high density emg based estimation of lower limb muscle

58
IN DEGREE PROJECT MEDICAL ENGINEERING, SECOND CYCLE, 30 CREDITS , STOCKHOLM SWEDEN 2021 High density EMG based estimation of lower limb muscle characteristics using feature extraction BALÁZS SZABÓ KTH ROYAL INSTITUTE OF TECHNOLOGY SCHOOL OF ENGINEERING SCIENCES IN CHEMISTRY, BIOTECHNOLOGY AND HEALTH

Upload: others

Post on 05-May-2022

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: High density EMG based estimation of lower limb muscle

IN DEGREE PROJECT MEDICAL ENGINEERING,SECOND CYCLE, 30 CREDITS

, STOCKHOLM SWEDEN 2021

High density EMG based estimation of lower limb muscle characteristics using feature extraction

BALÁZS SZABÓ

KTH ROYAL INSTITUTE OF TECHNOLOGYSCHOOL OF ENGINEERING SCIENCES IN CHEMISTRY, BIOTECHNOLOGY AND HEALTH

Page 2: High density EMG based estimation of lower limb muscle
Page 3: High density EMG based estimation of lower limb muscle

High density EMG basedestimation of lower limbmuscle characteristics usingfeature extraction

BALÁZS SZABÓ

MSc in Medical EngineeringDate: January 29, 2021Supervisor: Ruoli WangExaminer: Svein KleivenSchool of Engineering Sciences in Chemistry, Biotechnology andHealthSwedish title: Uppskattning av nedre extremiteternasmuskelegenskaper med högdensitets-EMG och funktionsextraktion

Page 4: High density EMG based estimation of lower limb muscle
Page 5: High density EMG based estimation of lower limb muscle

iii

AbstractElectromyography (EMG) is a common tool in electrical muscle activitymeasurement and can be used in multiple areas of clinical and biomedicalapplications, mainly in identifying neuromuscular diseases, analyzingmovement or in human machine interfaces. Traditionally a pair of electrodeswere used to measure the signals, but in recent years the use of high densitysurface EMG (HD-sEMG) gained more popularity as it can samplemyoelectric activities from multiple electrodes in an array on a single muscleand provide more information.

In this thesis a measurement setup and protocol is proposed that canprovide a reliably measurement, furthermore multiple features are extractedfrom the collected signals to characterise the major muscles around theankle. 5 healthy subjects were tested using an ankle dynamometer with 5HD-sEMG placed on the Tibialis Anterior, the Gastrocnemius Medialis, theSoleus, the Gastrocnemius Lateralis, and on the Peroneus Longus. Severaltests were conducted using different initial angle of the ankle joint anddifferent percentages of the maximum voluntary contraction. The reliabilityof the setup was assessed by comparing the variance between the collectedsignals of the same subject in a repeated test, and by comparing differentsubjects to each other. Results show a reasonably good reliability with lessthan 10% variance, and adequate selectivity as well. To examine the musclecharacteristics, 7 features were extracted from the collected and processedsignals, then the features were plotted and compared to signs for musclecharacteristics such as muscle fatigue, activation, and spatial distribution ofactivation. Correlations between features of mean average value (MAV) andzero crossing (ZC), and different muscle characteristics could be observed.

Page 6: High density EMG based estimation of lower limb muscle

iv

SammanfattningElektromyografi (EMG) är ett vanligt verktyg för mätning av elektriskmuskelaktivitet och kan användas inom flera områden av kliniska ochbiomedicinska tillämpningar, främst för att identifiera neuromuskulärasjukdomar, analysera rörelse eller i mänskliga maskingränssnitt. Traditionelltanvändes ett par elektroder för att mäta signalerna, men de senaste åren fickanvändningen av EMG med hög densitet (HD-sEMG) mer populariteteftersom det kan mäta myoelektriska aktiviteter från flera elektroder på enenda muskel och ge mer information.

I denna avhandling föreslås en mätinställning och ett protokoll som kange en pålitlig mätning. Dessutom extraheras flera funktioner från de samladesignalerna för att karakterisera de viktigaste musklerna runt fotleden. Femfriska försökspersoner testades med en fotledsdynamometer med 5HD-sEMG placerad på Tibialis Anterior, Gastrocnemius Medialis, Soleus,Gastrocnemius Lateralis och på Peroneus Longus. Flera tester utfördes medanvändning av olika initialvinklar på fotleden och olika procentsatser avmaximal frivillig sammandragning. Tillförlitligheten för installationenbedömdes genom att jämföra variansen mellan de insamlade signalerna frånsamma patient i ett upprepat test och genom att jämföra olika patienter medvarandra. Resultaten visar en ganska bra tillförlitlighet med mindre än 10%

avvikelse och adekvat selektivitet. För att undersöka muskelegenskapernaextraherades 7 funktioner från de samlade och bearbetade signalerna, sedanplottades funktionerna och jämfördes med tecken för muskelegenskaper sommuskeltrötthet, aktivering och rumslig fördelning av aktivering.Korrelationer mellan egenskaperna för medelvärdet (MAV) och zero crossing(ZC) och olika muskelegenskaper kunde observeras.

Page 7: High density EMG based estimation of lower limb muscle

Acknowledgement

This master thesis was performed at KTH MoveAbility Lab, School ofEngineering Sciences in Chemistry, Biotechnology and Health during thesecond year of the masters’ degree program in medical engineering at KTHRoyal Institute of Technology. I would like to take the opportunity to thankeveryone who helped and supported me during this time. First of all, I amvery thankful to Ruoli Wang for accepting me as a thesis student and givingme the opportunity to work with such an interesting topic, and for guidanceand support during the whole thesis. I want to thank Enrico Merlo from OTBioelettronica for technical support for the measuring devices and FedericaAresu for helping with the measurements. Finally, I want to thank my familyand friends for supporting me during my time at Royal Institute ofTechnology.

v

Page 8: High density EMG based estimation of lower limb muscle

Contents

1 Introduction 1

2 Methods 32.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . 32.2 Experiment protocol . . . . . . . . . . . . . . . . . . . . . . 52.3 Signal processing . . . . . . . . . . . . . . . . . . . . . . . . 62.4 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . 92.5 Evaluation of the two objectives . . . . . . . . . . . . . . . . 10

3 Results 113.1 Measurement reliability . . . . . . . . . . . . . . . . . . . . . 113.2 Muscle characteristics . . . . . . . . . . . . . . . . . . . . . . 12

4 Discussion 174.1 Measurement reliability . . . . . . . . . . . . . . . . . . . . . 174.2 Muscle characteristics . . . . . . . . . . . . . . . . . . . . . . 184.3 Limitations and future improvements . . . . . . . . . . . . . . 20

5 Conclusions 21

Bibliography 22

Appendices 24A State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . 25B Measurement protocol . . . . . . . . . . . . . . . . . . . . . 44

vi

Page 9: High density EMG based estimation of lower limb muscle

Introduction

Surface electromyography (sEMG) have been used widely to extract globalfeatures of myoelectrical activities and utilized in multiple areas of clinicaland biomedical applications, mainly in identifying neuromuscular diseases,analyzing movement or in human machine interfaces [1, 2]. The mostcommon application in human movement is to infer the global level ofmuscle activation as the amplitude or power of the sEMG signals usingbipolar electrodes. However, these global features are not really reliable forinvestigating neural signals that muscles receive because these features areinfluenced by the properties of the active motor units’ action potentials, e.g.,muscle condition and subject anatomy. Also, it measures superimposedmotor unit action potentials (MUAPs) from several motor units.

Instead of using a pair of electrodes, high density surface EMG(HD-sEMG) can sample myoelectric activities from multiple electrodes withlinear or bi-dimensional arrays on a single muscle. HD-sEMG has beenproposed to increase the amount and reliability of information extracted fromsEMG [3, 4] and it opens the possibility for measuring spatial muscleactivity besides temporal activity, as Drost et al. [5] explains, which allowsfor investigating new muscle characteristics. For instance, it can extractanatomical and physiological information at the muscle or at the motor unitlevel with potential application in neuromuscular physiology and control ofprosthetic device. Understanding the link of myoelectrical activities withjoint function, such as kinematics and kinetics are particularly attractive forclinician and researchers because the knowledge can be widely applied indifferent fields such as human machine interaction and neuro rehabilitationlike with stroke patients who, as a result of stroke, frequently can experiencespasticity [6, 7], an extensive activity in a muscle caused by neuronaldamage.

The aim of this master thesis is to create a reliable measurement setupand protocol that facilitates the measurement of the lower leg muscles during

1

Page 10: High density EMG based estimation of lower limb muscle

2 CHAPTER 1. INTRODUCTION

isometric contraction using multiple HD-sEMG electrodes. The secondobjective of the thesis is to extract feature from the collected signals that canbe used to characterise the investigated muscles and their contractions.

Page 11: High density EMG based estimation of lower limb muscle

Methods

2.1 Experimental setupThere are several different configurations in which EMG can be measureddepending on what kind of information is of interest. For this project, themuscles of the lower leg were examined especially the ones that have thehighest contribution to the movement around ankle joint [8]. The muscles inparticular are the Gastrocnemius Medialis, the Soleus, the Tibialis Anterior,the Gastrocnemius Lateralis and the Peroneus Longus and their location ofthe lower limb can be seen in Figure 2.1.

Figure 2.1: Skeletal muscles of the lower leg. (modified based on [8])

To measure the activation of the mentioned muscles five HD-sEMGelectrode grids were placed on the leg, corresponding to the standard EMG

3

Page 12: High density EMG based estimation of lower limb muscle

4 CHAPTER 2. METHODS

electrode placement position for each muscle based on recommendationsfrom SENIAM [9]. Three 64 channel HD-sEMG grids for the GastrocnemiusMedialis, Soleus and Tibialis Anterior, and two 32 channel HD-sEMG gridwere used for the Gastrocnemius Lateralis and the Peroneus Longus,respectively as seen in Figure 2.2. For signal acquisition OT-Bioelettronica’sQuattrocento [10] bioelectrical amplifier were used which offers analoguesignal amplification, filtration, and analogue to digital signal conversionintegrated in a single device, furthermore it is capable of handling all of thetotal 256 channels used in the experiment, simultaneously. All signals werecollected in a monopolar configuration with a sampling frequency of 2048Hz, using a reference point (without muscle activation) around the ankle forthe Soleus, Tibialis Anterior and Peroneus Longus, and another referencepoint around the knee for the Gastrocnemius Medialis and Lateralis.

Figure 2.2: Placement of the HD-sEMG electrode grids on the test subject.

One of the biggest challenges in a measurement involving an isometricmuscle contraction, especially an isometric contraction in the lower part of theleg, is to keep the body part, and therefore the muscles, in question fixed. Forthis reason, an ankle dynamometer, also provided by OT-Bioelettronica, wereslightly modified with additional straps and used to minimize the movementof the leg. The ankle dynamometer was also equipped with a single axis forcemeter to record the generated force by the muscle contractions, furthermore itallowed to set the initial angle of the ankle in both in positive (dorsiflexion)and negative (plantarflexion) direction. The full experimental setup can beseen on Figure 2.3 below.

Page 13: High density EMG based estimation of lower limb muscle

CHAPTER 2. METHODS 5

Figure 2.3: Complete measurement setup.

2.2 Experiment protocolIn this project 5 healthy people were tested with one repeated test on one ofthe test subjects (subject no. 1). Before each test, relevant physicalinformation was collected which is summarized in Table 2.1 below.

Table 2.1: Summary of completed test and relevant physical information.

Test no. 1 2 3 4 5 6Subject no. 1 2 1a 3 4 5Age 38 28 38 26 25 26Gender Female Female Female Male Female MaleHeight (cm) 162 164 162 181 162 176Weight (kg) 53 53 53 86 53 78Physical Fitness Level Moderate Moderate Moderate Moderate Moderate High

arepeated test

Each test constituted of 4 different ankle angles setting in randomised order:a neutral setting with 0◦; two setting in plantarflexion with −7.5◦ and −15◦;and one setting in dorsiflexion with 10◦. Within each setting both dorsi- and

Page 14: High density EMG based estimation of lower limb muscle

6 CHAPTER 2. METHODS

plantarflexion were examined, and in each direction first the maximumvoluntary contraction (MVC) was determined by the subject exertingmaximal force in the respective direction for 5 seconds and repeating it threetimes with a few seconds of rest in between, and then averaging the threemaximal force measured in the three consecutive runs. Then within eachdirection 3 runs were made where the subject was asked to follow atrapezoidal path with a 5 seconds ramp up, 4 seconds holding at therespective percentages (70%, 50%, 30%) of the MVC, and 5 seconds rampdown time. Then this trapezoidal path was repeated 5 times with 10 secondsof rest between them in each run. The trapezoidal path is visualized inFigure 2.4 and the full protocol can be found in Appendix B.

Figure 2.4: Example trapezoidal path with 50% MVC that the test subjectsneed to follow during measurement.

2.3 Signal processingCollected EMG signals tend to be contaminated by some noise mainlyoriginating from motion artifact and interference form the power lines [11,12]. A common practice to remove this noise is to use a band pass filter with

Page 15: High density EMG based estimation of lower limb muscle

CHAPTER 2. METHODS 7

a lower threshold of 10 Hz and a high threshold of 500 Hz, furthermore, touse a notch filter at 50 Hz to eliminate the effects of the interference [13, 14]Since the Quattrocento already have an integrated bandpass filter that can betuned to the required frequencies, only the notch filter was applied to thedigital signal during signal processing.

To estimate the quality of the collected signal, for each channel in allHD-sEMG electrode grids the signal to noise ratio (SNR) was calculated byseparating the collected signal in every channel into two parts: a signal part(Ssig) that corresponds to the contraction phase during the time the subjectfollows the trapezoidal path, and a noise part (Snoise) which is measuredduring the resting phases. The separation of the signal can be seen onFigure 2.5b. Then then SNR was calculated based on (2.1) [15].

SNR = 10 · log10(

RMSE(Ssig)

RMSE(Snosie)

)(2.1)

Where RMSE(Ssig) and RMSE(Snoise) are the root mean squared error ofthe signal and noise respectively.

All signals were collected in monopolar configuration and in order tofurther remove the common noise and to reduce the effect of crosstalk ofneighbouring muscles, differential EMG signals were calculated from twoadjacent electrode by taking the amplitude difference between the electrodepair. The conversion form monopolar to differential signal was made in sucha way that the resulting differential signal matrix would follow the directionof the underlying muscle fibers, as can be seen in Figure 2.5a.

Page 16: High density EMG based estimation of lower limb muscle

8 CHAPTER 2. METHODS

(a) Differential Calculation

(b) SNR calculation

Figure 2.5: Explanation for converting monopolar signals to differential signal(a) and calculation of signal to noise ration (b).

Page 17: High density EMG based estimation of lower limb muscle

CHAPTER 2. METHODS 9

Signals, originating from the same muscle group but from differentsubjects or even the same muscle on the same person but comparing twodifferent measurement in time, can have diverse characteristics, mainly inamplitude [16, 17]. To get signals that have more uniform characteristics andbe able to compare them better, all differential EMG signals were normalizedto the maximum amplitude measured during MVC run, respective to themuscle, direction and angle setting.

2.4 Feature extractionTo investigate the characteristics of the muscles during the contractions 7commonly used feature were extracted [18–20]. Before the features couldhave been calculated, the signal needed to be segmented in order to get acontinuous feature signal. For this, a siding window method was used with a500 ms wide window and 20% overlap [20, 21]. In each columns of allHD-sEMG electrode grids, one channel with the highest signal to noise ratio(SNR) was selected, which represents different muscle fibers in therespective muscles, and used for further processing. The extracted featuresand their equations can be seen below in (2.2-2.8):

• Integrated EMG (IEMG)

IEMG =N∑i=1

|xi| (2.2)

where xi is the differential EMG signal.

• Mean Average Variance (MAV)

MAV =1

N

N∑i=1

|xi| (2.3)

• Root Mean Square (RMS)

RMS =

√√√√ 1

N

N∑i=1

x2i (2.4)

• Variance (VAR)

V AR =1

N − 1

N∑i=1

x2i (2.5)

Page 18: High density EMG based estimation of lower limb muscle

10 CHAPTER 2. METHODS

• Wilson Amplitude (WAMP)

WAMP =N−1∑i=1

sign (|xi+1 − xi| − t) (2.6)

where t is a threshold of 0.05.

• Wavelet Length (WL)

WL =N−1∑i=1

|xi+1 − xi| (2.7)

• Zero-Crossing (ZC)

ZC =N−1∑i=1

sign (xi+1 · (−xi)) (2.8)

2.5 Evaluation of the two objectivesWith the help of the collected and processed signals, the two objectives of thethesis could be tested. To determine the reliability of the experimental setup,the variation between two measurements done in different times but on thesame subject is compared, and also the variation between the different subjectsis investigated.

For the second objective, to find a correlation between the extractedfeatures and muscle characteristics, the collected signals and the extractedfeatures were plotted and compared to signs for factors such as musclefatigue, activation, and spatial distribution of activation.

Page 19: High density EMG based estimation of lower limb muscle

Results

3.1 Measurement reliabilityTo investigate the reliability of the measurement, the signals collected indifferent times and in different test subjects were compared. Test 1 withsubject 1 was used as a base for comparison and the signal from the channelwith the highest SNR was selected in the major flexor muscles, the TibiaAnterior for dorsiflexion, the Gastrocnemius Medialis and the Soleus forplantarflexion. The SNR for the channels in each muscles is visualized byFigure 3.1.

(a) Tibia Anterior (b) Gastrocnemius Medialis (c) Soleus

Figure 3.1: Signal to noise ratio (SNR) visualized for each channel in the TibiaAnterior (a), the Gastrocnemius Medialis (b) and the Soleus (c) during MVC.

To assess the repeatability of the measurement, in Table 3.1 the relativevariance of the absolute deviation between the base test 1 and test 3 issummarised, i.e. the difference between two measurements done on the samesubject with the same measurement setup and protocols, but in differenttimes. The amplitude of the signals used for the calculations werenormalised beforehand, and the resulted relative variance of the absolutedeviation showed as a percentage of the maximal normalised amplitude.

11

Page 20: High density EMG based estimation of lower limb muscle

12 CHAPTER 3. RESULTS

Table 3.1: Shows the result for the repeatability test. Relative variance of theabsolute deviation was calculated between two test that was done on the samesubject but in different times.

Relative variance of the absolute deviationTest 3

Subject 1Tibia Anterior 1%Gastrocnemius Medialis 9%Soleus 7%

Table 3.2 shows a similar relative variance of deviation, but here 4 differentsubjects were compared to the base test. As before, the results are showed asa percentage of the maximal normalised amplitude.

Table 3.2: Shows the result for the selectivity test. Relative variance of theabsolute deviation was calculated between subject no. 1 and the other foursubjects.

Relative variance of the absolute deviationTest 2 Test 4 Test 5 Test 6

Subject 2 Subject 3 Subject 4 Subject 5Tibia Anterior 7% 8% 3% 11%Gastrocnemius Medialis 10% 14% 10% 13%Soleus 11% 12% 7% 8%

3.2 Muscle characteristicsIn Figure 3.2 two features out of the seven, MAV and ZC are plotted over time.Only these two are shows, since they are commonly used and the other fivefeatures show similar results. The features were extracted from the signalsof the Tibia Anterior collected during isometric dorsiflexion at 70% of theMVC. The same test is plotted and compared for the 4 different angles settingwhich allows for the examination of the effect of individual initial stretchingof the muscles. The MAV on Figure 3.2a reflects on the average amplitude ofthe muscle activation signals in different cases, while the ZC on Figure 3.2bindicates a frequency based change in the activation.

Page 21: High density EMG based estimation of lower limb muscle

CHAPTER 3. RESULTS 13

(a) MAV (b) ZC

Figure 3.2: Features MAV (a) and ZC (b) plotted over time extracted fromTibia Anterior during isometric dorsiflexion with 70% MVC in 4 differentpositions.

In Figure 3.3 the same features are plotted, except here the GastrocnemiusMedialis was tested during isometric plantarflexion.

(a) MAV (b) ZC

Figure 3.3: Features MAV (a) and ZC (b) plotted over time extracted fromGastrocnemius Medialis during isometric plantarflexion with 70% MVC in 4different positions.

In Figure 3.4 ZC feature extracted from the signals of the GastrocnemiusMedialis are showed as well. Here, instead of the different angles, the runswith various percentages of theMVC are compared fromwhich the connectionbetween the frequency of the signal and the generated force can be inferred.

Page 22: High density EMG based estimation of lower limb muscle

14 CHAPTER 3. RESULTS

Figure 3.4: Feature ZC plotted over time extracted from GastrocnemiusMedialis during isometric plantarflexion in −15◦ position with differentpercentages of MVC.

The following figures show so called “heatmaps”, a visualization of themeasured average differential EMG signal in each electrode of the gridduring the corresponding maximal contraction. This can provide animpression about the spatial distribution and magnitude of activation in themuscles over the measured area. In Figure 3.5 and Figure 3.6 heatmap of theGastrocnemius Medialis and Soleus presented respectively, in an increasingorder of percentage of MVC for the same subject during isometricplantarflexion in a neutral 0◦ angle. Figure 3.7 and Figure 3.8 shows theheatmaps of the Soleus and the Tibia Anterior for the 5 different test subjectsduring the tests with identical settings of 70%MVC and 0◦ angle.

(a) 30% MVC (b) 50%MVC (c) 70% MVC

Figure 3.5: Muscle activations in the Gastrocnemius Medialis duringisometric plantarflexion using 30%, 50%, and 70% of the MVC.

Page 23: High density EMG based estimation of lower limb muscle

CHAPTER 3. RESULTS 15

(a) 30%MVC (b) 50% MVC (c) 70%MVC

Figure 3.6: Muscle activations in the Soleus during isometric plantarflexionusing 30%, 50%, and 70% of the MVC.

(a) Test subject 1 (b) Test subject 2 (c) Test subject 3

(d) Test subject 4 (e) Test subject 5

Figure 3.7: Visualization of the spatial muscle activation in the Soleus of thefive different subjects during plantarflexion. In the tests 70% MVC and 0◦

angle were used.

Page 24: High density EMG based estimation of lower limb muscle

16 CHAPTER 3. RESULTS

(a) Test subject 1 (b) Test subject 2 (c) Test subject 3

(d) Test subject 4 (e) Test subject 5

Figure 3.8: Visualization of the spatial muscle activation in the Tibia Anteriorof the five different subjects during isometric dorsiflexion. In the tests 70%MVC and 0◦ angle were used.

Page 25: High density EMG based estimation of lower limb muscle

Discussion

4.1 Measurement reliabilityCreating a reliable measurement setup and protocol is important forinvestigating EMG signals when different muscles, subjects are tested, andoften in different times. The proposed experimental setup aims to addressthis requirement by using multiple adjustable straps that can hold the footand leg in place for the duration of the experiment. The proper fixation of theinvestigated limb serves multi purpose: it ensures that the leg is in the samepre-determined position in all of the measurements; stops the movement ofthe limb during muscle contractions which is essential in an experimentinvolving isometric contractions; also by keeping the leg in a stable positionit can help minimise the noise originating from motion artefacts duringmeasurements, which is a common issue in most of sEMG basedexperiments [22, 23].

Table 3.1 shows the result for a repeatability test. Among the 3 majorflexor muscles that was investigated, the Tibia Anterior shows a great resultof an average of 1% of difference between the base and repeated test, whichmeans that the collected signals are rather similar if the same configurationsare used in experiment setup even if the measurement is done several daysapart. The other two muscles which responsible for plantar flexion showsslightly worse results in terms of repeatability. One reason for this could bethe close location of these muscles relative to each other, and the fact thatthey are partially overlapping, compared to the Tibia Anterior which ismostly laying on top a bone and not as tightly surrounded with othermuscles. This proximity of the Gastrocnemius Medialis and Soleus canincrease the chance of cross-talk [24]. Additionally, the skin around the TibiaAnterior is generally thinner and contains less adipose tissue than theposterior side of the lower leg, which can also influence the effect of

17

Page 26: High density EMG based estimation of lower limb muscle

18 CHAPTER 4. DISCUSSION

cross-talk according to Solomonow et al. [25]. Considering the abovementioned factors, and other possible noise sources, the relative variance ofthe absolute deviation is still under 10% which indicates that the repeatabilityof the experiment setup and protocol are reasonably good.

To address the selectivity of the setup, the measurements of 4 differentsubjects were compared to test subject 1 as can be seen in Table 3.2. To get anacceptable selectivity, i.e. to be able to distinguish themuscle characteristics inthe collected signals for different subjects, the relative variance of the absolutedeviation for test subject 2-5 should be higher compared to the repeated testwith test subject 1. This condition is fulfilled with test subject 2, 3, and 5,furthermore subject 4 shows similar characteristics to the base line subject1. Beside the variation from subject 1, the 4 other subjects also show somedissimilarity compared to each other, which also indicates that the system iscapable of detecting differences between subjects.

Although the tested measurement setup shows promising results in termof repeatability and selectivity, it is important to note that the number of testsdone for reproducing the experiments, and also the number of test subject usedfor comparison is relatively low and probably not enough to reach the requiredstatistical significance. In order to give a confident answer about the reliability,more test need to be conducted.

4.2 Muscle characteristicsOn Figure 3.2a the relationship between the muscle activation in the TibiaAnterior during isometric dorsiflexion and the initial position of the foot, i.e.the initial stretching of the muscle, can be seen. The extracted MAV feature,which inversely related to the generated force in a way that the samegenerated force achieved with higher muscle activation, shows a linearconnection with the angles. By positioning the ankle in more negative anglesmean that the muscle fibers in the Tibia Anterior stretch more and inconsequence the magnitude of the extracted feature increases as well. Thisincrease seems proportional to the stretching with the muscle apart from themeasurement done with 10◦. However, considering the findings of Garcia,Dueweke, and Mendias [26] where they determined at 5◦ angle can the TibialAnterior produce the highest tension in isometric contraction, it can explainthis discrepancy. The explanation behind this is that according to the generalmuscle length vs tension plot, the relationship between them is not perfectlylinear, but the generated tension also decreases if the muscle inunder-stretched, as in the case at 10◦. In contrast, on Figure 3.3a the same

Page 27: High density EMG based estimation of lower limb muscle

CHAPTER 4. DISCUSSION 19

feature from the Gastrocnemius Medialis can be seen, where the correlationbetween angles and the feature is not that visible. Here, since plantarflexionis examined, the more negative angles mean more under-stretched muscleswhich can be a reason for the MAV of the different angles being close to eachother. Another explanation could be that during plantarflexion more majorprotagonist muscle are active at the same time, requiring one muscle togenerate less force and this way the difference between different angles is notas considerable. In Figure 3.2b and Figure 3.3b the feature ZC, whichcorrelates to the frequency of the activation signal as a higher number of ZCover the same time means higher frequency, is displayed where a similarconclusion can be drawn as for the MVC.

In Figure 3.4 the feature ZC extracted from Gastrocnemius Medialisplotted over time during isometric plantarflexion in -15◦ position withdifferent percentages of MVC. During the 70% and 50% MVC a slightincrease in the number of ZC, i.e. the frequency of the signal, can beobserved over the five peaks, which corresponds to the five repetition of thetrapezoid path during one measurement run. This increase in frequencycould infer a sign for muscle fatigue, as producing and holding the sametension during isometric contraction over time would require a higher stimulirate in a fatiguing muscle if the length and other factors are kept the same.On the other hand, this indication for fatigue can not be detected during 30%

MVC which could be rationalized by that the required tension for reaching30% of MVC cause insignificant fatigue or it requires more time toexperience muscle fatigue that the recorded time frame.

One of the biggest advantages of using HD-sEMG over regular EMG isthat it can provide additional information like the spatial distribution ofmuscle activity during contraction. This can be a helpful tool to assess thecondition of the muscles in patients, identify damaged or compromised areasof the muscles amongst others. In Figure 3.5 and Figure 3.6 this property ofthe HD-sEMG is exploited in the form of heatmaps. The figures show theheatmaps of the Gastrocnemius Medialis and Soleus of the same patientrespectively, during isometric plantarflexion using 30%, 50%, and 70% of theMVC. Here, the recruitment of more and more muscle fibers can be observedas a higher muscle tension is produced. In Figure 3.7 and Figure 3.8 theheatmaps for the Soleus and the Tibia Anterior is displayed for the five testsubjects. This can also provide useful information for the individualcharacteristics of the muscle, as the location of maximal activation and thegeneral distribution of activations can be identified.

Page 28: High density EMG based estimation of lower limb muscle

20 CHAPTER 4. DISCUSSION

4.3 Limitations and future improvementsBased on the presented results the proposed experimental setup seemspromising, however it is important to take into consideration the limitationsof the system and ways to improve it. Despite great efforts, the fixation of theleg into the dynamometer was not perfect which is important in isometriccontractions, since the foot was able to move a few millimetres in some cases,especially during plantarflexion where the subjects are able to generate highforces. Additionally, in order to get a good contact between the HD-sEMGelectrode grids and the skin and get a reliable signal, they need to be preparedand applied with great care, otherwise it can cause some loss of signalquality or additional noise. Finally, the number of subjects tested was onlyenough to give an indication for reliability of the setup, but it needs furthertesting to give a more confident conclusion. To improve on the reliability ofthe experiments and the results, the mentioned issues should be addressed.

Page 29: High density EMG based estimation of lower limb muscle

Conclusions

The aim of this master thesis was to create a reliable measurement setup andprotocol that facilitates the measurement of the lower leg muscles duringisometric contraction using multiple HD-sEMG electrodes, and to extractfeatures from the collected signals that can be used to characterise muscles.The conducted test and presented results show promising result in terms ofreliability, as the relative variance of the absolute deviation does not exceed10% when a subject is compared themselves, and the selectivity seemsacceptable as well. From the collected signals multiple features wereextracted and identified as a possible indicator for muscle characteristics suchas muscle fatigue, activation, and spatial distribution of activation.

21

Page 30: High density EMG based estimation of lower limb muscle

Bibliography

[1] Michael Wehner. “Man to Machine, Applications inElectromyography”. In: EMG Methods for Evaluating Muscle andNerve Function. InTech, Jan. 2012.

[2] Yunfen Wu, Mara ngeles Martnez Martnez, and Pedro OrizaolaBalaguer. “Overview of the Application of EMG Recording in theDiagnosis and Approach of Neurological Disorders”. In:Electrodiagnosis in New Frontiers of Clinical Research. InTech, May2013.

[3] Stegeman Dick F. et al. “High-density Surface EMG: Techniques andApplications at aMotor Unit Level”. In:Biocybernetics and BiomedicalEngineering 32.3 (Jan. 2012), pp. 3–27.

[4] Antonietta Stango, Francesco Negro, and Dario Farina. “SpatialCorrelation of High Density EMG Signals Provides Features Robustto Electrode Number and Shift in Pattern Recognition forMyocontrol”. In: IEEE Transactions on Neural Systems andRehabilitation Engineering 23.2 (Mar. 2015), pp. 189–198.

[5] Gea Drost et al. Clinical applications of high-density surface EMG: Asystematic review. Dec. 2006.

[6] Kyung Eun Nam et al. “When does spasticity in the upper limbdevelop after a first stroke? A nationwide observational study on 861stroke patients”. In: Journal of Clinical Neuroscience 66 (Aug. 2019),pp. 144–148.

[7] C L Watkins et al. “Prevalence of spasticity post stroke.” In: Clinicalrehabilitation 16.5 (Aug. 2002), pp. 515–22.

[8] OpenStax. Anatomy and Physiology. OpenStax CNX, Feb. 2016.

[9] SENAIM. Recommendations for sensor locations in lower leg or footmuscles.

22

Page 31: High density EMG based estimation of lower limb muscle

BIBLIOGRAPHY 23

[10] OT Bioelettronica. Quattrocento.

[11] E. A. Clancy, E. L. Morin, and R. Merletti. “Sampling,noise-reduction and amplitude estimation issues in surfaceelectromyography”. In: Journal of Electromyography andKinesiology. Vol. 12. 1. Elsevier, Feb. 2002, pp. 1–16.

[12] M. B.I. Reaz, M. S. Hussain, and F. Mohd-Yasin. “Techniques of EMGsignal analysis: Detection, processing, classification and applications”.In: Biological Procedures Online 8.1 (Mar. 2006), pp. 11–35.

[13] Mojtaba Malboubi, Farbod Razzazi, and Mahdi Aliyari Shoorehdeli.“Elimination of power line noise from EMG signals using an efficientadaptive Laguerre filter”. In: International Conference on Signals andElectronic Systems, ICSES’10 - Conference Proceeding (Oct. 2010),pp. 49–52.

[14] S. D.H. Soedirdjo, K. Ullah, and R. Merletti. “Power line interferenceattenuation in multi-channel sEMG signals: Algorithms and analysis”.In: Proceedings of the Annual International Conference of the IEEEEngineering in Medicine and Biology Society, EMBS.Vol. 2015-November. Institute of Electrical and Electronics EngineersInc., Nov. 2015, pp. 3823–3826.

[15] S Abbaspour and A Fallah. “Removing ECG Artifact from the SurfaceEMG Signal Using Adaptive Subtraction Technique.” In: Journal ofbiomedical physics & engineering 4.1 (Mar. 2014), pp. 33–8.

[16] Manuela Besomi et al. “Consensus for experimental design inelectromyography (CEDE) project: Amplitude normalization matrix”.In: Journal of Electromyography and Kinesiology 53 (Aug. 2020),p. 102438.

[17] Walaa M. Elsais et al. “Between-day repeatability of lower limb EMGmeasurement during running and walking”. In: Journal ofElectromyography and Kinesiology 55 (Dec. 2020).

[18] A. Phinyomark, C. Limsakul, and P. Phukpattaranont. “EMG featureextraction for tolerance of white Gaussian noise”. In: InternationalWorkshop and Symposium Science Technology. Nong Khai, Thailand,2008.

[19] Angkoon Phinyomark, Rami N Khushaba, and Erik Scheme. “FeatureExtraction and Selection for Myoelectric Control Based on WearableEMG Sensors.” In: Sensors (Basel, Switzerland) 18.5 (May 2018).

Page 32: High density EMG based estimation of lower limb muscle

24 BIBLIOGRAPHY

[20] Christopher Spiewak. “A Comprehensive Study on EMG FeatureExtraction and Classifiers”. In: Open Access Journal of BiomedicalEngineering and Biosciences 1.1 (Feb. 2018).

[21] Eugenio C. Orosco, Natalia M. Lopez, and Fernando Di Sciascio.“Bispectrum-based features classification for myoelectric control”. In:Biomedical Signal Processing and Control 8.2 (Mar. 2013),pp. 153–168.

[22] Hak W. Tam and John G. Webster. “Minimizing Electrode MotionArtifact by Skin Abrasion”. In: IEEE Transactions on BiomedicalEngineering BME-24.2 (1977), pp. 134–139.

[23] Lin Xu et al. “Use of power-line interference for adaptive motionartifact removal in biopotential measurements”. In: PhysiologicalMeasurement 37.1 (Jan. 2016), pp. 25–40.

[24] Taian Martins Vieira et al. “Specificity of surface EMG recordings forgastrocnemius during upright standing”. In: Scientific Reports 7.1 (Dec.2017).

[25] M. Solomonow et al. “Surface and wire EMG crosstalk in neighbouringmuscles”. In: Journal of Electromyography and Kinesiology 4.3 (1994),pp. 131–142.

[26] Stefan C. Garcia, Jeffrey J. Dueweke, and Christopher L. Mendias.“Optimal Joint Positions for Manual Isometric Muscle Testing”. In:Journal of Sport Rehabilitation 25.4 (Jan. 2016).

Page 33: High density EMG based estimation of lower limb muscle

Appendices

A State of the art

25

Page 34: High density EMG based estimation of lower limb muscle

Background

A.1 IntroductionElectromyography (EMG) is a common tool in electrical muscle activitymeasurement and can be used in multiple areas of clinical and biomedicalapplications, mainly in identifying neuromuscular diseases, analyzingmovement or in human machine interfaces [1, 2]. Traditionally a pair ofelectrodes were used to measure the signals, but in recent years the use ofHigh Density EMG (HD-EMG) gained more popularity, since it opens thepossibility for measuring spatial muscle activity besides temporal activity, asDrost et al. [3] explains, which allows for new muscle characteristicsdetection.

One of the applications of EMG is with stroke patients who, as a result ofstroke, frequently can experience spasticity [4, 5], which is an extensiveactivity in a muscle caused by neuronal damage. Spasticity and otherdisturbances in the regulation of muscle tone following a stroke can lead tocertain disabilities, it can for example disrupt the ability to walking if thespasticity happens in the lower limbs.

Electromyography (EMG) has been used in rehabilitation, especiallyamongst stroke patients, but the effectiveness of some of these treatmentmethods show moderate improvements [6]. However, in general a pair ofelectrodes are used in the treatment, which provides unreliable signals.Instead, High Density EMG (HD-EMG) has been proposed to increase theamount and reliability of information extracted from surface EMG (sEMG)which can be used widely in different fields [7–9].

1

Page 35: High density EMG based estimation of lower limb muscle

2 APPENDIX A. BACKGROUND

A.2 Biomechanics of the ankleProper muscle function is essential for many daily activities, especially formore mechanically demanding movements such as standing, walking or sit tostand movement. An impaired musculoskeletal system can limit a person indaily activities and degree of independence and therefore decrease the qualityof life [10]. Functional biomechanics of the ankle is essential for normalstability, especially in weightbearing positions like standing or walking.During a gait cycle the main movements of the ankle are plantarflexion anddorsiflexion. These movements, represented in Figure A.1, are controlled bythe muscles located in the lower leg and in the following sections theanatomical and physiological background of the basic muscle movements isexplained based on the book Anatomy and Physiology [11].

Figure A.1: Ankle movements in the saggital plane: plantar and dorsiflexion[11].

A.3 Lower leg musclesThe muscles in the lower leg can be grouped by their role in a biomechanicalpoint of view: muscles in the posterior compartment of the leg createplantarflexion, while muscles in the anterior part of the leg producedorsiflexion. The main plantarflexors (gastrocnemius, soleus and tibialisposterior) and plantarflexors (tibialis anterior, extensor hallucis longus andextensor digitorum longus) can be seen in Figure A.2.

Page 36: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 3

Figure A.2: Skeletal muscles of the lower leg [11].

A.3.1 GastrocnemiusThe Gastrocnemius is the most superficial muscle in the posterior part of theleg that runs from back of knee to the heel. It has two heads – medial andlateral – which attach to the medial and lateral sides of the femur.

A.3.2 SoleusThe Soleus is a flat and broad muscle located just beneath to the medial andlateral gastrocnemius muscle heads. It originates from the upper portions ofthe tibia and fibula, the bones of the lower leg, and then joins with thegastrocnemius to attach to the heel with the Achilles tendon.

A.3.3 Tibialis posteriorThe Tibialis posterior is in the deep compartment of the posterior leg andlocated between the flexor digitorum longus and the flexor hallucis longus. Itattaches to the plantar surfaces of the foot via a tendon.

A.3.4 Tibialis anteriorThe Tibialis anterior is a long, narrowmuscle lying superficially in the anteriorcompartment of the lower leg. It originates in the upper two-thirds of the

Page 37: High density EMG based estimation of lower limb muscle

4 APPENDIX A. BACKGROUND

lateral surface of the tibia and runs down the shin to the foot. This muscle actsas the main foot dorsiflexor.

A.3.5 Extensor hallucis longusThe Extensor hallucis longus is a thin muscle, situated on the lateral side ofthe leg, between the Tibialis anterior and the Extensor digitorum longus. Itarises from the anterior surface of the fibula, then the fibers pass downward,and end in a tendon inserted into the base of the distal phalanx of the great toe.

A.3.6 Extensor digitorum longusThe Extensor digitorum longus is a long, thin muscle that runs down the frontof the shin, across the ankle joint, and into foot. The muscle divides into fourslips, which run forward on the dorsum of the foot, and are inserted into thefour lesser toes.

A.4 Skeletal muscle structureThere are three different types of muscles that can be found in a human body:skeletal (striated), smooth, and cardiac. Skeletal muscle is the most abundantof the three making up close to 42% of the total bodymass [12] and its primaryfunction is to generate movements, but also to resist external forces – suchas gravity – to maintain certain poses. Apart from the skeletal movementsit contributes to protecting and supporting the skeleton, furthermore it alsohas a role in regulating the body’s glucose homeostasis. Skeletal muscles arecontrolled by the peripheral part of the central nervous system (CNS), allowingvoluntary control over a great range of movements.

Each skeletal muscle is an organ that vary considerably in size, shape,and arrangement of fibers. They consist of various integrated tissues such asskeletal muscle fibers, blood vessels, nerve fibers, and connective tissue.Inside each skeletal muscle, muscle fibers are organized into individualbundles, called a fascicle, which enables the nervous system to trigger aspecific movement by activating a subset of muscle fibers within a fascicle ofthe muscle. The structure of the skeletal muscles can be seen on Figure A.3.

Page 38: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 5

Figure A.3: Skeletal muscle structure [11].

Skeletal muscle cells, also called as muscle fibers based on their long andcylindrical shape, are multinucleated cells and consist of hundreds ofmyofibrils. Myofibrils consist of very fine contractile myofilaments, whichextend in parallel along the length of striated muscle fibers. The myofibrilsare made up of two types of protein myofilaments, which give the muscle itsstriped appearance. The thick filaments are composed of myosin, while thethin filaments of actin, along with their regulatory proteins, tropomyosin andtroponin. These filaments, along with the support proteins form thefunctional unit of a skeletal muscle fiber, the sarcomere.

Each sarcomere is bordered by structures called Z-discs, to which theactin myofilaments with its troponin-tropomyosin complex are anchored,while the thicker myosin filaments and their multiple heads located betweenthe actin filaments (Figure A.4). These two types of filaments structuredpartly overlapped in the sarcomere in a relaxed state and muscularcontraction is caused by the interaction between actin and myosin as theytemporarily bind to each other and are released causing a sliding movementbetween them.

Page 39: High density EMG based estimation of lower limb muscle

6 APPENDIX A. BACKGROUND

Figure A.4: Schematic view of the sarcomere [11].

A.5 Muscle activation

A.5.1 Neuromuscular junctionSkeletal muscles are voluntary controlled muscles which means they canonly functionally contract, apart from some special cases, through aconscious signal generated by motor neurons, originating from the centralnervous system. There are two types of motor neurons, upper motor neuronand lower motor neuron, depending on their locations. Upper motor neuronsare located in the motor cortex and the axon of these cells descend from thecortex to the spinal cord in a form of bundles of axons which are called nervetracts. These neurons mostly project directly onto the lower motor neuronslocated in the spinal cord, and they directly or indirectly innervate effectortargets, i.e. the muscle fibers. The structure of the upper and lower motor

Page 40: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 7

neurons can be seen on Figure A.5.

Figure A.5: Motor neuron structure [11].

Each lower motor neuron has a long axon that projects to the muscle thatit is controlling and at the end branches out to several terminal ending in axonterminals. The axon terminals connect to the muscle fibers which is called the

Page 41: High density EMG based estimation of lower limb muscle

8 APPENDIX A. BACKGROUND

neuromuscular junction (NMJ) as can be seen in Figure A.6.

Figure A.6: Neuromuscular junction [11].

A.5.2 Motor Unit (MU)The combination of a single motor neuron and the muscle fibers it innervatesis called a motor unit. The size of the motor units can vary depending onthe number of muscle fiber included in one motor unit, and it can range froma few dozens to thousands of muscle fibers depending on the nature of themuscle and the level of control, however one muscle fiber is only part of onemotor unit. Muscle units are usually recruited by the nervous system graduallyfrom smaller to larger to increase the generated force, but it is also possible tosimultaneously activate all the motor units to produce the maximal force. Thiscontraction however requires a lot of energy so it can’t be sustained for long,for this reason in normal cases only part of the available motor units are activewhile the remaining ones are resting.

Page 42: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 9

A.5.3 Activation PotentialThe signals originating from the motor neurons that are triggering thecontractions called an action potential which propagates to the muscle fibersconnected to that neuron. One action potential produces one contraction,also called as a twitch, in the muscle fibers, and it has three different phases:latent period, contraction period, and relaxation period. As Figure A.7shows, the maximal generated tension is at the end of the contraction phaseand only lasts about 10 - 20 milliseconds. This means that in order toproduce a contraction that can actually do some work, the motor units have tofire several times in a second. Consequently, the frequency in with theneurons fire is in direct correlation to the produced tension in the muscles, asseen in Figure A.7. If the firing rate reaches a frequency when the musclefibers receive a new stimulus before reaching the relaxation period, thetension increases continuously – until saturation – and this phenomenon iscalled tetanus.

Figure A.7: Muscle action potential (left) and tension (right) [11].

These stimuli origination from the motor neurons then trigger thedepolarization of the muscle fiber membrane, which propagates through thefiber, and the superimposed potential differences over time in the musclefibers of one motor unit is called the Motor Unit Action Potential (MUAP)[13].

A.5.4 Muscle contractionAs the MUAP propagates through the muscle fibers, it causes the sarcomeresto shorten and creates the contraction and creates a force calledmuscle tension.However, muscle also can generate force when there is no movement but onlya static load. These two type of contraction are called isotonic and isometriccontraction, respectively, and are explained in Figure A.8.

Page 43: High density EMG based estimation of lower limb muscle

10 APPENDIX A. BACKGROUND

Figure A.8: Types of Muscle Contractions [11].

A.6 EMG signalElectromyography (EMG) is a technique to measure the electromagneticfield around the muscle created by the constant depolarization andrepolarization of the muscle fiber membrane during activation. Thiselectromagnetic field can be detected and measured by an electrode that isplaced near this field. Ultimately, the measured EMG signal is the sum of theMUAPs and the background interference [14].

A.6.1 Electrode typesThere are two main types of electrodes used in EMG measurement:intramuscular, also known as needle electrode, and surface electrodes [15].

Page 44: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 11

Intramuscular electrode

As its name suggest, this type of electrode is an invasive method for measuringEMG. The electrode resembles a hollow needle, where the tip of the needleacts as the electrode, and the signal is conducted inside the cannula with aninsulated wire to the acquisition unit. There are two main advantage for thiselectrode, one is the small surface area of the electrode enables it to pick upthe signal of individual motor units and small contractions as well withoutsignificant crosstalk. The second is that since it is inserted into the muscledirectly, the proximity of the detection and signal source helps the acquisitionof a cleaner signal, because it is picked up before travelling through additionaltissues. The downside to this electrode is that only a small area of the musclecan be examined, it can be painful to use for the patients, and there is a risk ofinfections.

Surface electrode

Compared to the invasive needles, surface electrodes offer a more convenientsolution to measure EMG. In this case the electrode is simply placed on theskin, above the muscles, and the signal is measured utilizing the conductivityof the skin. There are some drawbacks to this method too, crosstalk is a majorproblem in this case, deep muscles are harder to measure since the electrodesare located on the surface, and the signal generally contains more noise. Toimprove the results proper preparation, like cleaning the skin and applyingconductive gel are common practices.

High-Density EMG (HD-EMG)

The most commonly used electrodes in EMG measurement are usually onechannel electrodes, meaning the signal is measured at one single point.Recently high-density EMG electrodes, i.e. a matrix of several smallelectrodes integrated in one unit, gained more popularity, as they provide awider surface coverage. This enables for a more precise measurement and fora more advanced analysis.

A.7 Factors affecting EMG signalEMG signals are an important source of information in examining musclefunction and movement, however the measurable signal usually differ fromthe ideal. There are a number of factors that can affect the properties of the

Page 45: High density EMG based estimation of lower limb muscle

12 APPENDIX A. BACKGROUND

signal, such as the amplitude, duration or shape of the action potential, and itcan be classified in three groups: causative, intermediate and deterministicfactors [16, 17].

Causative factors can be divided into intrinsic and extrinsic factors.Intrinsic factors are of physiological and anatomical origin, like number ofactive motor units in the measured muscle, the depth and fiber typecomposition of the muscle fibers, and the inconsistent thickness of the tissuesbetween the muscle and the electrodes. Extrinsic factors are howeverconnected to the properties and placement of electrodes. Differentconfiguration of the electrodes can result in different output signals, theshape of the electrodes, distance between them, the relative location of theelectrodes compared to the motor units, and orientation can affect the signal.

Intermediate and deterministic factors are ones that are influenced bycausative factors. Intermediate factor can be the superposition of actionpotentials of neighboring motor units, the propagation velocity of the actionpotential through the muscles, or crosstalk from another muscle. The reasonsbehind deterministic factors originates frequently form motor unit firing rateor interaction between muscle fibers.

A.8 EMG contaminantsApart from the factors that influence the EMG readings, there are severalartefacts that contaminates the signal and makes it harder to analyze [16].These contaminants are introduced mostly as noise or irregularity to thesignal. Common source of noise is the electrical noise from electricaldevices and from power lines. The human body acts similar to an antennaand picks up and amplifies the electromagnetic noise in the environment[18]. While the presence of electromagnetic radiation is unavoidable, withhigh quality devices its effect is usually unnoticeable, however in most casesthe 50-60 Hz interference from the power lines are usually picked up by theelectrodes, but with an appropriate filter its effect can be mitigated [19, 20].

Another typical contaminant is the motion artefact. Motion artifact iscaused by either the relative displacement between the sensor and the skin, orirregularities between the connection of the cable and the sensor. Withcareful fixation and a proper treatment of the skin before applying theelectrodes can however reduce the effect of this artefact [21].

Crosstalk is also a common source of contamination in EMGmeasurement. It is the result of superimposed signals propagating fromnearby muscles and it can depend on several physiological parameters [22]

Page 46: High density EMG based estimation of lower limb muscle

APPENDIX A. BACKGROUND 13

[23]. Crosstalk can lead to misinterpretation to the signal but with aminimalization of the surface electrode area and with the emerging ofhigh-definition surface electrodes its effect can be decreased [24, 25]. Aspecial case of crosstalk is the presence of the electrocardiogram (ECG)signal during measurements which usually higher in amplitude than theEMG signal and can distort its frequency components [26]. This artefactmostly affects the EMG measurements done on the upper body while it canbe insignificant on the lower limbs and also there have been recent methodsthat showed promising result isolating and removing this artefact [27].

Other anatomical and physiological factors can also affect the signalquality during EMG measurements. They are referred to as internal noise,and one of the most important one is the thickness of the skin and thesubcutaneous tissue between the muscle and the electrode [25]. Researchshows that the amount of body fat will directly affect the amplitude of thesignal and it can be considered as an internal noise but can be filtered to anextent by applying proper signal processing [28, 29].

A.9 EMG featuresHigh quality EMG signal is essential for extracting useful information, butreducing noise and avoiding other contaminants is just the first part in thesignal conditioning. Features, i.e. a measurable characteristics of the signal,are important aspect [30] in any machine learning and pattern recognitionapplication, such as the case in EMG signals as well as Gongfa et al [31]describes feature extraction as “one of the most important steps”.

In EMG signal analysis the features can be separated in three main groups[32]: time domain, frequency domain, and time-frequency domain signals.The most commonly used features are the ones represented in time domain[33], since they are relatively simple to calculate, and this allows for real timeanalysis as well. Frequency domain based features are mostly used for musclefatigue estimation and MU recruitment [33].

Phinyomark et al. [34] compared the performance of several features andin their research Willison amplitude (WAMP) showed the best result, whileWaveform length (WL), Root Mean Square (RMS), Mean absolute value(MAV), Integrated EMG (IEMG) also yielded good results. DespitePhinyomark et al. [35] proposed in a later study a more general approach forselecting feature sets, other studies suggests [36–38] that the performance ofthe features vary based on signal characteristics and application. Given thelimited number of studies on a EMG based measurement and estimation of

Page 47: High density EMG based estimation of lower limb muscle

14 APPENDIX A. BACKGROUND

the lower leg with high density EMG, it raises the question which featureswould give an optimal performance.

Page 48: High density EMG based estimation of lower limb muscle

Bibliography

[1] Yunfen Wu, Mara ngeles Martnez Martnez, and Pedro OrizaolaBalaguer. “Overview of the Application of EMG Recording in theDiagnosis and Approach of Neurological Disorders”. In:Electrodiagnosis in New Frontiers of Clinical Research. InTech, May2013.

[2] Michael Wehner. “Man to Machine, Applications inElectromyography”. In: EMG Methods for Evaluating Muscle andNerve Function. InTech, Jan. 2012.

[3] Gea Drost et al. Clinical applications of high-density surface EMG: Asystematic review. Dec. 2006.

[4] Kyung Eun Nam et al. “When does spasticity in the upper limbdevelop after a first stroke? A nationwide observational study on 861stroke patients”. In: Journal of Clinical Neuroscience 66 (Aug. 2019),pp. 144–148.

[5] C L Watkins et al. “Prevalence of spasticity post stroke.” In: Clinicalrehabilitation 16.5 (Aug. 2002), pp. 515–22.

[6] Henry Woodford and C. Price. EMG biofeedback for the recovery ofmotor function after stroke. 2007.

[7] Didier Staudenmann et al. “Independent component analysis ofhigh-density electromyography in muscle force estimation”. In: IEEETransactions on Biomedical Engineering 54.4 (Apr. 2007),pp. 751–754.

[8] A. Gallina et al. “Between-day reliability of triceps surae responses tostanding perturbations in people post-stroke and healthy controls: Ahigh-density surface EMG investigation”. In: Gait and Posture 44(Feb. 2016), pp. 103–109.

15

Page 49: High density EMG based estimation of lower limb muscle

16 BIBLIOGRAPHY

[9] Carlos Murillo et al. “High-Density Electromyography Provides NewInsights into the Flexion Relaxation Phenomenon in Individuals withLow Back Pain”. In: Scientific Reports 9.1 (Dec. 2019), pp. 1–9.

[10] Lori L. Ploutz-Snyder et al. “Functionally relevant thresholds ofquadriceps femoris strength”. In: Journals of Gerontology - Series ABiological Sciences and Medical Sciences 57.4 (2002), B144–B152.

[11] OpenStax. Anatomy and Physiology. OpenStax CNX, Feb. 2016.

[12] Robert C. Lee, Zi Mian Wang, and Steven B. Heymsfield. “Skeletalmuscle mass and aging: Regional and whole-body measurementmethods”. In: Canadian Journal of Applied Physiology 26.1 (2001),pp. 102–122.

[13] V. John Basmajian and Carlo J. de Luca. “Muscles Alive : theirfunctions revealed by electromyography”. In: Muscles Alive: TheirFunctions Revealed by Electromyography. 1985, pp. 65–70.

[14] John G. Webster, ed. Encyclopedia of Medical Devices andInstrumentation. Hoboken, NJ, USA: John Wiley & Sons, Inc., Apr.2006.

[15] Muhammad Zahak. “Signal Acquisition Using Surface EMG andCircuit Design Considerations for Robotic Prosthesis”. In:Computational Intelligence in Electromyography Analysis - APerspective on Current Applications and Future Challenges. InTech,Oct. 2012.

[16] M. B.I. Reaz, M. S. Hussain, and F. Mohd-Yasin. “Techniques of EMGsignal analysis: Detection, processing, classification and applications”.In: Biological Procedures Online 8.1 (Mar. 2006), pp. 11–35.

[17] Paul McCool et al. “Identification of contaminant type in surfaceelectromyography (EMG) signals”. In: IEEE Transactions on NeuralSystems and Rehabilitation Engineering 22.4 (July 2014),pp. 774–783.

[18] E. A. Clancy, E. L. Morin, and R. Merletti. “Sampling,noise-reduction and amplitude estimation issues in surfaceelectromyography”. In: Journal of Electromyography andKinesiology. Vol. 12. 1. Elsevier, Feb. 2002, pp. 1–16.

Page 50: High density EMG based estimation of lower limb muscle

BIBLIOGRAPHY 17

[19] Mojtaba Malboubi, Farbod Razzazi, and Mahdi Aliyari Shoorehdeli.“Elimination of power line noise from EMG signals using an efficientadaptive Laguerre filter”. In: International Conference on Signals andElectronic Systems, ICSES’10 - Conference Proceeding (Oct. 2010),pp. 49–52.

[20] S. D.H. Soedirdjo, K. Ullah, and R. Merletti. “Power line interferenceattenuation in multi-channel sEMG signals: Algorithms and analysis”.In: Proceedings of the Annual International Conference of the IEEEEngineering in Medicine and Biology Society, EMBS.Vol. 2015-Novem. Institute of Electrical and Electronics EngineersInc., Nov. 2015, pp. 3823–3826.

[21] Hak W. Tam and John G. Webster. “Minimizing Electrode MotionArtifact by Skin Abrasion”. In: IEEE Transactions on BiomedicalEngineering BME-24.2 (1977), pp. 134–139.

[22] S Viljoen, T Hanekom, and D Farina. “Effect of characteristics ofdynamic muscle contraction on crosstalk in surface electromyographyrecordings”. In: SAIEE Africa Research Journal 98.1 (2007),pp. 18–28.

[23] Madeleine M. Lowery, Nikolay S. Stoykov, and Todd A. Kuiken. “Asimulation study to examine the use of cross-correlation as an estimateof surface EMG cross talk”. In: Journal of Applied Physiology 94.4(Apr. 2003), pp. 1324–1334.

[24] Irsa Talib et al. A review on crosstalk in myographic signals. Jan. 2019.

[25] Rubana H. Chowdhury et al. Surface electromyography signalprocessing and classification techniques. Sept. 2013.

[26] Nienke W. Willigenburg et al. “Removing ECG contamination fromEMG recordings: A comparison of ICA-based and other filteringprocedures”. In: Journal of Electromyography and Kinesiology 22.3(June 2012), pp. 485–493.

[27] S Abbaspour and A Fallah. “Removing ECG Artifact from the SurfaceEMG Signal Using Adaptive Subtraction Technique.” In: Journal ofbiomedical physics & engineering 4.1 (Mar. 2014), pp. 33–8.

[28] Dario Farina and Alberto Rainoldi. “Compensation of the effect ofsub-cutaneous tissue layers on surface EMG: A simulation study”. In:Medical Engineering and Physics 21.6-7 (July 1999), pp. 487–497.

Page 51: High density EMG based estimation of lower limb muscle

18 BIBLIOGRAPHY

[29] Monica A. Hemingway, Heinz J. Biedermann, and James Inglis.“Electromyographic recordings of paraspinal muscles: Variationsrelated to subcutaneous tissue thickness”. In: Biofeedback andSelf-Regulation 20.1 (1995), pp. 39–49.

[30] K Rani, Dr G.Naga Rama Devi, and Lavanya Doddipalli. “Importanceof Feature Extraction for Classification of Breast Cancer Datasets – AStudy”. In: International Journal of Scientific and InnovativeMathematical Research 3 (July 2015), pp. 763–768.

[31] Gongfa Li et al. “A novel feature extraction method for machinelearning based on surface electromyography from healthy brain”. In:Neural Computing and Applications 31.12 (Dec. 2019),pp. 9013–9022.

[32] Angkoon Phinyomark, Pornchai Phukpattaranont, andChusak Limsakul. “Feature reduction and selection for EMG signalclassification”. In: Expert Systems with Applications 39.8 (June 2012),pp. 7420–7431.

[33] Mohammadreza Asghari Oskoei and Huosheng Hu. Myoelectriccontrol systems-A survey. Oct. 2007.

[34] A. Phinyomark et al. “Evaluation of EMG Feature Extraction forMovement Control of Upper Limb Prostheses Based on ClassSeparation Index”. In: IFMBE Proceedings. Vol. 35 IFMBE. 2011,pp. 750–754.

[35] Angkoon Phinyomark et al. “Navigating features: A topologicallyinformed chart of electromyographic features space”. In: Journal ofthe Royal Society Interface 14.137 (Dec. 2017).

[36] Angkoon Phinyomark et al. “EMG feature evaluation for improvingmyoelectric pattern recognition robustness”. In: Expert Systems withApplications 40.12 (2013), pp. 4832–4840.

[37] Adenike A. Adewuyi, Levi J. Hargrove, and Todd A. Kuiken.“Evaluating EMG Feature and Classifier Selection for Application toPartial-Hand Prosthesis Control”. In: Frontiers in Neurorobotics 10(Oct. 2016), p. 15.

[38] Sara Abbaspour et al. “Evaluation of surface EMG-based recognitionalgorithms for decoding hand movements”. In:Medical and BiologicalEngineering and Computing 58.1 (Jan. 2020), pp. 83–100.

Page 52: High density EMG based estimation of lower limb muscle

44 BIBLIOGRAPHY

B Measurement protocol

Page 53: High density EMG based estimation of lower limb muscle

1

Measurement Protocol

Basic information

Subject n° Date of Birth

Name Gender

Height (m) Weight (kg)

Physical fitness level*

Low Moderate High

Tibia Length (m)

Ankle width (m)

Side: Foot length (m)

Date of testing

Comments * Physical fitness level: low (no or very physical activities), moderate (training 1-2 times per week), high (training 3 times

per week or more often).

Measurements A. 100% MVC for 5s at ankle joint = -30°, -15°, 0° and 10° in a randomized order (-: plantarflexion, + dorsiflexion), each MVC

repeat 3 times B. % MVC profile (5s from relax to %MVC -> 4s %MVC hold -> 5s %MVC to relax) at ankle joint = -15°, -7.5°, 0° and 10° in a

randomized order each %MVC profile repeat 5 times.

• HD-EMG 3*64chs IN1: GAS MED, IN2: TA, IN3: SOL and 2*32chs IN1-2: GAS LAT, IN3-4: PL) • Three reference straps: two around the ankle and one around knee. GAS MED and GAS LAT connected with knee strap,

other three muscles connect with ankle strap.

Angle Flexion type MVC% File name Offset MVC value A

-15°

Plantarflexion

100% AN30_PF_100MVC B 70% AN30_PF_70MVC

B 50% AN30_PF_50MVC B 30% AN30_PF_30MVC A

Dorsiflexion

100% AN30_DF_100MVC B 70% AN30_DF_70MVC

B 50% AN30_DF_50MVC B 30% AN30_DF_30MVC A

-7.5°

Plantarflexion

100% AN15_PF_100MVC B 70% AN15_PF_70MVC

B 50% AN15_PF_50MVC B 30% AN15_PF_30MVC A

Dorsiflexion

100% AN15_DF_100MVC B 70% AN15_DF_70MVC

B 50% AN15_DF_50MVC B 30% AN15_DF_30MVC A

0° Plantarflexion

100% AP0_PF_100MVC B 70% AP0_PF_70MVC

B 50% AP0_PF_50MVC B 30% AP0_PF_30MVC

Page 54: High density EMG based estimation of lower limb muscle

2

A

Dorsiflexion

100% AP0_DF_100MVC B 70% AP0_DF_70MVC

B 50% AP0_DF_50MVC B 30% AP0_DF_30MVC A

10°

Plantarflexion

100% AP10_PF_100MVC B 70% AP10_PF_70MVC

B 50% AP10_PF_50MVC B 30% AP10_PF_30MVC A

Dorsiflexion

100% AP10_DF_100MVC B 70% AP10_DF_70MVC

B 50% AP10_DF_50MVC B 30% AP10_DF_30MVC

Notes • To record muscle activity during MVC isometric ankle plantar and dorsiflexion contractions will be recorded in the

tibialis anterior (TA), medial gastrocnemius (MG), soleus (SOL), lateral gastrocnemius (LG) and Perenous Longus (PL) muscles activity.

• Dorsiflexion: flexion of the foot upward (measurement of TA taken from a foot position of 30 degrees upward). • Plantarflexion: extension of the foot downward (measurement of TA, MG, SOL and PL) • Recording parameters: EMG signals and in the meanwhile joint torque from dynamometer.

Page 55: High density EMG based estimation of lower limb muscle

3

Skin preparation and Electrodes placement (seniam.org sensor placement recommendation for sEMG) Medial gastrocnemius

• Electrodes need to be placed on the most prominent bulge of the muscle • Orientation: In the direction of leg

Lateral gastrocnemius

• Electrodes need to be placed at 1/3 of the line between the head of the fibula and the heel. • Orientation: in the direction of the line between the head of the fibula and the heel.

Soleus

• The electrodes need to be placed at 2/3 of the line between the medial condylis of the femur to the medial malleolus

• Orientation: in the direction of the line between the medial condylis to the medial malleolus.

Page 56: High density EMG based estimation of lower limb muscle

4

Tibialis anterior

• The electrodes need to be placed at 1/3 on the line between the tip of the fibula and the tip of the medial malleolus.

• Orientation: in the direction of the line between the tip of the fibula and the tip of the medial malleolus.

Perenous Longus

• Electrodes need to be placed at 25% on the line between the tip of the head of the fibula to the tip of the lateral malleolus.

• Oreitnation: in the direction of the line between the tip of the head of the fibula to the tip of the lateral malleolus.

Page 57: High density EMG based estimation of lower limb muscle
Page 58: High density EMG based estimation of lower limb muscle

TRITA CBH-GRU-2020:290

www.kth.se