [25, 38]. The wavelet coefficients in the low-frequency scales then underwent a nonlinear thresholding process, where the absolute value of the coefficients greater than the threshold is set to zero. The inverse wavelet transform is then implemented using the new coefficients to obtain the denoised EMG signal [25, 38].


A new approach for an artifact removal process is to use single-channel ICA [61]. Empirical mode decomposition (EMD) and the wavelet transform are approaches that could enable the ICA method to be used in a single-channel analysis. The combined wavelet-ICA method could successfully remove the ECG artifacts [28]. However, EMD has a high computational cost, which is one of the drawbacks of this method [39]. The idea behind the combined wavelet transform and ICA method is to


decompose single-channel data into different components using the wavelet transform before applying the ICA technique. As is presented in Figure 4.3, in this method, a discrete wavelet transform is applied to a single-channel recording to create a wavelet decomposition of the raw signal. After calculating the wavelet coefficients, ICA is used to extract independent components from the multidimensional data produced by the wavelet transform. After finding the noisy component manually, the noisy component is set to zero [58].

Figure 4.3: The flowchart of the wavelet-ICA method.

4.1.2.9 Adaptive Noise Canceller

An adaptive noise canceller with recursive least squares (RLS) algorithm was used to adjust the amplitude and phase of the reference signal against the artifact signal and subtract it from the primary EMG signal [62-64]. Unlike the conventional bandpass filter, which fixes filter weights, the adaptive filter has adjustable weights ( in Figure 4.4) that are iteratively updated based on the characteristics of two input signals (the contaminated EMG and the reference ECG) [31].

Figure 4.4: The structure of the adaptive noise canceller.

In separating the EMG signal from the ECG artifact using an ANC, the contaminated EMG signal and ECG are the primary input and reference input, respectively. The

structure of the adaptive filter is presented in Figure 4.4, where is one sample

Single ChannelEMG signal

DenoisedEMG signalICAWavelet

Transform

InverseWavelet and ICA

Detecting noisy component manually Zero

( )y k

( )n k 1z 1z 1z

1kw 2kw Mkw

ˆ( )e k

ˆ( )s k


decompose single-channel data into different components using the wavelet transform before applying the ICA technique. As is presented in Figure 4.3, in this method, a discrete wavelet transform is applied to a single-channel recording to create a wavelet decomposition of the raw signal. After calculating the wavelet coefficients, ICA is used to extract independent components from the multidimensional data produced by the wavelet transform. After finding the noisy component manually, the noisy component is set to zero [58].

Figure 4.3: The flowchart of the wavelet-ICA method.


An adaptive noise canceller with recursive least squares (RLS) algorithm was used to adjust the amplitude and phase of the reference signal against the artifact signal and subtract it from the primary EMG signal [62-64]. Unlike the conventional bandpass filter, which fixes filter weights, the adaptive filter has adjustable weights ( in Figure 4.4) that are iteratively updated based on the characteristics of two input signals (the contaminated EMG and the reference ECG) [31].

Figure 4.4: The structure of the adaptive noise canceller.

In separating the EMG signal from the ECG artifact using an ANC, the contaminated EMG signal and ECG are the primary input and reference input, respectively. The

structure of the adaptive filter is presented in Figure 4.4, where is one sample



Transform


Detecting noisy component manually Zero

( )y k

( )n k 1z 1z 1z

1kw 2kw Mkw

ˆ( )e k

ˆ( )s k


delay, is a time instant, is the input to the ANC (contaminated EMG), is the reference input, is estimated noise that can be calculated using equation (5) and is the denoised EMG signal obtained from equation (6).

(5)

(6)

4.1.2.10 Artificial Neural Network

An artificial neural network with a back propagation network (BPN), which belongs to the category of nonlinear adaptive systems, has been proposed to estimate an unknown interference present in an EMG signal [29] (see Figure 4.5). A BPN is a feed forward, multilayer network that uses the supervised mode of learning. It uses a gradient descent algorithm to minimize the interference. The BPN architecture consists of an input layer, hidden layers, and output layer. The numbers of hidden layers and neurons in each layer depends on the application of the BPN [29].

Figure 4.5: Flowchart of the processing stages of removing ECG interferences from an EMG using an ANN.

To implement the BPN, the following steps are used: Specifying (1) the inputs and targets to the BPN, (2) the minimum and maximum values of the input ranges, (3) the number of layers and the number of units in each layer, (4) the activation functions for each layer, and (5) the training parameters such as the learning rate, momentum, performance goal, and number of epochs; (6) Creating a feed-forward back propagation network model by using the command called NEWFF; (7) simulating the

Contaminated EMG

ANN

Noise Estimated

Denoised EMG

ECG

D

+ -


delay, is a time instant, is the input to the ANC (contaminated EMG), is the reference input, is estimated noise that can be calculated using equation (5) and is the denoised EMG signal obtained from equation (6).

(5)

(6)


An artificial neural network with a back propagation network (BPN), which belongs to the category of nonlinear adaptive systems, has been proposed to estimate an unknown interference present in an EMG signal [29] (see Figure 4.5). A BPN is a feed forward, multilayer network that uses the supervised mode of learning. It uses a gradient descent algorithm to minimize the interference. The BPN architecture consists of an input layer, hidden layers, and output layer. The numbers of hidden layers and neurons in each layer depends on the application of the BPN [29].

Figure 4.5: Flowchart of the processing stages of removing ECG interferences from an EMG using an ANN.

To implement the BPN, the following steps are used: Specifying (1) the inputs and targets to the BPN, (2) the minimum and maximum values of the input ranges, (3) the number of layers and the number of units in each layer, (4) the activation functions for each layer, and (5) the training parameters such as the learning rate, momentum, performance goal, and number of epochs; (6) Creating a feed-forward back propagation network model by using the command called NEWFF; (7) simulating the

Contaminated EMG

ANN

Noise Estimated

Denoised EMG

ECG

D

+ -


network for plotting the network output; (8) training the network using the training function called TRAINLM; and (9) giving the conditions to stop training the network (the maximum number of epochs is reached or the maximum amount of time has been exceeded, etc.) [29].

4.1.2.11 Adaptive Neuro-Fuzzy Inference System

Adaptive noise cancellation using a linear filter was used successfully to remove ECG interference. However, it is clear that the idea of linear adaptive noise cancellation can be extended to nonlinear forms using nonlinear adaptive systems [26]. One of these nonlinear systems is ANFIS, which has been used to estimate noise [29, 35]. ANFIS is a network that combines the advantages of a neural network and a fuzzy system [65]. The general structure of ANFIS is with two inputs, one output and five hidden layers [35]. The first layer acts as a fuzzifier, the second layer represents the fuzzy rules, the third layer is used for the normalization of the membership functions, the fourth layer is used to identify effective parameters, and the fifth layer computes the output of the fuzzy system (defuzzification process) [49]. In this method, the noisy signal is applied as a target. An ECG signal and a delayed ECG signal are applied as inputs to the ANFIS. Collecting data (the ECG signal and the ECG artifact) from different sources causes a delay between recorded signals [28]; hence, to improve the noise estimation process, the delayed ECG signal is used as one of inputs to the ANFIS.

4.1.3 Proposed Novel Filtering Techniques

After evaluating the aforementioned methods using quantitative and qualitative criteria such as the SNR, relative error (RE), correlation coefficients (CC), elapsed time (ET) and PSD, new techniques (ANN-wavelet, ANFIS-wavelet, adaptive subtraction and automated wavelet-ICA) have been proposed to improve the ECG artifact removal process. These techniques are presented in papers A, B, C and D.

4.1.3.1 ANFIS-Wavelet (Paper A)

In the first step, a contaminated EMG and ECG signal were used as the input and the reference to the ANFIS, respectively, to remove ECG interference from the EMG signal. It was observed that low-frequency noise remained in the denoised signal. Since the wavelet transforms a signal into different frequency bands using different high-pass and low-pass filters, this feature of the wavelet transform was employed to find the low-frequency noise in the denoised EMG and remove it using a nonlinear thresholding approach.


network for plotting the network output; (8) training the network using the training function called TRAINLM; and (9) giving the conditions to stop training the network (the maximum number of epochs is reached or the maximum amount of time has been exceeded, etc.) [29].


Adaptive noise cancellation using a linear filter was used successfully to remove ECG interference. However, it is clear that the idea of linear adaptive noise cancellation can be extended to nonlinear forms using nonlinear adaptive systems [26]. One of these nonlinear systems is ANFIS, which has been used to estimate noise [29, 35]. ANFIS is a network that combines the advantages of a neural network and a fuzzy system [65]. The general structure of ANFIS is with two inputs, one output and five hidden layers [35]. The first layer acts as a fuzzifier, the second layer represents the fuzzy rules, the third layer is used for the normalization of the membership functions, the fourth layer is used to identify effective parameters, and the fifth layer computes the output of the fuzzy system (defuzzification process) [49]. In this method, the noisy signal is applied as a target. An ECG signal and a delayed ECG signal are applied as inputs to the ANFIS. Collecting data (the ECG signal and the ECG artifact) from different sources causes a delay between recorded signals [28]; hence, to improve the noise estimation process, the delayed ECG signal is used as one of inputs to the ANFIS.


After evaluating the aforementioned methods using quantitative and qualitative criteria such as the SNR, relative error (RE), correlation coefficients (CC), elapsed time (ET) and PSD, new techniques (ANN-wavelet, ANFIS-wavelet, adaptive subtraction and automated wavelet-ICA) have been proposed to improve the ECG artifact removal process. These techniques are presented in papers A, B, C and D.


In the first step, a contaminated EMG and ECG signal were used as the input and the reference to the ANFIS, respectively, to remove ECG interference from the EMG signal. It was observed that low-frequency noise remained in the denoised signal. Since the wavelet transforms a signal into different frequency bands using different high-pass and low-pass filters, this feature of the wavelet transform was employed to find the low-frequency noise in the denoised EMG and remove it using a nonlinear thresholding approach.


4.1.3.2 ANN-Wavelet (Paper B)

The procedure here is the same as ANFIS-wavelet, but instead of ANFIS, ANN was first employed for removing the ECG artifact from the contaminated EMG signal. In this step, a large amount of pollution was removed; however, the result shows that there are still low-frequency noise components in the denoised EMG signal. A wavelet transform with nonlinear thresholding was used to remove the residual noise from the output of the ANN method. Using this method, the low-frequency components (noise) remaining at the output of the neural network were deleted.

4.1.3.3 Adaptive Subtraction (Paper C)

This method contains of four steps; (1) QRS detection in the contaminated EMG signals, (2) formation of an ECG template by averaging the electrocardiogram complexes, (3) using a low-pass filter to remove undesirable artifacts from the ECG template, and (4) subtracting the template from the contaminated EMG signal where the QRS complexes are detected. The proposed technique is an online method that is somewhat effective in removing ECG artifacts.

4.1.3.4 Automated Wavelet-ICA (Paper D)

A contaminated EMG is first decomposed using a wavelet transform to provide a multichannel input to the ICA. In conventional ICA-based methods, after creating the ICA components, independent components are classified manually as an EMG signal or ECG artifact. In this proposed approach, an automated algorithm is used to determine the ECG artifact components automatically. Each channel of independent components is evaluated separately, and the component with an ECG artifact is selected automatically. After finding the noisy component in ICA-based methods, the noisy component usually is set to zero. This procedure removes useful information from the reconstructed signal. Therefore, an FIR high-pass filter, a Hamming window, order 100 with a cutoff frequency of 20 Hz, is used to remove ECG artifacts from the selected noisy component. Finally, an inverse ICA and then an inverse wavelet transform are applied to reconstruct the denoised EMG signal (See Figure 4.6).

Figure 4.6: The flowchart of the automated wavelet-ICA method.



Transform


Detecting noisy component automatically HPF



The procedure here is the same as ANFIS-wavelet, but instead of ANFIS, ANN was first employed for removing the ECG artifact from the contaminated EMG signal. In this step, a large amount of pollution was removed; however, the result shows that there are still low-frequency noise components in the denoised EMG signal. A wavelet transform with nonlinear thresholding was used to remove the residual noise from the output of the ANN method. Using this method, the low-frequency components (noise) remaining at the output of the neural network were deleted.


This method contains of four steps; (1) QRS detection in the contaminated EMG signals, (2) formation of an ECG template by averaging the electrocardiogram complexes, (3) using a low-pass filter to remove undesirable artifacts from the ECG template, and (4) subtracting the template from the contaminated EMG signal where the QRS complexes are detected. The proposed technique is an online method that is somewhat effective in removing ECG artifacts.


A contaminated EMG is first decomposed using a wavelet transform to provide a multichannel input to the ICA. In conventional ICA-based methods, after creating the ICA components, independent components are classified manually as an EMG signal or ECG artifact. In this proposed approach, an automated algorithm is used to determine the ECG artifact components automatically. Each channel of independent components is evaluated separately, and the component with an ECG artifact is selected automatically. After finding the noisy component in ICA-based methods, the noisy component usually is set to zero. This procedure removes useful information from the reconstructed signal. Therefore, an FIR high-pass filter, a Hamming window, order 100 with a cutoff frequency of 20 Hz, is used to remove ECG artifacts from the selected noisy component. Finally, an inverse ICA and then an inverse wavelet transform are applied to reconstruct the denoised EMG signal (See Figure 4.6).

Figure 4.6: The flowchart of the automated wavelet-ICA method.



Transform


Detecting noisy component automatically HPF


4.2 Offline Pattern Recognition (Paper E)


The required signals in this study (four-channel surface EMG signals from 20 healthy subjects) were obtained from the BioPatRec database [50]. The 11 hand gestures available in the selected database were open hand, closed hand, flex hand, extend hand, pronation, supination, side grip, fine grip, agree or thumb up, pointer or index extension, and relaxation.

4.2.2 Data Analysis

In this study, a total of 44 individual features (see Table 4.1) in the time, frequency, and time-frequency domains (25, 8 and 11, respectively) were extracted from surface EMG signals. Some of the features are commonly used in this area (and were chosen by extensively reviewing the literature), and some of them are new. Various combinations of the 44 individual features and six commonly used classifiers (LDA, KNN, RegTree, MLE, SVM, and MLP) were investigated to decode eleven individual movements. LDA, KNN, SVM and MLP were chosen by reviewing the state of the art; however, there is not enough investigation on RegTree and MLE’s performance on hand gesture recognition. In this study, in addition to investigating each individual feature in combination with the classifiers, 25 features in the TD, eight features in the FD and 11 features in the TFD were investigated as three different feature sets. Then, principal component analysis (PCA) was employed (because of its popularity in the state of the art) to reduce the dimensionality of the aforementioned feature sets. The resulting feature vectors were then fed into the classifiers (see Figure 4.7). I examined the classification accuracy and processing time, which are crucial for performance evaluations.

Afterwards, the accuracy rates of different combinations of the 25 TD features were determined by adding individual TD features one by one to the feature set from number one to number 25. If an added feature produced a higher accuracy rate, it would stay in the feature set; otherwise, it would be removed. Once all features were evaluated, the features in the obtained feature set were removed in inverted order if their subtraction did not negatively affect the accuracy.


4.2 Offline Pattern Recognition (Paper E)


The required signals in this study (four-channel surface EMG signals from 20 healthy subjects) were obtained from the BioPatRec database [50]. The 11 hand gestures available in the selected database were open hand, closed hand, flex hand, extend hand, pronation, supination, side grip, fine grip, agree or thumb up, pointer or index extension, and relaxation.

4.2.2 Data Analysis

In this study, a total of 44 individual features (see Table 4.1) in the time, frequency, and time-frequency domains (25, 8 and 11, respectively) were extracted from surface EMG signals. Some of the features are commonly used in this area (and were chosen by extensively reviewing the literature), and some of them are new. Various combinations of the 44 individual features and six commonly used classifiers (LDA, KNN, RegTree, MLE, SVM, and MLP) were investigated to decode eleven individual movements. LDA, KNN, SVM and MLP were chosen by reviewing the state of the art; however, there is not enough investigation on RegTree and MLE’s performance on hand gesture recognition. In this study, in addition to investigating each individual feature in combination with the classifiers, 25 features in the TD, eight features in the FD and 11 features in the TFD were investigated as three different feature sets. Then, principal component analysis (PCA) was employed (because of its popularity in the state of the art) to reduce the dimensionality of the aforementioned feature sets. The resulting feature vectors were then fed into the classifiers (see Figure 4.7). I examined the classification accuracy and processing time, which are crucial for performance evaluations.

Afterwards, the accuracy rates of different combinations of the 25 TD features were determined by adding individual TD features one by one to the feature set from number one to number 25. If an added feature produced a higher accuracy rate, it would stay in the feature set; otherwise, it would be removed. Once all features were evaluated, the features in the obtained feature set were removed in inverted order if their subtraction did not negatively affect the accuracy.


Figure 4.7: The processing stages of individual hand movement recognition for identifying new myoelectric control configurations.

I went through the same procedure for the eight FD and 11 TFD features. Finally, an efficient feature set (FS: waveform length, correlation coefficient and Hjorth parameters (activity, mobility and complexity)) was obtained to improve the motion recognition accuracy. To compare the results, I applied all the classifiers to a well-known feature set (Hudgins set of five time domain features). To show the significant differences between features, I conducted a statistical analysis.

Table 4.1: The forty-four extracted features (their names, abbreviations, descriptions and publication references) in the time, frequency and time-frequency domains from the collected EMG signals.

Time domain features Abbreviation Description References

1 Mean absolute value MAV Calculated by averaging the absolute value of a given signal in a window

[40]

2 Standard deviation STD Represents the difference between each sample of a signal and its mean value

[66]

3 Variance = Hjorth activity parameter

Var Represents the power of a signal [16, 67]

4 Waveform length WL Calculated by the sum of the absolute values of the difference between each sample of a signal

[68]

5 Zero crossing ZC Counts the number of times that the sign of the amplitude of the signal changes

[69]

6 Number of peaks NP Number of peak values that are higher than a signal’s RMS value

Paper E

7 Mean of peak values MPV Average of the peak values that are higher than a given signal’s RMS value

[70]


Preprocessing Feature Extraction

•  TD features •  FD features •  TFD features

•  LDA •  KNN •  RegTree •  MLE •  SVM •  MLP

Classification

Class Labels Dimensionality Reduction

Applied to Some Feature Sets

•  PCA •  Feature Selection

•  Filtering •  Windowing


Figure 4.7: The processing stages of individual hand movement recognition for identifying new myoelectric control configurations.

I went through the same procedure for the eight FD and 11 TFD features. Finally, an efficient feature set (FS: waveform length, correlation coefficient and Hjorth parameters (activity, mobility and complexity)) was obtained to improve the motion recognition accuracy. To compare the results, I applied all the classifiers to a well-known feature set (Hudgins set of five time domain features). To show the significant differences between features, I conducted a statistical analysis.

Table 4.1: The forty-four extracted features (their names, abbreviations, descriptions and publication references) in the time, frequency and time-frequency domains from the collected EMG signals.

Time domain features Abbreviation Description References

1 Mean absolute value MAV Calculated by averaging the absolute value of a given signal in a window

[40]

2 Standard deviation STD Represents the difference between each sample of a signal and its mean value

[66]

3 Variance = Hjorth activity parameter

Var Represents the power of a signal [16, 67]

4 Waveform length WL Calculated by the sum of the absolute values of the difference between each sample of a signal

[68]

5 Zero crossing ZC Counts the number of times that the sign of the amplitude of the signal changes

[69]

6 Number of peaks NP Number of peak values that are higher than a signal’s RMS value

Paper E

7 Mean of peak values MPV Average of the peak values that are higher than a given signal’s RMS value

[70]


Preprocessing Feature Extraction

•  TD features •  FD features •  TFD features

•  LDA •  KNN •  RegTree •  MLE •  SVM •  MLP

Classification

Class Labels Dimensionality Reduction

Applied to Some Feature Sets

•  PCA •  Feature Selection

•  Filtering •  Windowing


8 Mean firing velocity MFV Samples of a signal corresponding to the peak values higher than its RMS value

[70]

9 Slope sign changes SSC Number of times the slope of a signal in a time window changes sign

[14, 53]

10 Correlation coefficient CC Represents the linear relationship between two samples of a signal

[71]

11 Difference absolute mean value

DAMV Calculated by averaging the absolute value of difference between each sample of a signal

[72]

12 Fractal dimension FDim Measures the strength of muscle activity

[67]

13 Maximum fractal length MFL Measures the strength of muscle contraction

[67, 73]

14 Higuchi’s fractal dimension HFD A popular technique to calculate the fractal dimension

[73]

15 Skewness SkewDescribes asymmetry in a statistical distribution around the mean value

[74]

16 Integrated absolute value IAV Calculated by the summation of the absolute values of a signal in a time window

[75]

17 Hjorth mobility parameter HMob Represents the proportion of standard deviation of the power spectrum

[76, 77]

18 Hjorth complexity parameter

HCom Compares the similarity of the shape of a signal with a pure sine wave

[76, 77]

19 Multichannel energy ratio ER Describes the energy distribution in multiple channels

[71]

20 Difference absolute standard deviation value

DASDV Represents the STD value of the difference between each samples of a signal

[72, 78]

21 Willison amplitude WAM

The number of times the absolute value of the difference between two adjacent samples in a time window exceeds a predefined threshold

[16]

22 Mean absolute value slope MAVS The difference between the mean absolute values of adjacent time windows of a signal

[40, 68]

23 Kurtosis KurtCompares the shape of a statistical distribution with the normal distribution

[74]

24 Percentile Perc

Calculated by sorting all samples in a time window of a signal in ascending order and computing the position L = (K/100) * N, where N is the total

[74]


8 Mean firing velocity MFV Samples of a signal corresponding to the peak values higher than its RMS value

[70]

9 Slope sign changes SSC Number of times the slope of a signal in a time window changes sign

[14, 53]

10 Correlation coefficient CC Represents the linear relationship between two samples of a signal

[71]

11 Difference absolute mean value

DAMV Calculated by averaging the absolute value of difference between each sample of a signal

[72]

12 Fractal dimension FDim Measures the strength of muscle activity

[67]

13 Maximum fractal length MFL Measures the strength of muscle contraction

[67, 73]

14 Higuchi’s fractal dimension HFD A popular technique to calculate the fractal dimension

[73]

15 Skewness SkewDescribes asymmetry in a statistical distribution around the mean value

[74]

16 Integrated absolute value IAV Calculated by the summation of the absolute values of a signal in a time window

[75]

17 Hjorth mobility parameter HMob Represents the proportion of standard deviation of the power spectrum

[76, 77]

18 Hjorth complexity parameter

HCom Compares the similarity of the shape of a signal with a pure sine wave

[76, 77]

19 Multichannel energy ratio ER Describes the energy distribution in multiple channels

[71]

20 Difference absolute standard deviation value

DASDV Represents the STD value of the difference between each samples of a signal

[72, 78]

21 Willison amplitude WAM

The number of times the absolute value of the difference between two adjacent samples in a time window exceeds a predefined threshold

[16]

22 Mean absolute value slope MAVS The difference between the mean absolute values of adjacent time windows of a signal

[40, 68]

23 Kurtosis Kurt Compares the shape of a statistical distribution with the normal distribution

[74]

24 Percentile Perc

Calculated by sorting all samples in a time window of a signal in ascending order and computing the position L = (K/100) * N, where N is the total

[74]


number of samples in the time window and K is the K-th Percentile

25 Histogram Hist

Calculated by sorting all samples in a time window of a signal in descending order, segmenting the sorted values into several equally spaced frames and returning the number of samples in each segment

[69]

Frequency domain features Abbreviation Description References

26 Waveform length WL Calculating the WL of the fast Fourier transform of a signal in a time window

[70]

27 Mean frequency MNF

Calculated as the sum of the power spectrum of a signal multiplied by the frequency variable divided by the sum of the power spectrum of the signal

[14]

28 Median frequency MDF Is a frequency at which the EMG power spectrum is divided into two parts of equal amplitude

[14]

29 Mean of peaks MPK Average of peak values of a signal in frequency domain that are higher than a given signal’s RMS value

Paper E

30 Standard deviation of peaks STDPK Calculating the STD of the peak values of a signal in the frequency domain

Paper E

31 Frequency ratio FR Defined as the ratio of the power spectra at the low-frequency band and high-frequency band

[40]

32 Peak frequency PKF Is the frequency of the maximum power

[79]

33 Frequency energy FE Calculated by the sum of the squared amplitudes of the FFT of a signal

[12]

Time-frequency domain features

Abbreviation Description References

34 Standard deviation STD Calculating STD of wavelet coefficients in the last level

[80]

35 Variance Var Calculating the variance of wavelet coefficients in the last level

[80]

36 Waveform length WL Calculating WL of wavelet coefficients in the last level

Paper E

37 Energy Er Calculating the energy of wavelet coefficients in the last level

[81, 82]

38 Maximum absolute value MaxAV MaxAV of wavelet coefficients in the last level was calculated

[83]


number of samples in the time window and K is the K-th Percentile

25 Histogram Hist

Calculated by sorting all samples in a time window of a signal in descending order, segmenting the sorted values into several equally spaced frames and returning the number of samples in each segment

[69]

Frequency domain features Abbreviation Description References

26 Waveform length WL Calculating the WL of the fast Fourier transform of a signal in a time window

[70]

27 Mean frequency MNF

Calculated as the sum of the power spectrum of a signal multiplied by the frequency variable divided by the sum of the power spectrum of the signal

[14]

28 Median frequency MDF Is a frequency at which the EMG power spectrum is divided into two parts of equal amplitude

[14]

29 Mean of peaks MPK Average of peak values of a signal in frequency domain that are higher than a given signal’s RMS value

Paper E

30 Standard deviation of peaks STDPK Calculating the STD of the peak values of a signal in the frequency domain

Paper E

31 Frequency ratio FR Defined as the ratio of the power spectra at the low-frequency band and high-frequency band

[40]

32 Peak frequency PKF Is the frequency of the maximum power

[79]

33 Frequency energy FE Calculated by the sum of the squared amplitudes of the FFT of a signal

[12]

Time-frequency domain features

Abbreviation Description References

34 Standard deviation STD Calculating STD of wavelet coefficients in the last level

[80]

35 Variance Var Calculating the variance of wavelet coefficients in the last level

[80]

36 Waveform length WL Calculating WL of wavelet coefficients in the last level

Paper E

37 Energy Er Calculating the energy of wavelet coefficients in the last level

[81, 82]

38 Maximum absolute value MaxAV MaxAV of wavelet coefficients in the last level was calculated

[83]


39 Zero crossing ZC ZC of wavelet coefficients in the last level was calculated

[81, 84]

40 Mean Mean Mean of wavelet coefficients in the last level was calculated

Paper E

41 Mean absolute value MAV MAV of wavelet coefficients in the last level was calculated

Paper E

42 Logarithmic RMS LogRMS LogRMS of wavelet packet coefficient in the last subspace was calculated

[85]

43 Relative energy RE Calculated by total energy of a signal over energy in each subspace of wavelet packet transform

[86]

44 Normalized logarithmic energy

NLE

Calculated by applying the logarithmic operator to the accumulation of the squares of the wavelet packet coefficients divided by the number of coefficients in the subspace

[85]

4.3 Real-Time Pattern Recognition (Paper F)


To conduct this study, I recorded four channels of surface EMG signals representative of ten classes of motions from fifteen healthy subjects. The motion classes consisted of hand open/close, wrist flexion/extension, forearm pronation/supination, side grip, fine grip, agree and pointer. These movements were selected as they could be feasible in high-end commercial prostheses.

Eight disposable Ag/AgCl electrodes were placed equally distributed around the forearm in a bipolar configuration. Data were acquired using an open-source bioelectric signal acquisition system (ADS_BP) [87]. During the recording sessions, subjects were instructed to perform the ten selected movements. The data acquisition, offline analysis and real-time testing were done using BioPatRec software, which is a modular platform implemented in MATLAB [50]. Some modifications and implementations were made to BioPatRec to perform this study. This software ran on MATLAB R2014a.

4.3.2 Data Analysis

The recorded signals were filtered using a high-pass filter with a cutoff frequency of 20 Hz and a low-pass filter with a cutoff frequency of 500 Hz and a Notch filter at 50 Hz. After windowing (with a 200 ms length and 50 ms time increment) the signals,


39 Zero crossing ZC ZC of wavelet coefficients in the last level was calculated

[81, 84]

40 Mean Mean Mean of wavelet coefficients in the last level was calculated

Paper E

41 Mean absolute value MAV MAV of wavelet coefficients in the last level was calculated

Paper E

42 Logarithmic RMS LogRMS LogRMS of wavelet packet coefficient in the last subspace was calculated

[85]

43 Relative energy RE Calculated by total energy of a signal over energy in each subspace of wavelet packet transform

[86]

44 Normalized logarithmic energy

NLE

Calculated by applying the logarithmic operator to the accumulation of the squares of the wavelet packet coefficients divided by the number of coefficients in the subspace

[85]

4.3 Real-Time Pattern Recognition (Paper F)


To conduct this study, I recorded four channels of surface EMG signals representative of ten classes of motions from fifteen healthy subjects. The motion classes consisted of hand open/close, wrist flexion/extension, forearm pronation/supination, side grip, fine grip, agree and pointer. These movements were selected as they could be feasible in high-end commercial prostheses.

Eight disposable Ag/AgCl electrodes were placed equally distributed around the forearm in a bipolar configuration. Data were acquired using an open-source bioelectric signal acquisition system (ADS_BP) [87]. During the recording sessions, subjects were instructed to perform the ten selected movements. The data acquisition, offline analysis and real-time testing were done using BioPatRec software, which is a modular platform implemented in MATLAB [50]. Some modifications and implementations were made to BioPatRec to perform this study. This software ran on MATLAB R2014a.

4.3.2 Data Analysis

The recorded signals were filtered using a high-pass filter with a cutoff frequency of 20 Hz and a low-pass filter with a cutoff frequency of 500 Hz and a Notch filter at 50 Hz. After windowing (with a 200 ms length and 50 ms time increment) the signals,


the most commonly used set of features (four TD features reported by Hudgins et al. [43]: MAV, ZC, SSC, WL) was extracted from each window. Nine classification algorithms (see Table 4.2) chosen by extensively reviewing the state of the art were trained to decode the ten hand movements. Figure 4.8 illustrates the processing stages of real-time and offline myoelectric pattern recognition.

In total, each subject performed 12 trials (two for training the subjects (offline and real-time), one for offline training/testing the classifiers and nine for the real-time testing) to evaluate the nine classification algorithms. The recorded data in the third trial was used for training each classifier in combination with the selected feature set (nine combinations in total) using the offline training feature in BioPatRec.

Figure 4.8: The processing stages of real-time and offline myoelectric pattern recognition.

The motion test in BioPatRec was used to assess and compare the real-time classification of the selected algorithms in this study. During the motion test, subjects were requested to execute the ten selected motions three times in a randomized order. The selected hand and wrist movements for the real-time test are the same as in the offline test. The real-time accuracy, selection time, completion time and completion rate were used as the key indicators.


Preprocessing Feature Extraction Classification

Class Labels

Real-time Testing

Class Labels Offline Testing


the most commonly used set of features (four TD features reported by Hudgins et al. [43]: MAV, ZC, SSC, WL) was extracted from each window. Nine classification algorithms (see Table 4.2) chosen by extensively reviewing the state of the art were trained to decode the ten hand movements. Figure 4.8 illustrates the processing stages of real-time and offline myoelectric pattern recognition.

In total, each subject performed 12 trials (two for training the subjects (offline and real-time), one for offline training/testing the classifiers and nine for the real-time testing) to evaluate the nine classification algorithms. The recorded data in the third trial was used for training each classifier in combination with the selected feature set (nine combinations in total) using the offline training feature in BioPatRec.

Figure 4.8: The processing stages of real-time and offline myoelectric pattern recognition.

The motion test in BioPatRec was used to assess and compare the real-time classification of the selected algorithms in this study. During the motion test, subjects were requested to execute the ten selected motions three times in a randomized order. The selected hand and wrist movements for the real-time test are the same as in the offline test. The real-time accuracy, selection time, completion time and completion rate were used as the key indicators.


Preprocessing Feature Extraction Classification

Class Labels

Real-time Testing

Class Labels Offline Testing


Table 4.2: The nine investigated classifiers (their name, abbreviations, descriptions and publication references) in this study.

Classifiers Abbreviation Description References

Multilayer Perceptron MLP

A feedforward artificial neural network with an input layer, two hidden layers of 16 neurons each and an output layer

[50]

Regulatory Feedback Networks

RFN

Relatively new instance in pattern recognition instead of learning performs based on negative feedback.

[50]

Supervised Self-Organized Map

SSOM A type of artificial neural network that works based on competitive learning.

[88, 89]

Linear Discriminant Analysis LDA A simple, efficient and statistically-based classifier with a One-Vs-One (OVO) topology

[14, 51]

Support Vector Machine SVM

A very popular machine-learning algorithm that uses kernels to map data into separable hyperplanes.

[6, 10, 13]

K-Nearest Neighbor KNN

A simple machine-learning algorithm with a low training time that utilizes a distance measure to assign an unknown event to a given class.

[6, 69, 90]

Maximum Likelihood Estimation

MLE

Is used for parameter (mean and covariance of the probability density function) estimation in statistics

[72, 91]

Regression Tree RegTree Also known as decision tree, uses a set of features as input and outputs a decision

[92, 93]

Non-Negative Matrix Factorization

NMF

All the samples that do not have a class label are approximated by a non-negative linear combination of the training samples

[94, 95]


Table 4.2: The nine investigated classifiers (their name, abbreviations, descriptions and publication references) in this study.

Classifiers Abbreviation Description References

Multilayer Perceptron MLP

A feedforward artificial neural network with an input layer, two hidden layers of 16 neurons each and an output layer

[50]

Regulatory Feedback Networks

RFN

Relatively new instance in pattern recognition instead of learning performs based on negative feedback.

[50]

Supervised Self-Organized Map

SSOM A type of artificial neural network that works based on competitive learning.

[88, 89]

Linear Discriminant Analysis LDA A simple, efficient and statistically-based classifier with a One-Vs-One (OVO) topology

[14, 51]

Support Vector Machine SVM

A very popular machine-learning algorithm that uses kernels to map data into separable hyperplanes.

[6, 10, 13]

K-Nearest Neighbor KNN

A simple machine-learning algorithm with a low training time that utilizes a distance measure to assign an unknown event to a given class.

[6, 69, 90]

Maximum Likelihood Estimation

MLE

Is used for parameter (mean and covariance of the probability density function) estimation in statistics

[72, 91]

Regression Tree RegTree Also known as decision tree, uses a set of features as input and outputs a decision

[92, 93]

Non-Negative Matrix Factorization

NMF

All the samples that do not have a class label are approximated by a non-negative linear combination of the training samples

[94, 95]

Chapter 5

Results and Discussion

This chapter summarizes and discusses the results of the conventional methods and the proposed methods in papers A, B, C, D, E and F.

5.1 Filtering Techniques for Removing ECG Interference The conventional methods and the proposed methods in this study were applied to the contaminated EMG signals to remove ECG artifacts. For comparing the results of the proposed methods and other currently used methods, quantitative criteria such as the SNR, RE, CC, ET and PSD were used. The Welch method was used to estimate the PSD of the signals. The artifact removal methods were applied to the 10 channels of contaminated EMG signals recorded from five healthy subjects. The results for these signals using MATLAB R2014a on a PC with 2.6 GHz dual-core Intel Core i5 Turbo Boosted up to 3.1 GHz and 8GB 1600 MHz memory are presented in Table 5.1.


5.1.1.1 High-Pass Filter

A FIR high-pass filter with a Hamming window (length=100) and cutoff frequency of 30 Hz was implemented to remove ECG artifacts from surface EMG signals [17, 28, 54]. An HPF with a cutoff frequency of 30 Hz not only deletes all useful components below this frequency (Figures 5.1c and 5.1d) but also parts of the ECG artifact remain in the denoised EMG signal (Figure 5.1b). However, this method is straightforward and can be used in online applications. One of the advantages of this method is that the HPF does not require a separate ECG channel. The SNR, RE, CC and ET of this method are presented in Table 5.1.

Chapter 5

Results and Discussion

This chapter summarizes and discusses the results of the conventional methods and the proposed methods in papers A, B, C, D, E and F.

5.1 Filtering Techniques for Removing ECG Interference The conventional methods and the proposed methods in this study were applied to the contaminated EMG signals to remove ECG artifacts. For comparing the results of the proposed methods and other currently used methods, quantitative criteria such as the SNR, RE, CC, ET and PSD were used. The Welch method was used to estimate the PSD of the signals. The artifact removal methods were applied to the 10 channels of contaminated EMG signals recorded from five healthy subjects. The results for these signals using MATLAB R2014a on a PC with 2.6 GHz dual-core Intel Core i5 Turbo Boosted up to 3.1 GHz and 8GB 1600 MHz memory are presented in Table 5.1.


5.1.1.1 High-Pass Filter

A FIR high-pass filter with a Hamming window (length=100) and cutoff frequency of 30 Hz was implemented to remove ECG artifacts from surface EMG signals [17, 28, 54]. An HPF with a cutoff frequency of 30 Hz not only deletes all useful components below this frequency (Figures 5.1c and 5.1d) but also parts of the ECG artifact remain in the denoised EMG signal (Figure 5.1b). However, this method is straightforward and can be used in online applications. One of the advantages of this method is that the HPF does not require a separate ECG channel. The SNR, RE, CC and ET of this method are presented in Table 5.1.

32 Chapter 5. Result and Discussion

Figure 5.1: (a) Contaminated EMG, (b) denoised EMG, (c) estimated noise using

HPF and (d) PSD of raw EMG and denoised EMG using HPF.


To remove ECG artifacts from the EMG signals using the spike-clipping algorithm, a threshold value was set experimentally at 0.2. If the absolute value of the amplitude in the contaminated EMG signal is greater than the threshold, it sets to the threshold value. Spike clipping is a simple and fast method; however, it suffers from the loss of portions of EMG signal overlying QRS complexes (Figure 5.2) [18]. Figure 5.2d shows that this method removes high-frequency components of the original EMG signal, but low-frequency noise remains in the denoised signal. The performance of the spike-clipping method can be improved using an R peak detection algorithm; however, it is necessary to record an ECG signal simultaneously with the contaminated EMG.


The gating method is similar to a spike-clipping technique with a difference in the thresholding process. The absolute value of the amplitude greater than the threshold is set to zero. In this method, the threshold value was experimentally set to 0.15. As is presented in Figure 5.3b and 5.3c, parts of the EMG signal that have a high amplitude are removed, and a large part of the ECG artifact remains in the denoised signal. The PSD of the raw EMG and the denoised EMG using the gating method in Figure 5.3d show that this method removes high-frequency components of the EMG signal. It is



HPF and (d) PSD of raw EMG and denoised EMG using HPF.


To remove ECG artifacts from the EMG signals using the spike-clipping algorithm, a threshold value was set experimentally at 0.2. If the absolute value of the amplitude in the contaminated EMG signal is greater than the threshold, it sets to the threshold value. Spike clipping is a simple and fast method; however, it suffers from the loss of portions of EMG signal overlying QRS complexes (Figure 5.2) [18]. Figure 5.2d shows that this method removes high-frequency components of the original EMG signal, but low-frequency noise remains in the denoised signal. The performance of the spike-clipping method can be improved using an R peak detection algorithm; however, it is necessary to record an ECG signal simultaneously with the contaminated EMG.


The gating method is similar to a spike-clipping technique with a difference in the thresholding process. The absolute value of the amplitude greater than the threshold is set to zero. In this method, the threshold value was experimentally set to 0.15. As is presented in Figure 5.3b and 5.3c, parts of the EMG signal that have a high amplitude are removed, and a large part of the ECG artifact remains in the denoised signal. The PSD of the raw EMG and the denoised EMG using the gating method in Figure 5.3d show that this method removes high-frequency components of the EMG signal. It is

Chapter 5. Result and Discussion 33

possible to improve the performance of this method using an R peak detection algorithm at the cost of a higher computation time and more complex hardware needs.


spike clipping and (d) PSD of raw EMG and denoised EMG using spike clipping.


The hybrid technique is a combination of an adaptive spike clipping and an HPF. The first step is to apply the adaptive spike-clipping algorithm to the contaminated EMG signal. The second step is to use a fourth-order Butterworth HPF with a cutoff frequency of 30 Hz as a postprocessor. The contaminated EMG, denoised EMG, estimated noise and PSD of the raw EMG and denoised EMG using the hybrid technique are presented in Figure 5.4 The SNR, RE, CC and ET of this method are shown in Table 5.1. Using an HPF for removing low-frequency noise at the output of the spike-clipping algorithm improves the artifact removal process. However, this method removes some useful information from the original signal where the EMG has a high amplitude (Figure 5.4c). This approach is time-consuming (the adaptive spike-clipping part); therefore, it is not suitable for clinical applications (Table 5.1) [20].


possible to improve the performance of this method using an R peak detection algorithm at the cost of a higher computation time and more complex hardware needs.


spike clipping and (d) PSD of raw EMG and denoised EMG using spike clipping.


The hybrid technique is a combination of an adaptive spike clipping and an HPF. The first step is to apply the adaptive spike-clipping algorithm to the contaminated EMG signal. The second step is to use a fourth-order Butterworth HPF with a cutoff frequency of 30 Hz as a postprocessor. The contaminated EMG, denoised EMG, estimated noise and PSD of the raw EMG and denoised EMG using the hybrid technique are presented in Figure 5.4 The SNR, RE, CC and ET of this method are shown in Table 5.1. Using an HPF for removing low-frequency noise at the output of the spike-clipping algorithm improves the artifact removal process. However, this method removes some useful information from the original signal where the EMG has a high amplitude (Figure 5.4c). This approach is time-consuming (the adaptive spike-clipping part); therefore, it is not suitable for clinical applications (Table 5.1) [20].



gating method and (d) PSD of the raw EMG and denoised EMG using gating method.


hybrid technique and (d) PSD of the raw EMG and denoised EMG using hybrid

technique.



gating method and (d) PSD of the raw EMG and denoised EMG using gating method.


hybrid technique and (d) PSD of the raw EMG and denoised EMG using hybrid

technique.



In the first step, the Pan-Tompkins algorithm was used to detect R peaks in the contaminated EMG signal [55]. A window of a 2500-sample duration was used to segment the contaminated EMG around the detected R peaks (see Figure 5.5). The mean of all detected frames was taken to create an ECG template (Figure 5.6). Finally, the denoised EMG (Figure 5.7b) was obtained by subtracting the created ECG template from the contaminated EMG where the R peak was detected. Based on the results in Figure 5.7 and Table 5.1, this method has good performance in removing the ECG from surface EMG signals. Since the whole signal is required to obtain the ECG template, this method is not usable for online applications. Furthermore, it requires an additional ECG signal for the accurate detection of the QRS complexes.

Figure 5.5: Twenty selected frames including the QRS complexes in the

contaminated EMG signal.



In the first step, the Pan-Tompkins algorithm was used to detect R peaks in the contaminated EMG signal [55]. A window of a 2500-sample duration was used to segment the contaminated EMG around the detected R peaks (see Figure 5.5). The mean of all detected frames was taken to create an ECG template (Figure 5.6). Finally, the denoised EMG (Figure 5.7b) was obtained by subtracting the created ECG template from the contaminated EMG where the R peak was detected. Based on the results in Figure 5.7 and Table 5.1, this method has good performance in removing the ECG from surface EMG signals. Since the whole signal is required to obtain the ECG template, this method is not usable for online applications. Furthermore, it requires an additional ECG signal for the accurate detection of the QRS complexes.

Figure 5.5: Twenty selected frames including the QRS complexes in the

contaminated EMG signal.


Figure 5.6: ECG template created by averaging QRS complexes in the contaminated

EMG signal.


template subtraction and (d) PSD of the raw EMG and denoised EMG using template

subtraction.


Independent component analysis separates independent components from a number of mixed signals. The FastICA algorithm was used to separate the independent components from two channels of contaminated EMG signals (Figures 5.8a and


Figure 5.6: ECG template created by averaging QRS complexes in the contaminated

EMG signal.


template subtraction and (d) PSD of the raw EMG and denoised EMG using template

subtraction.


Independent component analysis separates independent components from a number of mixed signals. The FastICA algorithm was used to separate the independent components from two channels of contaminated EMG signals (Figures 5.8a and


5.8b). Then, an FIR high pass filter (Hamming window, order 100) with a cutoff frequency of 30 Hz was used to remove the ECG artifact from the noisy component (Figure 5.8c). The inverse ICA was applied to the denoised components to get the denoised EMG signals (Figure 5.9). It can be observed from Figure 5.9b that using ICA, small parts of the noise remain in the denoised signal, and parts of the desired signal are removed (Figure 5.9c and 5.9d). The requirement of an additional sensor is one of the drawbacks of this method since it adds to the complexity of the hardware system [96].

Figure 5.8: (a, b) Contaminated EMG signals, (c, d) independent components

obtained by ICA and (e, f) filtered components using HPF.


A discrete wavelet transform using a fourth-order Symlet wavelet was used to decompose the contaminated EMG signal to eight levels. Choosing this wavelet was due to the similarity of the mother wavelet to the ECG signal, which helps in removing ECG artifacts [97]. A threshold value was then defined to set the absolute value of the coefficients greater than the threshold to zero. This threshold was set experimentally. An inverse wavelet transform was then performed on the new coefficients to obtain the denoised EMG signal. It was observed that a considerable amount of noise remained in the original signal (Figure 5.10b). Figure 5.10d illustrates that the wavelet transform affects the EMG signal in high frequencies in a


5.8b). Then, an FIR high pass filter (Hamming window, order 100) with a cutoff frequency of 30 Hz was used to remove the ECG artifact from the noisy component (Figure 5.8c). The inverse ICA was applied to the denoised components to get the denoised EMG signals (Figure 5.9). It can be observed from Figure 5.9b that using ICA, small parts of the noise remain in the denoised signal, and parts of the desired signal are removed (Figure 5.9c and 5.9d). The requirement of an additional sensor is one of the drawbacks of this method since it adds to the complexity of the hardware system [96].

Figure 5.8: (a, b) Contaminated EMG signals, (c, d) independent components

obtained by ICA and (e, f) filtered components using HPF.


A discrete wavelet transform using a fourth-order Symlet wavelet was used to decompose the contaminated EMG signal to eight levels. Choosing this wavelet was due to the similarity of the mother wavelet to the ECG signal, which helps in removing ECG artifacts [97]. A threshold value was then defined to set the absolute value of the coefficients greater than the threshold to zero. This threshold was set experimentally. An inverse wavelet transform was then performed on the new coefficients to obtain the denoised EMG signal. It was observed that a considerable amount of noise remained in the original signal (Figure 5.10b). Figure 5.10d illustrates that the wavelet transform affects the EMG signal in high frequencies in a


negative way. However, one of the advantages of the wavelet transform is that it only requires a one-channel EMG.


ICA and (d) PSD of the raw EMG and denoised EMG using ICA.


wavelet transform and (d) PSD of the raw EMG and denoised EMG using wavelet

transform.


negative way. However, one of the advantages of the wavelet transform is that it only requires a one-channel EMG.


ICA and (d) PSD of the raw EMG and denoised EMG using ICA.


wavelet transform and (d) PSD of the raw EMG and denoised EMG using wavelet

transform.



A sixth-order Daubechies wavelet was used to decompose the contaminated EMG signal (60 seconds) into six levels to produce a multichannel EMG signal for ICA. After decomposing the signal, the FastICA algorithm was used to separate the desired signal from the artifact. The noisy component was then detected manually by the user and was set to zero. The inverse wavelet transform was used to obtain the denoised signal (Figure 5.11b). This technique does not require a multichannel signal, and it has better performance compared to the wavelet transform and the ICA technique (Table 5.1).


wavelet-ICA and (d) PSD of the raw EMG and denoised EMG using wavelet-ICA.


The contaminated EMG and the ECG signals were used as the input and reference input to the RLS algorithm, respectively. The algorithm was initialized with

, an RLS forgetting factor = 0.999 and filter length = 32. The results in Figure 5.12 and Table 5.1 show that the ANC is somewhat effective in removing the ECG from EMG signals. This method is capable of estimating the noise in the contaminated EMG signal using the reference ECG. However, the noise is not well estimated, and there is still noise in the denoised signal (Figure 5.12b and 5.12d). ANC is not suitable for clinical applications due to the heavy computational burden. It also takes a long time to get the denoised signal using this method (Table 5.1).



A sixth-order Daubechies wavelet was used to decompose the contaminated EMG signal (60 seconds) into six levels to produce a multichannel EMG signal for ICA. After decomposing the signal, the FastICA algorithm was used to separate the desired signal from the artifact. The noisy component was then detected manually by the user and was set to zero. The inverse wavelet transform was used to obtain the denoised signal (Figure 5.11b). This technique does not require a multichannel signal, and it has better performance compared to the wavelet transform and the ICA technique (Table 5.1).


wavelet-ICA and (d) PSD of the raw EMG and denoised EMG using wavelet-ICA.


The contaminated EMG and the ECG signals were used as the input and reference input to the RLS algorithm, respectively. The algorithm was initialized with

, an RLS forgetting factor = 0.999 and filter length = 32. The results in Figure 5.12 and Table 5.1 show that the ANC is somewhat effective in removing the ECG from EMG signals. This method is capable of estimating the noise in the contaminated EMG signal using the reference ECG. However, the noise is not well estimated, and there is still noise in the denoised signal (Figure 5.12b and 5.12d). ANC is not suitable for clinical applications due to the heavy computational burden. It also takes a long time to get the denoised signal using this method (Table 5.1).



ANC and (d) PSD of the raw EMG and denoised EMG using ANC.


The parameters used to train the ANN were two neurons in the input layer, 35 neurons in the only hidden layer and one neuron in the output layer. The ECG signal and delayed ECG signal (9 samples) were the inputs, and the contaminated EMG signal was the target to the network. Epochs = 1000, goal = 0.65, momentum = 0.9, show = 5, learning rate = 0.5 and time = infinity. The activation function for each layer was a hyperbolic tangent sigmoid, which is a transfer function. In this method, 120,000 samples (60 seconds) were selected from the dataset. 90,000 samples (45 seconds) were used for training and 30,000 samples (15 seconds) were used for testing. According to the quantitative evaluation criteria in Table 5.1, it can be observed that the performance of the ANN in removing noise from the EMG signal is relatively good. Using this method, a large amount of ECG artifacts in the EMG signal was removed (Figure 5.13); however, it can be seen in Figure 5.13d that low-frequency noise remained in the denoised signal. One of the drawbacks of this method is the requirement of an additional sensor to provide the reference input.



ANC and (d) PSD of the raw EMG and denoised EMG using ANC.


The parameters used to train the ANN were two neurons in the input layer, 35 neurons in the only hidden layer and one neuron in the output layer. The ECG signal and delayed ECG signal (9 samples) were the inputs, and the contaminated EMG signal was the target to the network. Epochs = 1000, goal = 0.65, momentum = 0.9, show = 5, learning rate = 0.5 and time = infinity. The activation function for each layer was a hyperbolic tangent sigmoid, which is a transfer function. In this method, 120,000 samples (60 seconds) were selected from the dataset. 90,000 samples (45 seconds) were used for training and 30,000 samples (15 seconds) were used for testing. According to the quantitative evaluation criteria in Table 5.1, it can be observed that the performance of the ANN in removing noise from the EMG signal is relatively good. Using this method, a large amount of ECG artifacts in the EMG signal was removed (Figure 5.13); however, it can be seen in Figure 5.13d that low-frequency noise remained in the denoised signal. One of the drawbacks of this method is the requirement of an additional sensor to provide the reference input.



ANN and (d) PSD of the raw EMG and denoised EMG using ANN.


For removing noise from the contaminated EMG signal, 120,000 samples (60 seconds) were selected. 90,000 samples (45 seconds) were used for training and 30,000 samples (15 seconds) were used for testing. The ECG signal and delayed ECG signal (9 samples) were inputs, and the contaminated EMG signal was the target to the system. The selection of appropriate parameters is essential for the performance of the ANFIS algorithm. In this study, the proposed parameters for training ANFIS were epochs = 1000, fuzzy rules = 9, inputs = 2, membership function = generalized bell (because of the advantages of smoothness and concise notation, it is used for tuning the fuzzy inference system parameters in nonlinear applications) and number of membership functions = 3, which were determined by a trial and error process. ANFIS was trained using the training database, and the ECG artifact was estimated using the testing database. The denoised signal was obtained by subtracting the estimated noise from the contaminated EMG signal. The performance of this method for estimating the ECG artifact in the EMG signal was similar to that for the ANN, but the elapsed time for the ANFIS method was lower (Table 5.1). ANFIS can be trained just once and after that can be used online quickly. However, the requirement of an additional sensor to provide the reference input is one of the drawbacks of this



ANN and (d) PSD of the raw EMG and denoised EMG using ANN.


For removing noise from the contaminated EMG signal, 120,000 samples (60 seconds) were selected. 90,000 samples (45 seconds) were used for training and 30,000 samples (15 seconds) were used for testing. The ECG signal and delayed ECG signal (9 samples) were inputs, and the contaminated EMG signal was the target to the system. The selection of appropriate parameters is essential for the performance of the ANFIS algorithm. In this study, the proposed parameters for training ANFIS were epochs = 1000, fuzzy rules = 9, inputs = 2, membership function = generalized bell (because of the advantages of smoothness and concise notation, it is used for tuning the fuzzy inference system parameters in nonlinear applications) and number of membership functions = 3, which were determined by a trial and error process. ANFIS was trained using the training database, and the ECG artifact was estimated using the testing database. The denoised signal was obtained by subtracting the estimated noise from the contaminated EMG signal. The performance of this method for estimating the ECG artifact in the EMG signal was similar to that for the ANN, but the elapsed time for the ANFIS method was lower (Table 5.1). ANFIS can be trained just once and after that can be used online quickly. However, the requirement of an additional sensor to provide the reference input is one of the drawbacks of this


method [98]. Furthermore, low-frequency noise remained in the denoised signal using ANFIS (Figure 5.14d).


ANFIS and (d) PSD of the raw EMG and denoised EMG using ANFIS.



In the first step, ANFIS with the parameters described in section 5.1.1.11 was employed to denoise the contaminated EMG signal. The result in Figure 5.14 shows that there are still low-frequency noise components (between 0 to 20 Hz) in the denoised signal using ANFIS. However, in the 5-20 Hz frequency range, the surface EMG spectrum contains information [99, 100] concerning the firing rates of the active motor unit that might be of interest in some applications such as movement analysis [100]; therefore, it is important to remove the noise from this particular frequency band. Hence, in the second step of the proposed method, a wavelet transform with nonlinear thresholding (Symlet wavelet family, fourth-order) was used to remove the residual noise at the output of the ANFIS method. The wavelet transform divides signals into different frequency components; therefore, it is possible to find the low-frequency noise in the original signal. Figure 5.15 illustrates that low-frequency noise was removed from the contaminated EMG signal when a wavelet transform was used as the postprocessor. It may be observed in Table 5.1 that


method [98]. Furthermore, low-frequency noise remained in the denoised signal using ANFIS (Figure 5.14d).


ANFIS and (d) PSD of the raw EMG and denoised EMG using ANFIS.



In the first step, ANFIS with the parameters described in section 5.1.1.11 was employed to denoise the contaminated EMG signal. The result in Figure 5.14 shows that there are still low-frequency noise components (between 0 to 20 Hz) in the denoised signal using ANFIS. However, in the 5-20 Hz frequency range, the surface EMG spectrum contains information [99, 100] concerning the firing rates of the active motor unit that might be of interest in some applications such as movement analysis [100]; therefore, it is important to remove the noise from this particular frequency band. Hence, in the second step of the proposed method, a wavelet transform with nonlinear thresholding (Symlet wavelet family, fourth-order) was used to remove the residual noise at the output of the ANFIS method. The wavelet transform divides signals into different frequency components; therefore, it is possible to find the low-frequency noise in the original signal. Figure 5.15 illustrates that low-frequency noise was removed from the contaminated EMG signal when a wavelet transform was used as the postprocessor. It may be observed in Table 5.1 that


combining ANFIS and a wavelet improved the SNR of the denoised signal with only a small increase in the processing time.


ANFIS-wavelet and (d) PSD of the raw EMG and denoised EMG using ANFIS-

wavelet.


The ECG artifact was first removed from the contaminated EMG signal using the ANN method. As can be seen in the PSD of the denoised EMG using ANN in Figure 5.13, low-frequency noise is remained in the signal. To remove the remaining noise from the output of the ANN, a wavelet transform with nonlinear thresholding (Symlet wavelet family, fourth-order) was used. Figure 5.16 shows the result of the ANN-wavelet method for one of the signals. In the PSD of the denoised EMG signal using ANN-wavelet in Figure 5.16, it can be seen that the low-frequency noise components have been removed by the second process using a wavelet transform. The parameters that were used to perform this method are the same as those for the ANN and wavelet transform that were described before. It was found that the proposed method, compared with the other methods, is capable of removing artifacts from the original signal, which is a promising result (Table 5.1). This method is fast and automated and can be used in clinical applications.

0 5 10 15

-0.5

0

0.5

Am

plitu

de (

mv)

a

0 5 10 15

-0.5

0

0.5

Am

plitu

de (

mv)

b

0 5 10 15Time(sec)

-0.5

0

0.5

Am

plitu

de (

mv)

c

0 50 100 150 200Frequency(Hz)

0

0.005

0.01

0.015

0.02

0.025

PS

D

d

Clean EMGDenoised EMG using ANFIS-wavelet


combining ANFIS and a wavelet improved the SNR of the denoised signal with only a small increase in the processing time.


ANFIS-wavelet and (d) PSD of the raw EMG and denoised EMG using ANFIS-

wavelet.


The ECG artifact was first removed from the contaminated EMG signal using the ANN method. As can be seen in the PSD of the denoised EMG using ANN in Figure 5.13, low-frequency noise is remained in the signal. To remove the remaining noise from the output of the ANN, a wavelet transform with nonlinear thresholding (Symlet wavelet family, fourth-order) was used. Figure 5.16 shows the result of the ANN-wavelet method for one of the signals. In the PSD of the denoised EMG signal using ANN-wavelet in Figure 5.16, it can be seen that the low-frequency noise components have been removed by the second process using a wavelet transform. The parameters that were used to perform this method are the same as those for the ANN and wavelet transform that were described before. It was found that the proposed method, compared with the other methods, is capable of removing artifacts from the original signal, which is a promising result (Table 5.1). This method is fast and automated and can be used in clinical applications.

0 5 10 15

-0.5

0

0.5

Am

plitu

de (

mv)

a

0 5 10 15

-0.5

0

0.5

Am

plitu

de (

mv)

b

0 5 10 15Time(sec)

-0.5

0

0.5

Am

plitu

de (

mv)

c

0 50 100 150 200Frequency(Hz)

0

0.005

0.01

0.015

0.02

0.025

PS

D

d

Clean EMGDenoised EMG using ANFIS-wavelet



ANN-wavelet and (d) PSD of the raw EMG and denoised EMG using ANN-wavelet.


After detecting the R peaks in the contaminated EMG signal, ECG templates were created by averaging the 30 past QRS complexes for each detected R wave. In this way, ECG templates are changed for each detected R to be adapted to each QRS complex. For removing undesirable artifacts from the created ECG templates, a low-pass filter with a cutoff frequency of 50 Hz was used. To obtain the denoised signal (Figure 5.17b), the created ECG templates were subtracted from the contaminated EMG signal where the R peaks were detected. The difference between this method and the template subtraction method is in the ECG template creation step. Unlike the template subtraction, this technique is an online method that can be used in clinical applications at a cost of a drop in the SNR value (Table 5.1).


A wavelet transform (Daubechies wavelet family, fourth-order) was used to decompose the contaminated EMG signal into eight levels. After decomposing the signal, the ICA algorithm was employed to separate the independent components. An algorithm based on a Hilbert transform was used to find the noisy component automatically (described in section 11.2.2). Then, an HPF with a cutoff frequency of 20 Hz was used to remove ECG artifacts from the detected noisy component. Finally, the inverse ICA and inverse wavelet transform were applied to obtain the denoised



ANN-wavelet and (d) PSD of the raw EMG and denoised EMG using ANN-wavelet.


After detecting the R peaks in the contaminated EMG signal, ECG templates were created by averaging the 30 past QRS complexes for each detected R wave. In this way, ECG templates are changed for each detected R to be adapted to each QRS complex. For removing undesirable artifacts from the created ECG templates, a low-pass filter with a cutoff frequency of 50 Hz was used. To obtain the denoised signal (Figure 5.17b), the created ECG templates were subtracted from the contaminated EMG signal where the R peaks were detected. The difference between this method and the template subtraction method is in the ECG template creation step. Unlike the template subtraction, this technique is an online method that can be used in clinical applications at a cost of a drop in the SNR value (Table 5.1).


A wavelet transform (Daubechies wavelet family, fourth-order) was used to decompose the contaminated EMG signal into eight levels. After decomposing the signal, the ICA algorithm was employed to separate the independent components. An algorithm based on a Hilbert transform was used to find the noisy component automatically (described in section 11.2.2). Then, an HPF with a cutoff frequency of 20 Hz was used to remove ECG artifacts from the detected noisy component. Finally, the inverse ICA and inverse wavelet transform were applied to obtain the denoised


signal. The results in Figure 5.18 show that small peaks remained in the denoised signal (Figure 5.186b) and some parts of the original signal are removed (Figures 5.18c and 5.18d). The calculated values (SNR, RE, CC and ET) for this method, as listed in Table 5.1, show that this algorithm is capable of eliminating the ECG artifacts. This technique is an automatic artifact removal method that does not require a multichannel signal and can operate online with a promising result for many applications.


adaptive subtraction, (d) PSD of the raw EMG and denoised EMG using adaptive

subtraction.

All the evaluated criteria in Table 5.1 confirmed that the proposed methods have good performance and efficiency. They are fast and could successfully remove ECG artifacts from surface EMG signals. Based on the results presented in Table 5.1, the proposed ANFIS-wavelet improved the SNR value of ANFIS by up to 22%, and the proposed ANN-wavelet improved the SNR value of ANN up to 31%. In the proposed automated wavelet-ICA, not only was the method converted to an automatic approach, but the SNR value was also improved by up to 20% by combining it with an HPF. The adaptive subtraction did not improve the results. The SNR, RE and CC were decreased, but it converted the template subtraction to an online technique. Table 5.1 also shows the elapsed time of each method in seconds. The HPF, spike-clipping and gating methods were faster than the other methods. The hybrid technique

0 5 10 15

-0.5

0

0.5

Am

plitu

de (m

v)

a

0 5 10 15

-0.5

0

0.5

Am

plitu

de (m

v)

b

0 5 10 15time(sec)

-0.5

0

0.5

Am

plitu

de (m

v)

c

0 50 100 150 200 250frequency(Hz)

0

0.005

0.01

0.015

0.02

0.025

PS

D

d

Clean EMGDenoised EMG using adaptive subtraction


signal. The results in Figure 5.18 show that small peaks remained in the denoised signal (Figure 5.186b) and some parts of the original signal are removed (Figures 5.18c and 5.18d). The calculated values (SNR, RE, CC and ET) for this method, as listed in Table 5.1, show that this algorithm is capable of eliminating the ECG artifacts. This technique is an automatic artifact removal method that does not require a multichannel signal and can operate online with a promising result for many applications.


adaptive subtraction, (d) PSD of the raw EMG and denoised EMG using adaptive

subtraction.

All the evaluated criteria in Table 5.1 confirmed that the proposed methods have good performance and efficiency. They are fast and could successfully remove ECG artifacts from surface EMG signals. Based on the results presented in Table 5.1, the proposed ANFIS-wavelet improved the SNR value of ANFIS by up to 22%, and the proposed ANN-wavelet improved the SNR value of ANN up to 31%. In the proposed automated wavelet-ICA, not only was the method converted to an automatic approach, but the SNR value was also improved by up to 20% by combining it with an HPF. The adaptive subtraction did not improve the results. The SNR, RE and CC were decreased, but it converted the template subtraction to an online technique. Table 5.1 also shows the elapsed time of each method in seconds. The HPF, spike-clipping and gating methods were faster than the other methods. The hybrid technique

0 5 10 15

-0.5

0

0.5

Am

plitu

de (m

v)

a

0 5 10 15

-0.5

0

0.5

Am

plitu

de (m

v)

b

0 5 10 15time(sec)

-0.5

0

0.5

Am

plitu

de (m

v)

c

0 50 100 150 200 250frequency(Hz)

0

0.005

0.01

0.015

0.02

0.025

PS

D

d

Clean EMGDenoised EMG using adaptive subtraction


and ANC were time-consuming. The elapsed time to run the proposed ANFIS-wavelet and ANN-wavelet were increased by 0.15 and 0.35 second, respectively, compared to ANFIS and ANN. The cost of making the wavelet-ICA method automated was a 0.24-second increase in the elapsed time to run the algorithm.


automated wavelet-ICA and (d) PSD of the raw EMG and denoised EMG using

automated wavelet-ICA.

In Table 5.2, different classes (e.g., very fast, fast, time-consuming, online, offline, automatic) were considered for artifact removal methods. Therefore, each method can be chosen based on a specific application. For example, for applications that require a speedy result, HPF is recommended among the very fast methods (HPF, spike-clipping and gating methods) because it has a higher SNR value. For applications that require very good performance, ANFIS-wavelet and ANN-wavelet are recommended. For applications that need a less complex hardware system, methods that do not require a multichannel input such as HPF, spike-clipping, the gating method, template subtraction, the wavelet transform, wavelet-ICA, adaptive subtraction and automated wavelet-ICA are good options, but among these methods, template subtraction, adaptive subtraction and automated wavelet-ICA are recommended because of their good performance (high SNR value and low ET).


and ANC were time-consuming. The elapsed time to run the proposed ANFIS-wavelet and ANN-wavelet were increased by 0.15 and 0.35 second, respectively, compared to ANFIS and ANN. The cost of making the wavelet-ICA method automated was a 0.24-second increase in the elapsed time to run the algorithm.


automated wavelet-ICA and (d) PSD of the raw EMG and denoised EMG using

automated wavelet-ICA.

In Table 5.2, different classes (e.g., very fast, fast, time-consuming, online, offline, automatic) were considered for artifact removal methods. Therefore, each method can be chosen based on a specific application. For example, for applications that require a speedy result, HPF is recommended among the very fast methods (HPF, spike-clipping and gating methods) because it has a higher SNR value. For applications that require very good performance, ANFIS-wavelet and ANN-wavelet are recommended. For applications that need a less complex hardware system, methods that do not require a multichannel input such as HPF, spike-clipping, the gating method, template subtraction, the wavelet transform, wavelet-ICA, adaptive subtraction and automated wavelet-ICA are good options, but among these methods, template subtraction, adaptive subtraction and automated wavelet-ICA are recommended because of their good performance (high SNR value and low ET).


Table 5.1: SNR, RE, CC and ET in seconds obtained for HPF, spike clipping, gating method, hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANC, ANN, ANFIS, ANFIS-wavelet, ANN-wavelet, adaptive subtraction and automated (auto) wavelet-ICA for ten contaminated EMG signals.

Filtering Methods SNR (dB) RE CC ET (seconds)

HPF 7.75±0.79 0.12±0.05 0.89±0.01 0.003

Spike clipping 3.14±0.44 0.36±0.04 0.77±0.02 0.007

Gating method 2.69±0.42 0.34±0.05 0.70±0.04 0.006

Hybrid technique 7.07±0.84 0.15±0.05 0.89±0.02 8.60

Template subtraction 11.41±0.91 0.05±0.02 0.96±0.01 0.42

ICA 7.93±0.62 0.09±0.04 0.92±0.03 1.64

Wavelet transform 5.36±0.81 0.15±0.03 0.86±0.02 0.37

Wavelet-ICA 8.64±1.92 0.09±0.04 0.92±0.03 1.85

ANC 8.09±1.29 0.12±0.04 0.92±0.02 126

ANN 11.85±1.16 0.05±0.02 0.96±0.01 0.31

ANFIS 12.27±1.06 0.04±0.01 0.97±0.01 0.12

ANFIS-wavelet 14.97±1.34 0.02±0.02 0.99±0.02 0.27

ANN-wavelet 15.53±1.14 0.01±0.00 0.98±0.00 0.66

Adaptive subtraction 9.58±0.83 0.07±0.02 0.94±1.92 1.20

Auto wavelet-ICA 10.41±0.67 0.06±0.04 0.95±0.03 2.09

Table 5.2: The advantages and disadvantages of HPF, spike clipping, gating method, hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANC, ANN, ANFIS, ANFIS-wavelet, ANN-wavelet, adaptive subtraction, and automated (auto) wavelet-ICA in removing ECG artifacts from EMG signals.

Filtering Methods Advantages Disadvantages

HPF 1. Very fast (Table 5.1)2

2. Not requiring multichannel signals

3. Automatic

4. Online

1. Removing a large amount of useful information from the original signal

2. Parts of artifacts remaining in the denoised signal

Spike clipping 1. Very fast (Table 5.1)


3. Automatic

4. Online

1. Removing parts of useful information from the original signal

2. A large amount of artifacts remaining in the denoised signal

Gating method 1. Very fast (Table 5.1) 1. Removing a large amount of useful

2 ET less than 0.01 sec.


Table 5.1: SNR, RE, CC and ET in seconds obtained for HPF, spike clipping, gating method, hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANC, ANN, ANFIS, ANFIS-wavelet, ANN-wavelet, adaptive subtraction and automated (auto) wavelet-ICA for ten contaminated EMG signals.

Filtering Methods SNR (dB) RE CC ET (seconds)

HPF 7.75±0.79 0.12±0.05 0.89±0.01 0.003

Spike clipping 3.14±0.44 0.36±0.04 0.77±0.02 0.007

Gating method 2.69±0.42 0.34±0.05 0.70±0.04 0.006

Hybrid technique 7.07±0.84 0.15±0.05 0.89±0.02 8.60

Template subtraction 11.41±0.91 0.05±0.02 0.96±0.01 0.42

ICA 7.93±0.62 0.09±0.04 0.92±0.03 1.64

Wavelet transform 5.36±0.81 0.15±0.03 0.86±0.02 0.37

Wavelet-ICA 8.64±1.92 0.09±0.04 0.92±0.03 1.85

ANC 8.09±1.29 0.12±0.04 0.92±0.02 126

ANN 11.85±1.16 0.05±0.02 0.96±0.01 0.31

ANFIS 12.27±1.06 0.04±0.01 0.97±0.01 0.12

ANFIS-wavelet 14.97±1.34 0.02±0.02 0.99±0.02 0.27

ANN-wavelet 15.53±1.14 0.01±0.00 0.98±0.00 0.66

Adaptive subtraction 9.58±0.83 0.07±0.02 0.94±1.92 1.20

Auto wavelet-ICA 10.41±0.67 0.06±0.04 0.95±0.03 2.09

Table 5.2: The advantages and disadvantages of HPF, spike clipping, gating method, hybrid technique, template subtraction, ICA, wavelet transform, wavelet-ICA, ANC, ANN, ANFIS, ANFIS-wavelet, ANN-wavelet, adaptive subtraction, and automated (auto) wavelet-ICA in removing ECG artifacts from EMG signals.

Filtering Methods Advantages Disadvantages

HPF 1. Very fast (Table 5.1)2


3. Automatic

4. Online



Spike clipping 1. Very fast (Table 5.1)


3. Automatic

4. Online



Gating method 1. Very fast (Table 5.1) 1. Removing a large amount of useful

2 ET less than 0.01 sec.



3. Automatic

4. Online

information from the original signal


Hybrid technique 1. Not requiring multichannel signals

2. Automatic

3. Online


2. Time-consuming (Table 5.1)

Template subtraction 1. Fast (Table 5.1)3


3. Automatic

4. Good performance (Table 5.1)4

1. Requiring predetermined QRS template

2. Offline

ICA 1. Fast (Table 5.1)

2. Automatic

3. Online

1. Requiring multichannel signals

2. A large amount of useful information removed from the original signal

Wavelet transform 1. Fast (Table 5.1)


3. Automatic

4. Online


2. Parts of artifact remaining in the denoised signal

Wavelet-ICA 1. Fast (Table 5.1)


3. Online

1. User-dependent



ANC 1. Automatic

2. Online


2. Requiring reference signal


ANN 1. Fast (Table 5.1)

2. Automatic

3. Online

4. Good performance (Table 5.1)


2. Part of low-frequency noise remaining in the denoised signal

ANFIS 1. Fast (Table 5.1)

2. Automatic

3. Online




ANFIS-wavelet 1. Fast (Table 5.1)

2. Automatic


3 ET less than 3 sec.SNR approximately equal to or more than 9.



3. Automatic

4. Online

information from the original signal


Hybrid technique 1. Not requiring multichannel signals

2. Automatic

3. Online



Template subtraction 1. Fast (Table 5.1)3


3. Automatic

4. Good performance (Table 5.1)4


2. Offline

ICA 1. Fast (Table 5.1)

2. Automatic

3. Online

1. Requiring multichannel signals

2. A large amount of useful information removed from the original signal

Wavelet transform 1. Fast (Table 5.1)


3. Automatic

4. Online


2. Parts of artifact remaining in the denoised signal

Wavelet-ICA 1. Fast (Table 5.1)


3. Online

1. User-dependent



ANC 1. Automatic

2. Online




ANN 1. Fast (Table 5.1)

2. Automatic

3. Online




ANFIS 1. Fast (Table 5.1)

2. Automatic

3. Online




ANFIS-wavelet 1. Fast (Table 5.1)

2. Automatic


3 ET less than 3 sec.SNR approximately equal to or more than 9.


3. Online

4. Very good performance (Table 5.1)

ANN-wavelet 1. Fast (Table 5.1)

2. Automatic

3. Online

4. The best performance based on the SNR value (Table 5.1)


Adaptive subtraction 1. Fast (Table 5.1)


3. Automatic

4. Online



Auto wavelet-ICA 1. Fast (Table 5.1)


3. Automatic

4. Online


1. Removing useful information from the original signal

2. Artifacts remaining in the denoised signal

5.2 Offline Pattern Recognition (Paper E) The offline analysis of 44 individual features in the TD, FD and TFD in combination with the six classification algorithms (LDA, KNN, RegTree, MLE, SVM and MLP) to decode individual hand movements showed that:

• Among the 25 individual TD features, MAV, standard deviation (STD), WL, mean of peek value (MPV), difference absolute mean value (DAMV), maximum fractal length (MFL), integrated absolute value (IAV), difference absolute standard deviation value (DASDV) and percentile (Perc) performed better (based on the obtained classification accuracy rates). Among the classifiers, KNN and MLP had higher accuracy rates, in combination with the individual TD features, while KNN was slightly better than MLP, with an average accuracy rate of 93.9% for the MFL/KNN combination. Among the investigated TD features, mean absolute value slope (MAVS) obtained the lowest accuracy rate with an average of 11.5% in combination with the LDA classifier (see Table 5.3).

• Most of the investigated FD features in this study obtained low accuracy rates; however, the highest rate was obtained by frequency energy (FE) in combination with the KNN classifier, with an average accuracy rate of 90.0% (see Table 5.4).


3. Online

4. Very good performance (Table 5.1)

ANN-wavelet 1. Fast (Table 5.1)

2. Automatic

3. Online

4. The best performance based on the SNR value (Table 5.1)


Adaptive subtraction 1. Fast (Table 5.1)


3. Automatic

4. Online



Auto wavelet-ICA 1. Fast (Table 5.1)


3. Automatic

4. Online


1. Removing useful information from the original signal

2. Artifacts remaining in the denoised signal

5.2 Offline Pattern Recognition (Paper E) The offline analysis of 44 individual features in the TD, FD and TFD in combination with the six classification algorithms (LDA, KNN, RegTree, MLE, SVM and MLP) to decode individual hand movements showed that:

• Among the 25 individual TD features, MAV, standard deviation (STD), WL, mean of peek value (MPV), difference absolute mean value (DAMV), maximum fractal length (MFL), integrated absolute value (IAV), difference absolute standard deviation value (DASDV) and percentile (Perc) performed better (based on the obtained classification accuracy rates). Among the classifiers, KNN and MLP had higher accuracy rates, in combination with the individual TD features, while KNN was slightly better than MLP, with an average accuracy rate of 93.9% for the MFL/KNN combination. Among the investigated TD features, mean absolute value slope (MAVS) obtained the lowest accuracy rate with an average of 11.5% in combination with the LDA classifier (see Table 5.3).

• Most of the investigated FD features in this study obtained low accuracy rates; however, the highest rate was obtained by frequency energy (FE) in combination with the KNN classifier, with an average accuracy rate of 90.0% (see Table 5.4).


• Among the eleven TFD features in this study, eight of them were extracted from wavelet coefficients, and three of them were extracted from wavelet packet (WP) coefficients. The WT features did not perform well in decoding the hand movements, and the highest accuracy rate was obtained by WL, with an average of 67.7% in combination with MLP. Among the WP features, logarithmic RMS (LogRMS) and normalized logarithmic energy (NLE) produced the highest accuracy rates, with averages of 95.2% and 95.1%, respectively, in combination with the LDA classifier. KNN and SVM also obtained accuracies above 93% in combination with the abovementioned WP features (see Table 5.5).

Table 5.3: The average classification accuracy and standard deviation of individual TD features that obtained higher accuracy rates in combinations with the six classifiers across 20 subjects.

Average classification accuracy (%) ± STD

TDF LDA KNN RegTree MLE SVM MLP

MAV 84.6c±7.73 93.2ab±4.60 90.0abc±4.92 89.9bcd±6.21 84.0bc±8.47 91.1bcdef±4.69

STD 84.7c±7.95 93.5a±3.56 89.7abcd±5.01 89.4bcd±6.05 84.6bc±7.62 91.0cdef±4.20

WL 84.7c±7.43 93.2ab±4.38 90.2abc±5.47 90.4bc±5.88 84.9bc±8.08 91.2bcde±4.69

MPV 83.5cd±7.81 92.3abc ±4.36 88.8abcd±5.01 87.8bcde±6.64 84.1bc±7.11 89.8defg±4.69

DAMV 84.8c±7.45 93.3ab ±4.50 90.3abc±5.40 90.4bc±5.84 84.0bc±8.11 91.2bcdef±4.72

MFL 84.9c±7.47 93.9a±4.41 90.2abc±5.50 90.4bc±5.98 84.2bc±7.73 91.5abcde±4.41

IAV 84.6c±7.75 93.2ab±4.49 90.0abc±5.13 89.0bc±6.04 84.0bc±8.08 91.6abcde±4.46

DASDV 83.9cd±7.95 92.3abc±4.50 88.9abcd±5.31 88.2bcde±5.98 85.0bc±7.44 89.7defg±4.29 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Table 5.4: The average classification accuracy and standard deviation of FE in combinations with the six classifiers across 20 subjects.


FDF LDA KNN RegTree MLE SVM MLP

FE 64.1ghi±9.23 90.0abc±6.41 82.5efgh±7.87 16.8u±6.76 62.5fg±9.71 61.6mno±10.3 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Different combinations of features in the TD, FD and TFD were investigated using feature selection and feature projection (PCA) methods to develop a set of features that improves the performance of hand gesture recognition based on their obtained accuracy and processing time. A set of TD features (FS) including WL, CC, Hjorth parameters (activity, mobility and complexity) was found to improve the results. The proposed feature set significantly (p<0.0001) increased the accuracy rate of LDA


• Among the eleven TFD features in this study, eight of them were extracted from wavelet coefficients, and three of them were extracted from wavelet packet (WP) coefficients. The WT features did not perform well in decoding the hand movements, and the highest accuracy rate was obtained by WL, with an average of 67.7% in combination with MLP. Among the WP features, logarithmic RMS (LogRMS) and normalized logarithmic energy (NLE) produced the highest accuracy rates, with averages of 95.2% and 95.1%, respectively, in combination with the LDA classifier. KNN and SVM also obtained accuracies above 93% in combination with the abovementioned WP features (see Table 5.5).

Table 5.3: The average classification accuracy and standard deviation of individual TD features that obtained higher accuracy rates in combinations with the six classifiers across 20 subjects.


TDF LDA KNN RegTree MLE SVM MLP

MAV 84.6c±7.73 93.2ab±4.60 90.0abc±4.92 89.9bcd±6.21 84.0bc±8.47 91.1bcdef±4.69

STD 84.7c±7.95 93.5a±3.56 89.7abcd±5.01 89.4bcd±6.05 84.6bc±7.62 91.0cdef±4.20

WL 84.7c±7.43 93.2ab±4.38 90.2abc±5.47 90.4bc±5.88 84.9bc±8.08 91.2bcde±4.69

MPV 83.5cd±7.81 92.3abc ±4.36 88.8abcd±5.01 87.8bcde±6.64 84.1bc±7.11 89.8defg±4.69

DAMV 84.8c±7.45 93.3ab ±4.50 90.3abc±5.40 90.4bc±5.84 84.0bc±8.11 91.2bcdef±4.72

MFL 84.9c±7.47 93.9a±4.41 90.2abc±5.50 90.4bc±5.98 84.2bc±7.73 91.5abcde±4.41

IAV 84.6c±7.75 93.2ab±4.49 90.0abc±5.13 89.0bc±6.04 84.0bc±8.08 91.6abcde±4.46

DASDV 83.9cd±7.95 92.3abc±4.50 88.9abcd±5.31 88.2bcde±5.98 85.0bc±7.44 89.7defg±4.29 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Table 5.4: The average classification accuracy and standard deviation of FE in combinations with the six classifiers across 20 subjects.


FDF LDA KNN RegTree MLE SVM MLP

FE 64.1ghi±9.23 90.0abc±6.41 82.5efgh±7.87 16.8u±6.76 62.5fg±9.71 61.6mno±10.3 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Different combinations of features in the TD, FD and TFD were investigated using feature selection and feature projection (PCA) methods to develop a set of features that improves the performance of hand gesture recognition based on their obtained accuracy and processing time. A set of TD features (FS) including WL, CC, Hjorth parameters (activity, mobility and complexity) was found to improve the results. The proposed feature set significantly (p<0.0001) increased the accuracy rate of LDA


from 84.9% to 95.1%, MLE from 90.4% to 97.4%, and SVM from 85.0% to 92.0% compared to the individual TD features. However, the elapsed time slightly increased which is still acceptable for real-time applications. It can be observed in Table 5.6 that our proposed feature set outperformed the other sets and obtained accuracy rates up to 97%. No significant improvement (p<0.0001) was found by applying PCA as the feature reduction method to the three different feature sets (the 25 TD features, 8 FD features and 11 TFD features) in comparison with the individual features and our proposed feature set (see Table 5.6). The result of our proposed feature set was then compared with that of a well-known feature set (Hudgins’ set of five TD features). The results in Table 5.6 demonstrated that among the classifiers, LDA, KNN, MLE and SVM obtained significantly lower (p<0.0001) accuracy rates in combination with the Hudgins’ set when compared to our proposed set. Table 5.6 shows that less variability was found in the obtained classification accuracy by our proposed feature set and Hudgins’ set.

Table 5.5: The average classification accuracy and standard deviation of individual TFD features that obtained higher accuracy rates in combinations with the six classifiers across 20 subjects.


TFDF LDA KNN RegTree MLE SVM MLP

LogRMS 95.2a±3.63 93.3a±4.94 86.6bcde±6.04 76.1h±9.02 93.5a±5.51 95.5a±3.51

NLE 95.1a±3.76 93.3ab±5.00 86.5bcde±6.29 76.2h±9.56 93.8a±5.29 95.2ab±3.61 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Table 5.6: The average classification accuracy and standard deviation of feature sets in combinations with the six classifiers across 20 subjects.


LDA KNN RegTree MLE SVM MLP

FS 95.1a±2.85 94.1a±3.77 91.4a±4.56 97.4a±1.70 92.0a±3.82 93.7abcd±3.22

Hudgins set 89.9b±5.23 88.7bc±7.02 90.6ab±5.62 91.3b±5.12 86.3b±7.53 91.6abcde±4.26

TD/PCA 96.2a±2.44 92.5abc±5.02 79.7h±10.49 83.8efg±7.09 94.7a±3.20 94.4abc±3.18

FD/PCA 81.9cde±8.37 90.8abc±6.44 78.5hi±9.42 19.0tu±8.00 63.1f±7.91 85.2hi±5.07

TFD/PCA 94.9a±3.97 90.3abc±6.06 73.9i±8.76 79.4gh±9.65 93.3a±5.55 85.4hi±6.36 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Regarding the processing time, the training and testing time of all the classifiers were measured in combination with the features and feature sets that obtained good accuracies. The result showed that KNN and MLE are faster in training but MLP is


from 84.9% to 95.1%, MLE from 90.4% to 97.4%, and SVM from 85.0% to 92.0% compared to the individual TD features. However, the elapsed time slightly increased which is still acceptable for real-time applications. It can be observed in Table 5.6 that our proposed feature set outperformed the other sets and obtained accuracy rates up to 97%. No significant improvement (p<0.0001) was found by applying PCA as the feature reduction method to the three different feature sets (the 25 TD features, 8 FD features and 11 TFD features) in comparison with the individual features and our proposed feature set (see Table 5.6). The result of our proposed feature set was then compared with that of a well-known feature set (Hudgins’ set of five TD features). The results in Table 5.6 demonstrated that among the classifiers, LDA, KNN, MLE and SVM obtained significantly lower (p<0.0001) accuracy rates in combination with the Hudgins’ set when compared to our proposed set. Table 5.6 shows that less variability was found in the obtained classification accuracy by our proposed feature set and Hudgins’ set.

Table 5.5: The average classification accuracy and standard deviation of individual TFD features that obtained higher accuracy rates in combinations with the six classifiers across 20 subjects.


TFDF LDA KNN RegTree MLE SVM MLP

LogRMS 95.2a±3.63 93.3a±4.94 86.6bcde±6.04 76.1h±9.02 93.5a±5.51 95.5a±3.51

NLE 95.1a±3.76 93.3ab±5.00 86.5bcde±6.29 76.2h±9.56 93.8a±5.29 95.2ab±3.61 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Table 5.6: The average classification accuracy and standard deviation of feature sets in combinations with the six classifiers across 20 subjects.


LDA KNN RegTree MLE SVM MLP

FS 95.1a±2.85 94.1a±3.77 91.4a±4.56 97.4a±1.70 92.0a±3.82 93.7abcd±3.22

Hudgins set 89.9b±5.23 88.7bc±7.02 90.6ab±5.62 91.3b±5.12 86.3b±7.53 91.6abcde±4.26

TD/PCA 96.2a±2.44 92.5abc±5.02 79.7h±10.49 83.8efg±7.09 94.7a±3.20 94.4abc±3.18

FD/PCA 81.9cde±8.37 90.8abc±6.44 78.5hi±9.42 19.0tu±8.00 63.1f±7.91 85.2hi±5.07

TFD/PCA 94.9a±3.97 90.3abc±6.06 73.9i±8.76 79.4gh±9.65 93.3a±5.55 85.4hi±6.36 a,b,… means within a column not sharing a common superscript are significantly different (p<0.0001)

Regarding the processing time, the training and testing time of all the classifiers were measured in combination with the features and feature sets that obtained good accuracies. The result showed that KNN and MLE are faster in training but MLP is


very slow. The fastest combination (MAV/MLE) took 0.008 seconds. KNN and RegTree were faster in testing, but LDA was more time-consuming. The FD and TFD features were computationally expensive since the TD EMG signal needs to be transformed to the frequency and time-frequency domains.

5.3 Real-Time Pattern Recognition (Paper F) The real-time and offline performance of nine classification algorithms (MLP, RFN, supervised self-organized map [SSOM], LDA, SVM, KNN, MLE, RegTree and non-negative matrix factorization [NMF]) in combination with the Hudgins set of four TD features were compared using their obtained accuracy rates and processing time to classify ten individual hand movements. The offline performance was evaluated using the accuracy rate, training time in seconds and testing time in milliseconds (ms). To evaluate the real-time performance of the classifiers, the real-time accuracy, selection time in seconds, completion time in seconds and completion rate were used as the performance indicators. The selection time was the time between the first prediction different than “rest” and the first correct prediction. The completion time was the time between the first prediction different than “res”’ and the 20th correct prediction. The completion rate was the percentage of movements that reached 20 percent correct predictions before timeout.

Table 5.7 illustrates that among the classifiers, LDA and MLE performed significantly better (p<0.05) based on their obtained offline accuracy rates (above 97%). SSOM and KNN obtained the lowest accuracy rates, which were statistically significant (p<0.05). Among the classifiers, MLE was the fastest (0.171 s), and NMF was the slowest (46.3 s) in training; MLP was the fastest (0.115 ms) and NMF the slowest (286.6 ms) classifiers in offline testing.

Based on the real-time accuracy obtained by the motion test, MLP, LDA and MLE performed better than the other classifiers, with an accuracy rate of above 68%. MLP obtained the lowest selection time (0.657 s), and NMF obtained the highest selection time (1.99 s). The lowest completion time belongs to KNN, with an average of 3.85 seconds, and MLP has the highest completion time, with an average of 5.06 seconds. The highest completion rates were obtained by LDA and MLE, with average rates of 71.3% and 72.4%, respectively. SSOM and KNN obtained the lowest completion rates, with average rates of 42.9% and 38.9%.

The result illustrated that each classifier obtained higher offline accuracy rates in comparison with that of its real time. Some hand movements were easier to detect offline such as “Extend hand” since most of the classifiers were able to detect it with


very slow. The fastest combination (MAV/MLE) took 0.008 seconds. KNN and RegTree were faster in testing, but LDA was more time-consuming. The FD and TFD features were computationally expensive since the TD EMG signal needs to be transformed to the frequency and time-frequency domains.

5.3 Real-Time Pattern Recognition (Paper F) The real-time and offline performance of nine classification algorithms (MLP, RFN, supervised self-organized map [SSOM], LDA, SVM, KNN, MLE, RegTree and non-negative matrix factorization [NMF]) in combination with the Hudgins set of four TD features were compared using their obtained accuracy rates and processing time to classify ten individual hand movements. The offline performance was evaluated using the accuracy rate, training time in seconds and testing time in milliseconds (ms). To evaluate the real-time performance of the classifiers, the real-time accuracy, selection time in seconds, completion time in seconds and completion rate were used as the performance indicators. The selection time was the time between the first prediction different than “rest” and the first correct prediction. The completion time was the time between the first prediction different than “res”’ and the 20th correct prediction. The completion rate was the percentage of movements that reached 20 percent correct predictions before timeout.

Table 5.7 illustrates that among the classifiers, LDA and MLE performed significantly better (p<0.05) based on their obtained offline accuracy rates (above 97%). SSOM and KNN obtained the lowest accuracy rates, which were statistically significant (p<0.05). Among the classifiers, MLE was the fastest (0.171 s), and NMF was the slowest (46.3 s) in training; MLP was the fastest (0.115 ms) and NMF the slowest (286.6 ms) classifiers in offline testing.

Based on the real-time accuracy obtained by the motion test, MLP, LDA and MLE performed better than the other classifiers, with an accuracy rate of above 68%. MLP obtained the lowest selection time (0.657 s), and NMF obtained the highest selection time (1.99 s). The lowest completion time belongs to KNN, with an average of 3.85 seconds, and MLP has the highest completion time, with an average of 5.06 seconds. The highest completion rates were obtained by LDA and MLE, with average rates of 71.3% and 72.4%, respectively. SSOM and KNN obtained the lowest completion rates, with average rates of 42.9% and 38.9%.

The result illustrated that each classifier obtained higher offline accuracy rates in comparison with that of its real time. Some hand movements were easier to detect offline such as “Extend hand” since most of the classifiers were able to detect it with


a higher accuracy rate. Based on the obtained real-time accuracy, “Flex hand” and “Extend hand” were the easiest and “Side grip” was the hardest movement to detect in real time. By looking at the obtained accuracy rates per subject both offline and in real time, it was observed that practicing the movements improved the performance of the classifiers. During the data acquisition of one of the subjects, it was observed that despite filtering the signal with a notch filter at 50 Hz, the power line noise still affected the signal. The result showed that this subject obtained the lowest offline and real-time accuracy rates for the SSOM and KNN classifiers. Therefore, it is likely that these classifiers are not very robust against baseline noise, though it is difficult to make a definitive statement based on only one subject.

Table 5.7: Average and standard deviation of offline and real-time accuracy rates for all classifiers over 15 subjects.

Classifiers Offline Accuracy (%) Real-time Accuracy (%)

MLP 91.7b ±5.47 69.8 a ±8.16

RFN 81.0c ±5.25 60.9 bc ±10.4

SSOM 64.6e ±7.20 35.5 ef ±12.4

LDA 97.9a ±2.69 69.3 ab ±10.8

SVM 84.8c ±5.84 43.0 de ±14.8

KNN 63.1e ±6.63 33.3 f ±11.9

MLE 97.0a ±2.25 68.4 abc ±11.0

RegTree 92.2b ±4.33 59.8 c ±15.8

NMF 70.3d ±8.49 47.6 d ±11.7

a,b,… means within a column not sharing a common superscript are significantly different (p<0.05)


a higher accuracy rate. Based on the obtained real-time accuracy, “Flex hand” and “Extend hand” were the easiest and “Side grip” was the hardest movement to detect in real time. By looking at the obtained accuracy rates per subject both offline and in real time, it was observed that practicing the movements improved the performance of the classifiers. During the data acquisition of one of the subjects, it was observed that despite filtering the signal with a notch filter at 50 Hz, the power line noise still affected the signal. The result showed that this subject obtained the lowest offline and real-time accuracy rates for the SSOM and KNN classifiers. Therefore, it is likely that these classifiers are not very robust against baseline noise, though it is difficult to make a definitive statement based on only one subject.

Table 5.7: Average and standard deviation of offline and real-time accuracy rates for all classifiers over 15 subjects.

Classifiers Offline Accuracy (%) Real-time Accuracy (%)

MLP 91.7b ±5.47 69.8 a ±8.16

RFN 81.0c ±5.25 60.9 bc ±10.4

SSOM 64.6e ±7.20 35.5 ef ±12.4

LDA 97.9a ±2.69 69.3 ab ±10.8

SVM 84.8c ±5.84 43.0 de ±14.8

KNN 63.1e ±6.63 33.3 f ±11.9

MLE 97.0a ±2.25 68.4 abc ±11.0

RegTree 92.2b ±4.33 59.8 c ±15.8

NMF 70.3d ±8.49 47.6 d ±11.7

a,b,… means within a column not sharing a common superscript are significantly different (p<0.05)

Chapter 6

Conclusion

This doctoral thesis is based on surface EMG signal processing and consists of two parts: 1) improving the ECG interference removal from surface EMG signals and 2) investigating a wide range of offline and real-time myoelectric pattern recognition algorithms and proposing new configurations to improve hand movement recognition.

In the first part, the efficiencies of different artifact removal methods were studied to remove ECG interference from surface EMG signals. The investigated methods were HPF, the gating technique, spike clipping, hybrid approach, subtraction method, ICA, wavelet transform, wavelet-ICA, ANN, ANC and ANFIS. The experimental results of these methods were analyzed and compared using common evaluation criteria such as SNR, RE, CC, ET and PSD. Finally, to address the first research question and improve the filtering process of EMG signals, four novel filtering approaches, ANFIS-wavelet, ANN-wavelet, adaptive subtraction and automated wavelet-ICA, were proposed. The experimental results revealed that using the wavelet transform as the postprocessor for ANN and ANFIS produced better results (based on their SNR values) compared to the other investigated methods in this thesis. The proposed ANN-wavelet and ANFIS-wavelet achieved the highest SNR value for all data sets. These techniques are fast and could successfully remove ECG artifacts from EMG signals. The importance of these two proposed approaches lies in the removal of the ECG without distorting the EMG signal. The denoised EMG signal may be used to predict and control hand movements; hence, a low computational time is very important. The template subtraction algorithm is successful in removing ECG artifacts from EMG signals, but it is not feasible for real-time applications. I overcome this limitation by proposing an online technique, adaptive subtraction,

Chapter 6

Conclusion

This doctoral thesis is based on surface EMG signal processing and consists of two parts: 1) improving the ECG interference removal from surface EMG signals and 2) investigating a wide range of offline and real-time myoelectric pattern recognition algorithms and proposing new configurations to improve hand movement recognition.

In the first part, the efficiencies of different artifact removal methods were studied to remove ECG interference from surface EMG signals. The investigated methods were HPF, the gating technique, spike clipping, hybrid approach, subtraction method, ICA, wavelet transform, wavelet-ICA, ANN, ANC and ANFIS. The experimental results of these methods were analyzed and compared using common evaluation criteria such as SNR, RE, CC, ET and PSD. Finally, to address the first research question and improve the filtering process of EMG signals, four novel filtering approaches, ANFIS-wavelet, ANN-wavelet, adaptive subtraction and automated wavelet-ICA, were proposed. The experimental results revealed that using the wavelet transform as the postprocessor for ANN and ANFIS produced better results (based on their SNR values) compared to the other investigated methods in this thesis. The proposed ANN-wavelet and ANFIS-wavelet achieved the highest SNR value for all data sets. These techniques are fast and could successfully remove ECG artifacts from EMG signals. The importance of these two proposed approaches lies in the removal of the ECG without distorting the EMG signal. The denoised EMG signal may be used to predict and control hand movements; hence, a low computational time is very important. The template subtraction algorithm is successful in removing ECG artifacts from EMG signals, but it is not feasible for real-time applications. I overcome this limitation by proposing an online technique, adaptive subtraction,

56 Chapter 6. Conclusion

which is usable for real-time applications. To improve the performance of this proposed approach, a low-pass filter was applied to remove noise from the obtained ECG template. The wavelet-ICA has been proposed to improve the performance of the wavelet transform and the ICA method. This method is user-dependent, and a significant part of the EMG is removed while setting the noisy component to zero. In this thesis, an algorithm was proposed to convert this method to an automated approach in which the noisy component is found automatically. I also used an HPF for removing the noisy component instead of setting the noisy component to zero.

In the second part, to address the second research question, I first investigated the offline performance of a wide range of myoelectric pattern recognition algorithms (44 individual features and six classifiers) to decode 11 individual hand movements. The classification accuracy and processing time were used as the performance indicators. By applying a feature selection technique to the 44 individual features, a new set of time domain features (FS) was proposed. All the possible combinations of features and classifiers were also investigated, and new configurations were found to improve the result. The experimental result presented that the MAV, STD, WL, DAMV, IAV and FS features in the time domain in combination with the KNN classifier and FS/MLE combination are recommended to obtain higher recognition accuracy rates with a lower processing time. The LogRMS and NLE features in combination with LDA and MLP are suitable when the time consumption is not a key requirement.

Afterwards, to address the third research question, the offline and real-time performance of nine classification algorithms to decode ten individual hand movements was investigated and compared. The results revealed that there is a great difference (approximately 14%) between the highest classification accuracy rates obtained offline and in real time (97.9% vs 69.8%). LDA and MLE obtained the highest offline accuracy rates. However, in real time, MLP had the highest accuracy rate, but the differences were not statistically significant against LDA and MLE. MLP, RFN, LDA, SVM, MLE and RegTree had good performance in decoding most of the movements offline, but several movements from the same muscle group were not easy to detect in real time. NMF was the most time-consuming algorithm based on its processing time (training, testing and selection time). Consequently, MLP, LDA and MLE, with high accuracy and completion rates and low processing times, are more feasible for prosthetic control.

56 Chapter 6. Conclusion

which is usable for real-time applications. To improve the performance of this proposed approach, a low-pass filter was applied to remove noise from the obtained ECG template. The wavelet-ICA has been proposed to improve the performance of the wavelet transform and the ICA method. This method is user-dependent, and a significant part of the EMG is removed while setting the noisy component to zero. In this thesis, an algorithm was proposed to convert this method to an automated approach in which the noisy component is found automatically. I also used an HPF for removing the noisy component instead of setting the noisy component to zero.

In the second part, to address the second research question, I first investigated the offline performance of a wide range of myoelectric pattern recognition algorithms (44 individual features and six classifiers) to decode 11 individual hand movements. The classification accuracy and processing time were used as the performance indicators. By applying a feature selection technique to the 44 individual features, a new set of time domain features (FS) was proposed. All the possible combinations of features and classifiers were also investigated, and new configurations were found to improve the result. The experimental result presented that the MAV, STD, WL, DAMV, IAV and FS features in the time domain in combination with the KNN classifier and FS/MLE combination are recommended to obtain higher recognition accuracy rates with a lower processing time. The LogRMS and NLE features in combination with LDA and MLP are suitable when the time consumption is not a key requirement.

Afterwards, to address the third research question, the offline and real-time performance of nine classification algorithms to decode ten individual hand movements was investigated and compared. The results revealed that there is a great difference (approximately 14%) between the highest classification accuracy rates obtained offline and in real time (97.9% vs 69.8%). LDA and MLE obtained the highest offline accuracy rates. However, in real time, MLP had the highest accuracy rate, but the differences were not statistically significant against LDA and MLE. MLP, RFN, LDA, SVM, MLE and RegTree had good performance in decoding most of the movements offline, but several movements from the same muscle group were not easy to detect in real time. NMF was the most time-consuming algorithm based on its processing time (training, testing and selection time). Consequently, MLP, LDA and MLE, with high accuracy and completion rates and low processing times, are more feasible for prosthetic control.

Chapter 7

Future Work

Several research topics could be investigated further on the basis of the work presented in this doctoral thesis. Some of the possible research directions are as follow:

• Combining electroencephalogram signals from the scalp (reflecting underlying brain activity) with EMG signals to increase the accuracy of the classification of motor patterns, with potential applications in brain-computer interfaces.

• Exploring EMG as a biofeedback tool for rehabilitating muscle function in post stroke patients together with the electroencephalogram signal.

• Extending our results to other types of movements (e.g., discriminating finger movements).

• Replicating the results presented in this thesis in trials with actual amputees.

• Conducting long-term trials, i.e., test and train data are collected over a period of days.

• Using deep learning algorithms to decode different hand movements and comparing their results with those of machine learning algorithms.

Chapter 7

Future Work

Several research topics could be investigated further on the basis of the work presented in this doctoral thesis. Some of the possible research directions are as follow:

• Combining electroencephalogram signals from the scalp (reflecting underlying brain activity) with EMG signals to increase the accuracy of the classification of motor patterns, with potential applications in brain-computer interfaces.

• Exploring EMG as a biofeedback tool for rehabilitating muscle function in post stroke patients together with the electroencephalogram signal.

• Extending our results to other types of movements (e.g., discriminating finger movements).

• Replicating the results presented in this thesis in trials with actual amputees.

• Conducting long-term trials, i.e., test and train data are collected over a period of days.

• Using deep learning algorithms to decode different hand movements and comparing their results with those of machine learning algorithms.

Bibliography

[1] H. Rehbaum and D. Farina, "Adaptive common average filtering for myocontrol applications," Medical & Biological Engineering & Computing, vol. 53, pp. 179-186, 2015.

[2] F. Strasser, M. Muma, and A. Zoubir, "Motion artifact removal in ECG signal using multi-resolution thresholding," presented at the 20th European Signal Processing Conference (EUSIPCO), Bucharest, August, 2012.

[3] K. B. Englehart, B. S. Hudgins, P. Parker, and M. Stevenson, "Classification of the myoelectric signal using time-frequency based representations," Medical Engineering & Physics, vol. 21, no. 6-7, pp. 431-438, 1999.

[4] E. Scheme and K. B. Englehart, "Electromyogram pattern recognition for control of powered upper-limb prostheses: State of the art and challenges for clinical use," Journal of Rehabilitation Research & Development, vol. 48, no. 6, pp. 643-660, 2011.

[5] A. O. Andrade et al., "Bridging the gap between robotic technology and health care," Biomedical Signal Processing and Control, vol. 10, pp. 65-78, 2014/03/01/ 2014.

[6] E. J. Scheme, K. B. Englehart, and B. S. Hudgins, "Selective classification for improved robustness of myoelectric control under nonideal conditions," IEEE Transactions on Biomedical Engineering, vol. 58, no. 6, pp. 1698-1705, 2011.

[7] Y. Hu, J. Mak, H. Liu, and K. D. K. Luk, "ECG cancellation for surface electromyography measurement using independent component analysis," presented at the International Symposium on Circuits and Systems, New Orleans, 2007.

[8] N. Ghaeb, "Simulation study for electrocardiography contamination in surface electromyography," presented at the Biomedical Engineering, Cairo, 2008.

[9] C. J. De Luca, L. D. Gilmore, M. Kuznetsov, and S. H. Roy, "Filtering the surface EMG signal: Movement artifact and baseline noise contamination," Biomechanics, vol. 43, pp. 1573-1579, 2010.

Bibliography

[1] H. Rehbaum and D. Farina, "Adaptive common average filtering for myocontrol applications," Medical & Biological Engineering & Computing, vol. 53, pp. 179-186, 2015.

[2] F. Strasser, M. Muma, and A. Zoubir, "Motion artifact removal in ECG signal using multi-resolution thresholding," presented at the 20th European Signal Processing Conference (EUSIPCO), Bucharest, August, 2012.

[3] K. B. Englehart, B. S. Hudgins, P. Parker, and M. Stevenson, "Classification of the myoelectric signal using time-frequency based representations," Medical Engineering & Physics, vol. 21, no. 6-7, pp. 431-438, 1999.

[4] E. Scheme and K. B. Englehart, "Electromyogram pattern recognition for control of powered upper-limb prostheses: State of the art and challenges for clinical use," Journal of Rehabilitation Research & Development, vol. 48, no. 6, pp. 643-660, 2011.

[5] A. O. Andrade et al., "Bridging the gap between robotic technology and health care," Biomedical Signal Processing and Control, vol. 10, pp. 65-78, 2014/03/01/ 2014.

[6] E. J. Scheme, K. B. Englehart, and B. S. Hudgins, "Selective classification for improved robustness of myoelectric control under nonideal conditions," IEEE Transactions on Biomedical Engineering, vol. 58, no. 6, pp. 1698-1705, 2011.

[7] Y. Hu, J. Mak, H. Liu, and K. D. K. Luk, "ECG cancellation for surface electromyography measurement using independent component analysis," presented at the International Symposium on Circuits and Systems, New Orleans, 2007.

[8] N. Ghaeb, "Simulation study for electrocardiography contamination in surface electromyography," presented at the Biomedical Engineering, Cairo, 2008.

[9] C. J. De Luca, L. D. Gilmore, M. Kuznetsov, and S. H. Roy, "Filtering the surface EMG signal: Movement artifact and baseline noise contamination," Biomechanics, vol. 43, pp. 1573-1579, 2010.

60 Bibliography

[10] N. Nazmi, M. A. Abdul Rahman, S. Yamamoto, S. A. Ahmad, H. Zamzuri, and S. A. Mazlan, "A review of classification techniques of EMG signals during isotonic and isometric contractions," Sensors, vol. 16, no. 8, 2016.

[11] A. Phinyomark, F. Quaine, S. Charbonnier, C. Serviere, F. Tarpin-Bernard, and Y. Laurillau, "EMG feature evaluation for improving myoelectric pattern recognition robustness," Expert Systems with Applications, vol. 40, no. 12, pp. 4832-4840, 2013.

[12] S. Saponas, D. Tan, D. Morris, and R. Balakrishnan, "Demonstrating the feasibility of using forearm electromyography for muscle-computer interfaces," Presenetd at the SIGCHI Conference on Human Factors in Computing Systems, Italy, pp. 515-524, 2008.

[13] A. C. Tsai, T. H. Hsieh, J. J. Luh, and T. T. Lin, "A comparison of upper-limb motion pattern recognition using EMG signals during dynamic and isometric muscle contractions," Biomedical Signal Processing and Control, vol. 11, no. 1, pp. 17-26, 2014.

[14] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, "Feature reduction and selection for EMG signal classification," Expert Systems with Applications, vol. 39, no. 8, pp. 7420-7431, 2012.

[15] K. Xing, P. Yang, J. Huang, Y. Wang, and Q. Zhu, "A real-time EMG pattern recognition method for virtual myoelectric hand control," Neurocomputing, vol. 136, pp. 345-355, 2014.

[16] F. V. Tenore, A. Ramos, A. Fahmy, S. Acharya, R. Etienne-Cummings, and N. V. Thakor, "Decoding of individuated finger movements using surface electromyography," IEEE Transactions on Biomedical Engineering, vol. 56, no. 5, pp. 1427-1434, 2009.

[17] M. Redfern, R. Hughes, and D. Chaffin, "High-pass filtering to remove electrocardiographic interference from torso EMG recordings," Clinical Biomechanics, vol. 8, no. 1, pp. 44-48, 1993.

[18] H. Liang, Z. Lin, and F. Yin, "Removal of ECG contamination from diaphragmatic EMG by nonlinear filtering," Nonlinear Analysis, vol. 63, no. 5-7, pp. 745-753, 2005.

[19] J. S. Black and D. F. Lovely, "A comparison of ECG cancellation techniques applied to the surface recording of somatosensory evoked potentials," Medical Engineering & Physics, vol. 19, no. 3, pp. 279-285, 1997.

[20] B. Rhou, M. Sawan, T. Desilets, and F. Bellemare, "Real-time filtering technique to remove ECG interference from recorded esophageal EMG," presented at the Biomedical Circuits and Systems Conference, Baltimore, 2008.

60 Bibliography

[10] N. Nazmi, M. A. Abdul Rahman, S. Yamamoto, S. A. Ahmad, H. Zamzuri, and S. A. Mazlan, "A review of classification techniques of EMG signals during isotonic and isometric contractions," Sensors, vol. 16, no. 8, 2016.

[11] A. Phinyomark, F. Quaine, S. Charbonnier, C. Serviere, F. Tarpin-Bernard, and Y. Laurillau, "EMG feature evaluation for improving myoelectric pattern recognition robustness," Expert Systems with Applications, vol. 40, no. 12, pp. 4832-4840, 2013.

[12] S. Saponas, D. Tan, D. Morris, and R. Balakrishnan, "Demonstrating the feasibility of using forearm electromyography for muscle-computer interfaces," Presenetd at the SIGCHI Conference on Human Factors in Computing Systems, Italy, pp. 515-524, 2008.

[13] A. C. Tsai, T. H. Hsieh, J. J. Luh, and T. T. Lin, "A comparison of upper-limb motion pattern recognition using EMG signals during dynamic and isometric muscle contractions," Biomedical Signal Processing and Control, vol. 11, no. 1, pp. 17-26, 2014.

[14] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, "Feature reduction and selection for EMG signal classification," Expert Systems with Applications, vol. 39, no. 8, pp. 7420-7431, 2012.

[15] K. Xing, P. Yang, J. Huang, Y. Wang, and Q. Zhu, "A real-time EMG pattern recognition method for virtual myoelectric hand control," Neurocomputing, vol. 136, pp. 345-355, 2014.

[16] F. V. Tenore, A. Ramos, A. Fahmy, S. Acharya, R. Etienne-Cummings, and N. V. Thakor, "Decoding of individuated finger movements using surface electromyography," IEEE Transactions on Biomedical Engineering, vol. 56, no. 5, pp. 1427-1434, 2009.

[17] M. Redfern, R. Hughes, and D. Chaffin, "High-pass filtering to remove electrocardiographic interference from torso EMG recordings," Clinical Biomechanics, vol. 8, no. 1, pp. 44-48, 1993.

[18] H. Liang, Z. Lin, and F. Yin, "Removal of ECG contamination from diaphragmatic EMG by nonlinear filtering," Nonlinear Analysis, vol. 63, no. 5-7, pp. 745-753, 2005.

[19] J. S. Black and D. F. Lovely, "A comparison of ECG cancellation techniques applied to the surface recording of somatosensory evoked potentials," Medical Engineering & Physics, vol. 19, no. 3, pp. 279-285, 1997.

[20] B. Rhou, M. Sawan, T. Desilets, and F. Bellemare, "Real-time filtering technique to remove ECG interference from recorded esophageal EMG," presented at the Biomedical Circuits and Systems Conference, Baltimore, 2008.

Bibliography 61

[21] A. Bartolo, C. Roberts, R. Dzwonczyk, and E. Goldman, "Analysis of diaphragm EMG signal: comparison of gating vs. subtraction for removal of ECG contamination," Applied Physiology, vol. 80, no. 6, pp. 1898-1902, 1996.

[22] S. C. Ragupathy, D. K. Kumar, B. Polus, and K. Kamei, "Electrocardiogram removal from electromyogram of the lumbar muscles," presented at the 26th Conference on Engineering in Medicine and Biology Society, 2004.

[23] J. N. F. Mak, Y. Hu, and K. D. K. Luk, "ICA-based ECG removal from surface electromyography and its effect on low back pain assessment," presented at the 3rd International IEEE/EMBS Conference on Neural Engineering, Kohala Coast, 2007.

[24] S. R. Alty, W. D. Man, J. Moxham, and K. C. Lee, "Denoising of diaphragmatic electromyogram signals for respiratory control and diagnostic purposes," presented at the 30th International Conference on Engineering in Medicine and Biology Society, 2008.

[25] P. Zhou, M. Lowery, R. Weir, and T. Kuiken, "Elimination of ECG artifacts from myoelectric prosthesis control signals developed by targeted muscle reinnervation," presented at the 27th Conference on Engineering in Medicine and Biology, Shanghai, September, 2005.

[26] E. Aithocine, B. Guméry, S. Meignen, L. Heyer, Y. Lavault, and S. B. Gottfried, "Contribution to Structural intensity tool: application to the cancellation of ECG interference in diaphragmatic EMG," presented at the 28th International Conference on Engineering in Medicine and Biology Society, 2006.

[27] C. Zhan, L. F. Yeung, and Z. Yang, "A wavelet-based adaptive filter for removing ECG interference in EMGdi signals," Electromyography and Kinesiology, vol. 20, pp. 542-549, 2010.

[28] J. Tealman, S. Van Huffel, and A. Spaepen, "Wavelet-independent component analysis to remove electrocardiography contamination in surface electromyography," presented at the 29th International Conference on Engineering in Medicine and Biology Society, United States, 2007.

[29] C. K. S. Vijila and C. E. S. Kumar, "Cancellation of ECG in electromyogram using back Propagation network," presented at the International Conference on advances in recent technologies in communication and computing, Kottayam, Kerala, 2009.

[30] S. Abbaspour, A. Fallah, and A. Maleki, "A comparison of adaptive filter and artificial neural network results in removing electrocardiogram contamination

Bibliography 61

[21] A. Bartolo, C. Roberts, R. Dzwonczyk, and E. Goldman, "Analysis of diaphragm EMG signal: comparison of gating vs. subtraction for removal of ECG contamination," Applied Physiology, vol. 80, no. 6, pp. 1898-1902, 1996.

[22] S. C. Ragupathy, D. K. Kumar, B. Polus, and K. Kamei, "Electrocardiogram removal from electromyogram of the lumbar muscles," presented at the 26th Conference on Engineering in Medicine and Biology Society, 2004.

[23] J. N. F. Mak, Y. Hu, and K. D. K. Luk, "ICA-based ECG removal from surface electromyography and its effect on low back pain assessment," presented at the 3rd International IEEE/EMBS Conference on Neural Engineering, Kohala Coast, 2007.

[24] S. R. Alty, W. D. Man, J. Moxham, and K. C. Lee, "Denoising of diaphragmatic electromyogram signals for respiratory control and diagnostic purposes," presented at the 30th International Conference on Engineering in Medicine and Biology Society, 2008.

[25] P. Zhou, M. Lowery, R. Weir, and T. Kuiken, "Elimination of ECG artifacts from myoelectric prosthesis control signals developed by targeted muscle reinnervation," presented at the 27th Conference on Engineering in Medicine and Biology, Shanghai, September, 2005.

[26] E. Aithocine, B. Guméry, S. Meignen, L. Heyer, Y. Lavault, and S. B. Gottfried, "Contribution to Structural intensity tool: application to the cancellation of ECG interference in diaphragmatic EMG," presented at the 28th International Conference on Engineering in Medicine and Biology Society, 2006.

[27] C. Zhan, L. F. Yeung, and Z. Yang, "A wavelet-based adaptive filter for removing ECG interference in EMGdi signals," Electromyography and Kinesiology, vol. 20, pp. 542-549, 2010.

[28] J. Tealman, S. Van Huffel, and A. Spaepen, "Wavelet-independent component analysis to remove electrocardiography contamination in surface electromyography," presented at the 29th International Conference on Engineering in Medicine and Biology Society, United States, 2007.

[29] C. K. S. Vijila and C. E. S. Kumar, "Cancellation of ECG in electromyogram using back Propagation network," presented at the International Conference on advances in recent technologies in communication and computing, Kottayam, Kerala, 2009.

[30] S. Abbaspour, A. Fallah, and A. Maleki, "A comparison of adaptive filter and artificial neural network results in removing electrocardiogram contamination

62 Bibliography

from surface EMGs," presented at the Proceeding of 20th Iranian Conference on Electrical Engineering, 2012.

[31] J. D. Z. Chen, Z. Y. Lin, M. Ramahi, and R. K. Mittal, "Adaptive cancellation of ECG artifacts in the diaphragm electromyographic signals obtained through intraoesophageal electrodes during swallowing and inspiration," Neurogastroenterology & Motility, vol. 6, no. 4, pp. 279-288, 1994.

[32] C. Marque, C. Bisch, R. Dantas, S. Elayoubi, V. Brosse, and C. Perot, "Adaptive filtering for ECG rejection from surface EMG recordings," Electromyography and Kinesiology, vol. 15, no. 3, pp. 310-315, 2005.

[33] G. Lu et al., "Removing ECG noise from surface EMG signals using adaptive filtering," Neuroscience Letters, vol. 462, no. 1, pp. 14-19, 2009.

[34] K. h. Assaleh, "Adaptive neuro-fuzzy inference systems for extracting fetal electrocardiogram," presented at the International Symposium on Signal Processing and Information Technology, 2006.

[35] C. K. S. Vijila and C. E. S. Kumar, "Interference cancellation in EMG signal using ANFIS," Recent Trends in Engineering, vol. 2, no. 5, pp. 244-248, 2009.

[36] R. Swarnalatha and D. V. Prasad, "Maternal ECG cancellation in abdominal signal using ANFIS and wavelets," applied science, vol. 10, pp. 868-877, 2010.

[37] S. Abbaspour, A. Fallah, and A. Maleki, "A comparison of adaptive neuro-fuzzy inference system and real-time filtering in cancellation ECG artifact from surface EMGs," presented at the Proceeding of 20th Iranian Conference on Electrical Engineering, 2012.

[38] S. Abbaspour and A. Fallah, "A combination method for electrocardiogram rejection from surface electromyogram," Open Biomedical Engineering, vol. 8, pp. 13-19, 2014.

[39] S. Abbaspour, H. Gholamhosseini, and M. Linden, "Evaluation of wavelet based methods in removing motion artifact from ECG signal," presented at the 16th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics, Gotenburg, 2014.

[40] M. Zecca, S. Micera, M. C. Carrozza, and P. Dario, "Control of multifunctional prosthetic hands by processing the electromyographic signal," Critical Reviews in Biomedical Engineering, vol. 30, no. 4-6, pp. 459-485, 2002.

[41] S. Guo, M. Pang, B. Gao, H. Hirata, and H. shihara, "Comparison of sEMG-based feature extraction and motion classification methods for upper-limb movement," Sensors, vol. 15, pp. 9022-9038, 2015.

62 Bibliography

from surface EMGs," presented at the Proceeding of 20th Iranian Conference on Electrical Engineering, 2012.

[31] J. D. Z. Chen, Z. Y. Lin, M. Ramahi, and R. K. Mittal, "Adaptive cancellation of ECG artifacts in the diaphragm electromyographic signals obtained through intraoesophageal electrodes during swallowing and inspiration," Neurogastroenterology & Motility, vol. 6, no. 4, pp. 279-288, 1994.

[32] C. Marque, C. Bisch, R. Dantas, S. Elayoubi, V. Brosse, and C. Perot, "Adaptive filtering for ECG rejection from surface EMG recordings," Electromyography and Kinesiology, vol. 15, no. 3, pp. 310-315, 2005.

[33] G. Lu et al., "Removing ECG noise from surface EMG signals using adaptive filtering," Neuroscience Letters, vol. 462, no. 1, pp. 14-19, 2009.

[34] K. h. Assaleh, "Adaptive neuro-fuzzy inference systems for extracting fetal electrocardiogram," presented at the International Symposium on Signal Processing and Information Technology, 2006.

[35] C. K. S. Vijila and C. E. S. Kumar, "Interference cancellation in EMG signal using ANFIS," Recent Trends in Engineering, vol. 2, no. 5, pp. 244-248, 2009.

[36] R. Swarnalatha and D. V. Prasad, "Maternal ECG cancellation in abdominal signal using ANFIS and wavelets," applied science, vol. 10, pp. 868-877, 2010.

[37] S. Abbaspour, A. Fallah, and A. Maleki, "A comparison of adaptive neuro-fuzzy inference system and real-time filtering in cancellation ECG artifact from surface EMGs," presented at the Proceeding of 20th Iranian Conference on Electrical Engineering, 2012.

[38] S. Abbaspour and A. Fallah, "A combination method for electrocardiogram rejection from surface electromyogram," Open Biomedical Engineering, vol. 8, pp. 13-19, 2014.

[39] S. Abbaspour, H. Gholamhosseini, and M. Linden, "Evaluation of wavelet based methods in removing motion artifact from ECG signal," presented at the 16th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics, Gotenburg, 2014.

[40] M. Zecca, S. Micera, M. C. Carrozza, and P. Dario, "Control of multifunctional prosthetic hands by processing the electromyographic signal," Critical Reviews in Biomedical Engineering, vol. 30, no. 4-6, pp. 459-485, 2002.

[41] S. Guo, M. Pang, B. Gao, H. Hirata, and H. shihara, "Comparison of sEMG-based feature extraction and motion classification methods for upper-limb movement," Sensors, vol. 15, pp. 9022-9038, 2015.

Bibliography 63

[42] M. A. Oskoei and H. Hu, "GA-based feature subset selection for myoelectric classification," in IEEE International Conference on Robotics and Biomimetics, China, pp. 1465-1470, 2006.

[43] B. S. Hudgins, P. Parker, and R. N. Scott, "The recognition of myoelectric patterns for prosthetic limb control," in Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 13, no. 5, pp. 2040-2041, 1991.

[44] L. J. Hargrove, K. B. Englehart, and B. S. Hudgins, "A comparison of surface and intramuscular myoelectric signal classification," IEEE Transactions on Biomedical Engineering, vol. 54, no. 5, pp. 847-853, 2007.

[45] A. Phinyomark and E. Scheme, "An investigation of temporally inspired time domain features for electromyographic pattern recognition," in the 40th International Conference of the IEEE Engineering in Medicine and Biology, Honolulu, HI, pp. 5236-5240, 2018.

[46] A. A. Al-Taee and A. Al-Jumaily, "Optimal feature set for finger movement classification based on sEMG," in the 40th International Conference of the IEEE Engineering in Medicine and Biology Honolulu, HI, pp. 5228-5231, 2018.

[47] P. Shenoy, K. J. Miller, B. Crawford, and R. P. N. Roa, "Online electromyographic control of a robotic prosthesis," IEEE Transaction on Biomedical Engineering, vol. 55, no. 3, pp. 1128-1135, 2008.

[48] F. Wang, M. A. Oskoei, and H. Hu, "Multi-finger myoelectric signals for controlling a virtual robotic prosthetic hand," International Journal of Modeling, Identification and Control, vol. 27, no. 3, pp. 181-190, 2017.

[49] F. Sebelius, M. Axelsson, N. Danielson, J. Schouenborg, and T. Laurell, "Real-time control of a virtual hand," Technology and Disability, vol. 17, no. 3, pp. 131-141, 2005.

[50] M. Ortiz-Catalan, R. Brånemark, and B. Håkansson, "BioPatRec: a modular research platform for the control of artificial limbs based on pattern recognition algorithms," Source Code for Biology and Medicine, vol. 8, no. 11, 2013.

[51] M. Ortiz-Catalan, B. Håkansson, and R. Brånemark, "Real-time and simultaneous control of artificial limbs based on pattern recognition algorithms," IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 22, no. 4, pp. 756-764, 2014.

[52] Y. geng, A. M. Sabatini, R. Weir, and G. Li, "Improving the robustness of real-time myoelectric pattern recognition against arm position changes in

Bibliography 63

[42] M. A. Oskoei and H. Hu, "GA-based feature subset selection for myoelectric classification," in IEEE International Conference on Robotics and Biomimetics, China, pp. 1465-1470, 2006.

[43] B. S. Hudgins, P. Parker, and R. N. Scott, "The recognition of myoelectric patterns for prosthetic limb control," in Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 13, no. 5, pp. 2040-2041, 1991.

[44] L. J. Hargrove, K. B. Englehart, and B. S. Hudgins, "A comparison of surface and intramuscular myoelectric signal classification," IEEE Transactions on Biomedical Engineering, vol. 54, no. 5, pp. 847-853, 2007.

[45] A. Phinyomark and E. Scheme, "An investigation of temporally inspired time domain features for electromyographic pattern recognition," in the 40th International Conference of the IEEE Engineering in Medicine and Biology, Honolulu, HI, pp. 5236-5240, 2018.

[46] A. A. Al-Taee and A. Al-Jumaily, "Optimal feature set for finger movement classification based on sEMG," in the 40th International Conference of the IEEE Engineering in Medicine and Biology Honolulu, HI, pp. 5228-5231, 2018.

[47] P. Shenoy, K. J. Miller, B. Crawford, and R. P. N. Roa, "Online electromyographic control of a robotic prosthesis," IEEE Transaction on Biomedical Engineering, vol. 55, no. 3, pp. 1128-1135, 2008.

[48] F. Wang, M. A. Oskoei, and H. Hu, "Multi-finger myoelectric signals for controlling a virtual robotic prosthetic hand," International Journal of Modeling, Identification and Control, vol. 27, no. 3, pp. 181-190, 2017.

[49] F. Sebelius, M. Axelsson, N. Danielson, J. Schouenborg, and T. Laurell, "Real-time control of a virtual hand," Technology and Disability, vol. 17, no. 3, pp. 131-141, 2005.

[50] M. Ortiz-Catalan, R. Brånemark, and B. Håkansson, "BioPatRec: a modular research platform for the control of artificial limbs based on pattern recognition algorithms," Source Code for Biology and Medicine, vol. 8, no. 11, 2013.

[51] M. Ortiz-Catalan, B. Håkansson, and R. Brånemark, "Real-time and simultaneous control of artificial limbs based on pattern recognition algorithms," IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 22, no. 4, pp. 756-764, 2014.

[52] Y. geng, A. M. Sabatini, R. Weir, and G. Li, "Improving the robustness of real-time myoelectric pattern recognition against arm position changes in

64 Bibliography

transradial amputees," BioMed Research International, vol. 2017, pp. 1-10, 2017.

[53] S. Micera, J. Carpaneto, and S. Raspopovic, "Control of hand prostheses using peripheral information," IEEE Reviews in Biomedical Engineering, vol. 3, pp. 48-68, 2010.

[54] J. D. M. Drake and J. P. Callaghan, "Elimination of electrocardiogram contamination from electromyogram signals: an evaluation of currently used removal techniques," Electromyography and Kinesiology, vol. 16, no. 2, pp. 175–187, 2006.

[55] J. Pan and W. Tompkins, "A real-time QRS detection algorithm," IEEE Transaction on Biomedical Engineering, vol. 32, pp. 230-236, 1985.

[56] P. S. Hamilton and W. Tompkins, "Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database," IEEE Transaction on Biomedical Engineering, vol. 33, pp. 1157-1165, 1986.

[57] J. Nurminen, "Electrophysiological studies of neonatal sleep stage," presented at the Proceeding of International Conference on advances in recent technologies in communication and computing, 2009.

[58] J. N. F. Mak, Y. Hu, and K. D. K. Luk, "An automated ECG-artifact removal method for trunk muscle surface EMG recordings," Medical Engineering & Physics vol. 8, pp. 840-848, 2010.

[59] P. Zhou and T. A. Kuiken, "Eliminating cardiac contamination from myoelectric control signals developed by targeted muscle reinnervation," Physiological Measurment, vol. 27, pp. 1311-1327, 2006.

[60] I. Daubechies, Ten lectures on wavelets. CBMS, SIAM, 1994. [61] J. Taelman, B. Mijovic, S. Van Huffel, S. Devuyst, and T. Dutoit, "ECG

artifact removal from surface EMG signals by combining emperical mode decomposition and independent component analysis," presented at the In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, 2011.

[62] M. Akay, Biomedical signal processing. San Diego: Academic press, 1994. [63] M. G. Bellanger, Adaptive digital filters. New York: Marcel Dekker, 2001. [64] E. A. Wan, "Adjoint LMS: an efficient alternative to the filtered-x LMS and

multiple error LMS algorithms," presented at the Proceeding of International Conference on Acoustics, Speech and Signal Processing 1996.

[65] M. A. Shoorehdeli, M. Teshnehlab, and A. Sedigh, "Training ANFIS as an identifier with intelligent hybrid stable learning algorithm based on particle swarm optimization and extended Kalman filter," Fuzzy Sets and System, vol. 160, pp. 922-948, 2009.

64 Bibliography

transradial amputees," BioMed Research International, vol. 2017, pp. 1-10, 2017.

[53] S. Micera, J. Carpaneto, and S. Raspopovic, "Control of hand prostheses using peripheral information," IEEE Reviews in Biomedical Engineering, vol. 3, pp. 48-68, 2010.

[54] J. D. M. Drake and J. P. Callaghan, "Elimination of electrocardiogram contamination from electromyogram signals: an evaluation of currently used removal techniques," Electromyography and Kinesiology, vol. 16, no. 2, pp. 175–187, 2006.

[55] J. Pan and W. Tompkins, "A real-time QRS detection algorithm," IEEE Transaction on Biomedical Engineering, vol. 32, pp. 230-236, 1985.

[56] P. S. Hamilton and W. Tompkins, "Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database," IEEE Transaction on Biomedical Engineering, vol. 33, pp. 1157-1165, 1986.

[57] J. Nurminen, "Electrophysiological studies of neonatal sleep stage," presented at the Proceeding of International Conference on advances in recent technologies in communication and computing, 2009.

[58] J. N. F. Mak, Y. Hu, and K. D. K. Luk, "An automated ECG-artifact removal method for trunk muscle surface EMG recordings," Medical Engineering & Physics vol. 8, pp. 840-848, 2010.

[59] P. Zhou and T. A. Kuiken, "Eliminating cardiac contamination from myoelectric control signals developed by targeted muscle reinnervation," Physiological Measurment, vol. 27, pp. 1311-1327, 2006.

[60] I. Daubechies, Ten lectures on wavelets. CBMS, SIAM, 1994. [61] J. Taelman, B. Mijovic, S. Van Huffel, S. Devuyst, and T. Dutoit, "ECG

artifact removal from surface EMG signals by combining emperical mode decomposition and independent component analysis," presented at the In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, 2011.

[62] M. Akay, Biomedical signal processing. San Diego: Academic press, 1994. [63] M. G. Bellanger, Adaptive digital filters. New York: Marcel Dekker, 2001. [64] E. A. Wan, "Adjoint LMS: an efficient alternative to the filtered-x LMS and

multiple error LMS algorithms," presented at the Proceeding of International Conference on Acoustics, Speech and Signal Processing 1996.

[65] M. A. Shoorehdeli, M. Teshnehlab, and A. Sedigh, "Training ANFIS as an identifier with intelligent hybrid stable learning algorithm based on particle swarm optimization and extended Kalman filter," Fuzzy Sets and System, vol. 160, pp. 922-948, 2009.

Bibliography 65

[66] W. M. B. Wan Daud, A. B. Yahya, C. S. Horng, M. F. Sulaima, and R. Sudirman, "Features extraction of electromyography signals in time domain on biceps brachii muscle " International Journal of Modeling and Optimization, vol. 3, no. 6, pp. 515-519, 2013.

[67] S. P. Arjunan and D. K. Kumar, "Decoding subtle forearm flexions using fractal features of surface Electromyogram from single and multiple sensors," NeuroEngineering and Rehabilitation, vol. 7, no. 53, 2010.

[68] B. S. Hudgins, P. Parker, and R. N. Scott, "A new strategy for multifunction myoelectric control," IEEE Transactions on Biomedical Engineering, vol. 40, no. 1, pp. 82-94, 1993.

[69] M. Zardoshti-Kermani, B. C. Wheeler, K. Badie, and R. M. Hashemi, "EMG feature evaluation for movement control of upper extremity prostheses," IEEE Transactions on Rehabilitation Engineering, vol. 3, no. 4, pp. 324-333, 1995.

[70] M. Ortiz-Catalan, R. Brånemark, and B. Håkansson, "Biologically inspired algorithms applied to prosthetic control," in BioMed, Austria, pp. 35-44, 2012.

[71] X. Tang, Y. Liu, C. Lv, and D. Sun, "Hand motion classification using a multi-channel surface electromyography sensor," Sensors, vol. 12, no. 2, pp. 1130-1147, 2012.

[72] S. Yu, E. Jeong, K. Hong, and S. Lee, "Classification of nine directions using the maximum likelihood estimation based on electromyogram of both forearms," Biomedical Engineering Letters, vol. 2, no. 2, pp. 129-137, 2012.

[73] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, "Fractal analysis features for weak and single-channel upper-limb EMG signals," Expert Systems with Applications, vol. 39, no. 12, pp. 11156-11163, 2012.

[74] L. Zoubek, S. Charbonnier, S. Lesecq, A. Buguet, and F. Chapotot, "Feature selection for sleep/wake stages classification using data driven methods," Biomedical Signal Processing and Control, vol. 2, no. 3, pp. 171-179, 2007.

[75] S. Micera, A. M. Sabatini, P. Dario, and B. Rossi, "A hybrid approach to EMG pattern analysis for classification of arm movements using statistical and fuzzy techniques," Medical Engineering & physics, vol. 21, no. 5, pp. 303-311, 1999.

[76] C. Vidaurre, N. Kramer, B. Blankertz, and A. Schlogl, "Time domain parameters as a feature for EEG-based brain–computer interfaces," Neural Network, vol. 22, no. 9, pp. 1313-1319, 2009.

[77] S. H. Oh, Y. R. Lee, and H. N. Kim, "A novel EEG Feature extraction method using hjorth parameter," International Journal of Electronics and Electrical Engineering, vol. 2, no. 2, pp. 106-110, 2014.

Bibliography 65

[66] W. M. B. Wan Daud, A. B. Yahya, C. S. Horng, M. F. Sulaima, and R.Sudirman, "Features extraction of electromyography signals in time domainon biceps brachii muscle " International Journal of Modeling andOptimization, vol. 3, no. 6, pp. 515-519, 2013.

[67] S. P. Arjunan and D. K. Kumar, "Decoding subtle forearm flexions usingfractal features of surface Electromyogram from single and multiple sensors,"NeuroEngineering and Rehabilitation, vol. 7, no. 53, 2010.

[68] B. S. Hudgins, P. Parker, and R. N. Scott, "A new strategy for multifunctionmyoelectric control," IEEE Transactions on Biomedical Engineering, vol. 40,no. 1, pp. 82-94, 1993.

[69] M. Zardoshti-Kermani, B. C. Wheeler, K. Badie, and R. M. Hashemi, "EMGfeature evaluation for movement control of upper extremity prostheses," IEEETransactions on Rehabilitation Engineering, vol. 3, no. 4, pp. 324-333, 1995.

[70] M. Ortiz-Catalan, R. Brånemark, and B. Håkansson, "Biologically inspiredalgorithms applied to prosthetic control," in BioMed, Austria, pp. 35-44,2012.

[71] X. Tang, Y. Liu, C. Lv, and D. Sun, "Hand motion classification using amulti-channel surface electromyography sensor," Sensors, vol. 12, no. 2, pp.1130-1147, 2012.

[72] S. Yu, E. Jeong, K. Hong, and S. Lee, "Classification of nine directions usingthe maximum likelihood estimation based on electromyogram of bothforearms," Biomedical Engineering Letters, vol. 2, no. 2, pp. 129-137, 2012.

[73] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, "Fractal analysisfeatures for weak and single-channel upper-limb EMG signals," ExpertSystems with Applications, vol. 39, no. 12, pp. 11156-11163, 2012.

[74] L. Zoubek, S. Charbonnier, S. Lesecq, A. Buguet, and F. Chapotot, "Featureselection for sleep/wake stages classification using data driven methods,"Biomedical Signal Processing and Control, vol. 2, no. 3, pp. 171-179, 2007.

[75] S. Micera, A. M. Sabatini, P. Dario, and B. Rossi, "A hybrid approach toEMG pattern analysis for classification of arm movements using statisticaland fuzzy techniques," Medical Engineering & physics, vol. 21, no. 5, pp.303-311, 1999.

[76] C. Vidaurre, N. Kramer, B. Blankertz, and A. Schlogl, "Time domainparameters as a feature for EEG-based brain–computer interfaces," NeuralNetwork, vol. 22, no. 9, pp. 1313-1319, 2009.

[77] S. H. Oh, Y. R. Lee, and H. N. Kim, "A novel EEG Feature extraction methodusing hjorth parameter," International Journal of Electronics and ElectricalEngineering, vol. 2, no. 2, pp. 106-110, 2014.

66 Bibliography

[78] K. S. Kim, H. H. Choi, C. S. Moon, and C. W. Mun, "Comparison of k-nearest neighbor, quadratic discriminant and linear discriminant analysis in classification of electromyogram signals based on the wrist-motion directions," Current Applied Physics, vol. 11, no. 3, pp. 740-745, 2011.

[79] Application Note, 2010. [80] J. Rafiee, M. A. Rafiee, F. Yavari, and M. P. Schoen, "Feature extraction of

forearm EMG signals for prosthetics," Expert Systems with Applications, vol. 38, no. 4, pp. 4.58-4067, 2011.

[81] R. Boostani and M. M.H., "Evaluation of the forearm EMG signal features for the control of a prosthetic hand," Physiological Measurement, vol. 24, no. 2, pp. 309-319, 2003.

[82] H. M. Al-Angari, G. Kanitz, S. Tarantino, and C. Cipriani, "Distance and mutual information methods for EMG feature and channel subset selection for classification of hand movements," Biomedical Signal Processing and Control, vol. 27, pp. 24-31, 2016.

[83] F. Duan, L. Dai, W. Chang, Z. Chen, C. Zhu, and W. Li, "sEMG-based identification of hand motion commands using wavelet neural network combined with discrete wavelet transform," IEEE Transactions on Industrial Electronics, vol. 63, no. 3, pp. 1923-1934, 2016.

[84] M. Khezri and M. Jahed, "A neuro–fuzzy inference system for sEMG-based identification of hand motion commands," IEEE Transactions on Industrial Electronics, vol. 58, no. 5, pp. 1952-1960, 2011.

[85] R. N. Khushaba, S. Kodagoda, S. Lal, and G. Dissanayake, "Driver drowsiness classification using fuzzy wavelet-packet-based feature-extraction algorithm," IEEE Transactions on Biomedical Engineering, vol. 58, no. 1, pp. 121-131, 2011.

[86] G. Wang, Z. Wang, W. Chen, and J. Zhuang, "Classification of surface EMG signals using optimal wavelet packet method based on Davies-Bouldin criterion," Medical and Biological Engineering and Computing, vol. 44, no. 10, pp. 865-872, 2006.

[87] E. Mastinu, M. Ortiz-Catalan, and B. Håkansson, "Low-cost, open source bioelectric signal acquisition system," in 14th International Conference on Wearable and Implantable Body Sensor Networks (BSN), Eindhoven, Netherlands, 2017.

[88] S. Haykin, Neural Networks: A Comprehensive Foundation, 2 ed. NJ, USA: Prentice Hall: Upper Saddle River, 1999.

66 Bibliography

[78] K. S. Kim, H. H. Choi, C. S. Moon, and C. W. Mun, "Comparison of k-nearest neighbor, quadratic discriminant and linear discriminant analysis inclassification of electromyogram signals based on the wrist-motiondirections," Current Applied Physics, vol. 11, no. 3, pp. 740-745, 2011.

[79] Application Note, 2010.[80] J. Rafiee, M. A. Rafiee, F. Yavari, and M. P. Schoen, "Feature extraction of

forearm EMG signals for prosthetics," Expert Systems with Applications, vol.38, no. 4, pp. 4.58-4067, 2011.

[81] R. Boostani and M. M.H., "Evaluation of the forearm EMG signal features forthe control of a prosthetic hand," Physiological Measurement, vol. 24, no. 2,pp. 309-319, 2003.

[82] H. M. Al-Angari, G. Kanitz, S. Tarantino, and C. Cipriani, "Distance andmutual information methods for EMG feature and channel subset selection forclassification of hand movements," Biomedical Signal Processing andControl, vol. 27, pp. 24-31, 2016.

[83] F. Duan, L. Dai, W. Chang, Z. Chen, C. Zhu, and W. Li, "sEMG-basedidentification of hand motion commands using wavelet neural networkcombined with discrete wavelet transform," IEEE Transactions on IndustrialElectronics, vol. 63, no. 3, pp. 1923-1934, 2016.

[84] M. Khezri and M. Jahed, "A neuro–fuzzy inference system for sEMG-basedidentification of hand motion commands," IEEE Transactions on IndustrialElectronics, vol. 58, no. 5, pp. 1952-1960, 2011.

[85] R. N. Khushaba, S. Kodagoda, S. Lal, and G. Dissanayake, "Driverdrowsiness classification using fuzzy wavelet-packet-based feature-extractionalgorithm," IEEE Transactions on Biomedical Engineering, vol. 58, no. 1, pp.121-131, 2011.

[86] G. Wang, Z. Wang, W. Chen, and J. Zhuang, "Classification of surface EMGsignals using optimal wavelet packet method based on Davies-Bouldincriterion," Medical and Biological Engineering and Computing, vol. 44, no.10, pp. 865-872, 2006.

[87] E. Mastinu, M. Ortiz-Catalan, and B. Håkansson, "Low-cost, open sourcebioelectric signal acquisition system," in 14th International Conference onWearable and Implantable Body Sensor Networks (BSN), Eindhoven,Netherlands, 2017.

[88] S. Haykin, Neural Networks: A Comprehensive Foundation, 2 ed. NJ, USA:Prentice Hall: Upper Saddle River, 1999.

Bibliography 67

[89] D. Ballabio, M. Vasighi, and P. Filzmoser, "Effects of supervised Self Organising Maps parameters on classification performance," Analytica Chimica Acta, vol. 765, pp. 45-53, 2013.

[90] S. Shin, R. Tafreshi, and R. Langari, "A performance comparison of hand motion EMG classification," in Middle East Conference on Biomedical Engineering (MECBME), Qatar, pp. 353-356, 2014.

[91] R. N. Khushaba, S. Kodagoda, D. Liu, and G. Dissanayake, "Electromyogram (EMG) based fingers movement recognition using neighborhood preserving analysis with QR-decomposition," in 7th International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), Australia, 2011.

[92] P. Rani, N. Sarkar, and C. Liu, "Maintaining optimal challenge in computer games through real-time physiological feedback," in 1st Internasional Conference on Augmented Cognition, Las Vegas, pp. 22-27, 2005.

[93] P. Kokol, M. Mernik, J. Zavrsnik, K. Kancler, and I. Maleiæ, "Decision trees based on automatic learning and their use in cardiology " Medical Systems, vol. 9, no. 4, pp. 201-206, 1994.

[94] Y. Li and A. Ngom, "The non-negative matrix factorization toolbox for biological data mining," Source Code for Biology and Medicine, vol. 8, no. 10, pp. 1-15, 2013.

[95] Y. Li and A. Ngom, "Classification approach based on non-negative least squares," Neurocomputing, vol. 118, pp. 41-57, 2013.

[96] K. T. Sweeney, T. E. Ward, and S. F. McLoone, "Artifact removal in physiological signals-practices and possibilities," IEEE Transaction on Information Technology in Biomedicine, vol. 16, pp. 488-500, 2012.

[97] F. R. Hashim, J. Soraghan, L. Petropoulakis, and S. I. Safie, "Wavelet based motion artifact removal for ECG signals," presented at the IEEE EMBS International Conference on Biomedical Engineering and Sciences (IECBES), Langkawi, December, 2012.

[98] P. He, G. Wilson, and C. Russell, "Removal of ocular artifacts from electro-encephalogram by adaptive filtering," Medical & Biological Engineering & Computing, vol. 42, no. 3, pp. 407-412, 2004.

[99] D. F. Stegeman and H. J. Hermens, Recommendation for Sensor Location on Individual Muscles, Institute of Neurology, Department of Clinical Neurophysiology, http://www.seniam.org, 2007. [Online]. Available.

[100] R. Merletti, "Standards for reporting EMG data," Electromyography and Kinesiology, vol. 9, no. 1, pp. 3-4, 1999.

Bibliography 67

[89] D. Ballabio, M. Vasighi, and P. Filzmoser, "Effects of supervised SelfOrganising Maps parameters on classification performance," AnalyticaChimica Acta, vol. 765, pp. 45-53, 2013.

[90] S. Shin, R. Tafreshi, and R. Langari, "A performance comparison of handmotion EMG classification," in Middle East Conference on BiomedicalEngineering (MECBME), Qatar, pp. 353-356, 2014.

[91] R. N. Khushaba, S. Kodagoda, D. Liu, and G. Dissanayake, "Electromyogram(EMG) based fingers movement recognition using neighborhood preservinganalysis with QR-decomposition," in 7th International Conference onIntelligent Sensors, Sensor Networks and Information Processing (ISSNIP),Australia, 2011.

[92] P. Rani, N. Sarkar, and C. Liu, "Maintaining optimal challenge in computergames through real-time physiological feedback," in 1st InternasionalConference on Augmented Cognition, Las Vegas, pp. 22-27, 2005.

[93] P. Kokol, M. Mernik, J. Zavrsnik, K. Kancler, and I. Maleiæ, "Decision treesbased on automatic learning and their use in cardiology " Medical Systems,vol. 9, no. 4, pp. 201-206, 1994.

[94] Y. Li and A. Ngom, "The non-negative matrix factorization toolbox forbiological data mining," Source Code for Biology and Medicine, vol. 8, no.10, pp. 1-15, 2013.

[95] Y. Li and A. Ngom, "Classification approach based on non-negative leastsquares," Neurocomputing, vol. 118, pp. 41-57, 2013.

[96] K. T. Sweeney, T. E. Ward, and S. F. McLoone, "Artifact removal inphysiological signals-practices and possibilities," IEEE Transaction onInformation Technology in Biomedicine, vol. 16, pp. 488-500, 2012.

[97] F. R. Hashim, J. Soraghan, L. Petropoulakis, and S. I. Safie, "Wavelet basedmotion artifact removal for ECG signals," presented at the IEEE EMBSInternational Conference on Biomedical Engineering and Sciences (IECBES),Langkawi, December, 2012.

[98] P. He, G. Wilson, and C. Russell, "Removal of ocular artifacts from electro-encephalogram by adaptive filtering," Medical & Biological Engineering &Computing, vol. 42, no. 3, pp. 407-412, 2004.

[99] D. F. Stegeman and H. J. Hermens, Recommendation for Sensor Location onIndividual Muscles, Institute of Neurology, Department of ClinicalNeurophysiology, http://www.seniam.org, 2007. [Online]. Available.

[100] R. Merletti, "Standards for reporting EMG data," Electromyography andKinesiology, vol. 9, no. 1, pp. 3-4, 1999.

electromyogram signal enhancement and upper-limb...

Documents