big_data_and_machine_learning_in_plastic_surgery__.45 (5)

Copyright © 2016 American Society of Plastic Surgeons. Unauthorized reproduction of this article is prohibited.

www.PRSJournal.com890e

In the era of evidence-based medicine, a vast amount of information is collected on patients.1,2 This information has become

increasingly useful in guiding treatment and opti-mizing clinical outcomes in medical care. The result is an ever-expanding volume of data con-taining complex patterns that may extend beyond the physician’s ability to use traditional data pro-cessing techniques such as regression and multi-variate analysis for interpretation.1,2 As innovators, plastic surgeons must then adapt to the grow-ing trend of “big data,” and find ways to tap its resources to deliver more efficient health care and improved surgical outcomes.

The answer may lie in “machine learning.” A subfield of artificial intelligence, machine learning involves generating algorithms capa-ble of knowledge acquisition through historical examples. Machine learning has already been

applied successfully to big data problems in vari-ous sectors, with applications including speech recognition and search engine optimization.3 In medicine, the IBM Watson Health (International Business Machines Corp., Armonk, N.Y.) cogni-tive computing system has used machine learning approaches to create a decision support system for physicians treating cancer patients, with the intention of improving diagnostic accuracy and reducing costs. Initially trained at Memorial Sloan Kettering Cancer Center using large volumes of patient cases and over 1 million scholarly arti-cles, the project now has 14 participating cancer centers.4,5 All of these centers contribute to an ever-expanding corpus of information that helps

Disclosure: None of the authors has a financial interest in any of the products, devices, drugs or procedures mentioned in this article.

Copyright © 2016 by the American Society of Plastic Surgeons

DOI: 10.1097/PRS.0000000000002088

Jonathan Kanevsky, M.D.Jason Corban, B.Sc.

Richard Gaster, M.D., Ph.D.Ari Kanevsky

Samuel Lin, M.D.Mirko Gilardino, M.D.,

M.Sc.

Montreal, Quebec, Canada; Boston, Mass.; and Albany, N.Y.

Summary: Medical decision-making is increasingly based on quantifiable data. From the moment patients come into contact with the health care system, their entire medical history is recorded electronically. Whether a patient is in the operating room or on the hospital ward, technological advancement has facilitated the expedient and reliable measurement of clinically relevant health metrics, all in an effort to guide care and ensure the best possible clinical outcomes. However, as the volume and complexity of biomedical data grow, it becomes challenging to effectively process “big data” using conventional techniques. Physicians and scientists must be prepared to look beyond clas-sic methods of data processing to extract clinically relevant information. The purpose of this article is to introduce the modern plastic surgeon to machine learning and computational interpretation of large data sets. What is machine learning? Machine learning, a subfield of artificial intelligence, can address clinically relevant problems in several domains of plastic surgery, including burn surgery; microsurgery; and craniofacial, peripheral nerve, and aesthetic surgery. This article provides a brief introduction to current research and sug-gests future projects that will allow plastic surgeons to explore this new frontier of surgical science. (Plast. Reconstr. Surg. 137: 890e, 2016.)

From the Division of Plastic and Reconstructive Surgery, Faculty of Medicine, McGill University; the Division of Plastic and Reconstructive Surgery, Harvard University, the Division of Plastic and Reconstructive Surgery, Beth Israel Deaconess Medical Center; and the Department of Biological Sciences, University at Albany.Received for publication May 22, 2015; accepted December 22, 2015.

Big Data and Machine Learning in Plastic Surgery: A New Frontier in Surgical Innovation

Supplemental digital content is available for this article. Direct URL citations appear in the text; simply type the URL address into any Web browser to access this content. Clickable links to the material are provided in the HTML text of this article on the Journal’s Web site (www.PRSJournal.com).

SUPPLEMENTAL DIGITAL CONTENT IS AVAIL-ABLE IN THE TEXT.

TECHNOLOGY & INNOvaTIONs

http://www.PRSJournal.com

http://www.PRSJournal.com


Volume 137, Number 5 • Big Data and Machine Learning

891e

Watson fine tune its ability to suggest optimal treatment options for cancer patients based on the nature of their specific illness.4,5

Although similar to data mining, which tradi-tionally involves knowledge acquisition through analysis of preexisting data sets, machine learning places a greater emphasis on descriptive modeling and outcome prediction for novel data.5 Further-more, machine learning algorithms are capable of improving or “learning” when exposed to more information.5,6 As the algorithm attempts to find the most appropriate hypotheses for a given data set, within the computational boundaries of the spe-cific machine learning approach used, it statistically assesses how each model compares to each other and models that have been assessed previously.6 The result of this process is the creation of data models that are either predictive or descriptive in nature.

Predictive machine learning models fall under the domain of supervised learning, where the algorithm has been trained using examples of both inputs and desired outputs, allowing for mapping of future inputs to outputs.6,7 The goal of this process is a unique mathematical model capable of predicting desired target values from novel data (Fig. 1). For example, a recent surgi-cal application involved the use of data from the American College of Surgeons National Surgical Quality Improvement Program to train a support vector machine to quantify procedural complex-ity and risk associated with different procedures

based on Current Procedural Terminology codes.8 Using data from 2005 to 2009, the support vector machine was trained to determine the association between Current Procedural Terminology and mortality, morbidity, Clavien classification type IV complications, and surgical-site infection to produce an algorithm capable of generating pro-cedural risk scores.8 When tested using National Surgical Quality Improvement Program data from 2010, the support vector machine approach achieved a greater level of discrimination for determining surgical complications compared with other measures of procedural complexity.8

Descriptive models, in contrast, fall under the category of “unsupervised learning.” Unsupervised learning analyzes data that are unlabeled, and the system discovers structure in the data for interpre-tation (Fig. 2).6,9 (See Video, Supplemental Digital Content 1, which highlights the difference between supervised and unsupervised machine learning using hypothetical machine learning algorithms capable of processing visual data for the detection and differentiation of different types of craniosyn-ostosis. Available in the “Related Videos” section of the full-text article on PRSJournal.com or, for Ovid users, available at http://links.lww.com/PRS/B706.) This type of machine learning has been applied in the field of molecular genetics and genomics to organize and interpret vast amounts of genetic information.10 Following the application of leu-kemic blasts from pediatric acute lymphoblastic

Fig. 1. Graphic representation of supervised machine learning. In supervised learning, original prepro-cessed data sets, containing known variables and targets, are divided into training data and test data. (Above) During the training phase, the training data are used to train a learning algorithm in an attempt to develop an accurate predictive model. (Center) To validate the model, the test data are then applied to the model and predictive accuracy is assessed. (Below) Once validated, new data are input into the model in an attempt to make new predictions.

http://links.lww.com/PRS/B706


892e

Plastic and Reconstructive Surgery • May 2016

leukemia patients to microarrays, unsupervised clustering of the data identified six new clinical subtypes of acute lymphoblastic leukemia.10

As illustrated above, machine learning has already been applied, with great success, to pro-cess large amounts of complex data in medicine

and surgery. With the volumes of patient data gen-erated in all domains of plastic surgery and the emergence of large databases such as the National Surgical Quality Improvement Program and Track-ing Operations and Outcomes for Plastic Surgeons for storing this information, plastic surgeons stand to benefit from similar objective and data-driven machine learning approaches. This article presents a selection of preliminary investigations in the fields of burn surgery; microsurgery; and craniofacial, peripheral nerve, hand, and aesthetic surgery, and proposes future applications in an effort to demon-strate how machine learning may be used to lever-age complex, clinically derived data into improved efficiency and better clinical outcomes in plastic surgery. Institutional review board exemption was granted by our medical center review board.

CURRENT AND FUTURE APPLICATIONS OF MACHINE LEARNING IN PLASTIC

SURGERY

Burn SurgeryAn early application of machine learning

related to plastic surgery was the development of a method to accurately determine healing time in burn injury.11 Using reflectance spectrometry and an artificial neural network, researchers devel-oped a model to predict whether a burn would take more or less than 14 days to heal, ultimately

Fig. 2. Graphic representation of unsupervised machine learning. In unsupervised learning, raw data, containing unknown patterns and targets, are presented to an algorithm. The algorithm attempts to develop descriptive models for the data based on regularities detected. (adapted from Hudson Legal. Unsupervised learning. available at: http://us.hudson.com/portals/Us/images/blogs/legal/wp/2011/09/Unsupervised-Learning1.jpg. accessed October 6, 2014.)

Video. supplemental Digital Content 1 highlights the difference between supervised and unsupervised machine learning using hypothetical machine learning algorithms capable of process-ing visual data for the detection and differentiation of different types of craniosynostosis. available in the “Related videos” sec-tion of the full-text article on PRsJournal.com or, for Ovid users, available at http://links.lww.com/PRS/B706.

http://us.hudson.com/portals/US/images/blogs/legal/wp/2011/09/Unsupervised-Learning1.jpg


http://links.lww.com/PRS/B706



893e

serving as a proxy for the assessment of burn depth for surgical planning. Artificial neural net-works consist of input nodes, representing the data to be used for prediction, intermediate or “hidden nodes” that calculate predictions based on the inputs, and output nodes that represent the predictions themselves.6 During training, arti-ficial neural networks are tuned through a pro-cess of “back-propagation” where the accuracy of the output values is compared to the actual target values (Fig. 3).6 In this investigation, normalized spectral data served as input nodes, whereas the two output nodes distinguished between spec-tra associated with burns that healed in less than 14 days and those associated with burns that took more than 14 days to heal.11 After examining reflectance spectrometry data from 41 wounds, the investigators determined that their model had an average predictive accuracy of 86 percent, sug-gesting that it may serve as an effective screening tool for assessing burn depth and a superior alter-native to direct visualization.11

Another task within burn care that might lend itself to machine learning is the accurate and pre-cise quantification of burn size (total body sur-face area). Current methods, such as the “rule of nines,” are limited by the asymmetry of burns, surface area variations related to patient age, and

interobserver variability. By pairing images of burns to precise measurements of the percent-age of body area affected, an algorithm could be trained to rapidly and accurately predict the per-centage of burned tissue (Fig. 4). From these mea-surements, more accurate resuscitation protocols could be generated in addition to surgical plan-ning strategies for autografting or allografting.

MicrosurgeryPostoperative monitoring after microsurgery is

crucial for achieving desired clinical outcomes. In light of this, researchers have recently developed a postoperative microsurgery monitoring applica-tion, the SilpaRamanitor, capable of quantifying free-flap tissue perfusion.12 Using 60 smartphone images of middle and index fingers exposed to varying degrees of vascular compromise to mimic vascular occlusion, a k-nearest neighbor algorithm was trained to categorize flap tissue into different classes: normal, venous outflow occlusion, and arterial inflow occlusion.12 In cases of occlusion, the degree was further categorized as partial or complete. The k-nearest neighbor algorithm is a non–parameter supervised learning approach whereby the classification rules are generated by the training samples themselves and do not require the input of additional information.13 The

Fig. 3. Graphic representation of an artificial neural network. Modeled after biological neural networks, artificial neural networks use input nodes, representing data input into the model; hidden nodes, responsible for making the predictions); and output nodes, representative of the predictions being made (Oncologists partner with Watson on genomics. Cancer Discov. 2015;5:788). During training, artificial neural networks, in a fashion similar to biological neurons, take part in a process called back-propagation, whereby the weight of the connections between nodes is adjusted based on the difference between the artificial neural networks output values and known target values. This process ensures that the output of the artificial neural network is as close as possible to the desired target values. (adapted from Meyfroidt G, Güiza F, Ramon J, Bruynooghe M. Machine learning techniques to examine large patient databases. Best Pract Res Clin Anaesthesiol. 2009;23:127–143.)


894e


algorithm assigns a test sample, which in this study was an image of a free flap, to a category, which was the type and degree of occlusion.12 This classi-fication is based on features of the test sample that are the most similar, or the “nearest neighbors,” to those from the training set.12 The overall sensitiv-ity and specificity of this application were found to be 94 percent and 98 percent, respectively, with an accuracy of 95 percent.12 Thus, through the accurate and rapid monitoring of free-flap perfusion, the SilpaRamanitor application is an example of how machine learning can be used to potentially increase the success of detecting early anastomotic failure or thrombotic issues and con-comitant free-flap salvage.12

However, the spectrum of potential applica-tions of machine learning for microsurgery are not only limited to postoperative monitoring. Machine learning could also benefit preoperative consultation and surgical planning for microsur-gery. Through the collection of detailed infor-mation such as the size and location of defects, the type of flap used, patient age, body mass index, smoking status, and resultant complica-tions in large-scale databases (such as Tracking Operations and Outcomes for Plastic Surgeons), machine learning algorithms could be trained to assess a particular defect in a selected patient and suggest the reconstructive approach with the highest chance of a favorable outcome.

Craniofacial SurgeryCurrently, machine learning is being explored

to facilitate the automated diagnosis of non-syndromic craniosynostosis. Examining com-puted tomographic scans from 141 subjects, of which 50 had either sagittal, metopic, or coronal

craniosynostoses, a regularized linear discrimi-nant analysis algorithm was trained to diagnose craniosynostosis and distinguish between differ-ent types using an index of cranial suture fusion along with deformation and curvature discrep-ancy averages across five cranial bones and six suture regions.14 Regularized linear discriminant analysis is frequently used for high-dimensional data when there are a small number of samples, as was the case in this investigation.15 The result of this machine learning process is an automated classifier capable of differentiating types of cra-niosynostosis based on computed tomographic scans, with a sensitivity of 92.3 percent and a speci-ficity of 98.9 percent.14 With accuracy comparable to trained radiologists, the authors propose that their algorithm may provide an objective standard capable of reducing interobserver variability and providing a quantitative measure of procedural success.14

Although the automation of this process has many benefits, we propose that machine learn-ing theoretically has the potential to enable surgeons to bypass the use of computed tomo-graphic imaging for routine diagnostics. Using three-dimensional surface photographs of differ-ent plagiocephalies, an algorithm could be devel-oped to differentiate between cases of synostotic and deformational plagiocephaly. Together with clinical examination, the goal would be to further reduce the need for ionizing radiation in these children.

Another potential machine learning applica-tion for craniofacial surgery involves the identifi-cation of candidate genes in nonsyndromic cases of cleft lip and palate.16 Through a combination of genomewide association studies and other

Fig. 4. By pairing images of burns to precise measurements of the per-centage of body area affected, an algorithm could be trained to use images to rapidly and accurately predict the percentage of burned tissue and automate part of assessment of burn patients.



895e

investigations, several genetic factors have been elucidated for this condition for which the cause is poorly understood.16 However, our understanding of the molecular pathogenesis of nonsyndromic cleft lip and palate remains far from comprehen-sive. Using methods similar to those used previ-ously in the field of genomics, machine learning has the potential to uncover previously unknown candidate genes and regulatory sequences for nonsyndromic cleft lip and palate, allowing for an improved understanding of the pathogenesis of this condition.16

Hand and Peripheral Nerve SurgeryResearch in the field of hand and peripheral

nerve surgery has the potential to benefit from machine learning. As an example, investigators have recently developed an artificial neural net-work capable of predicting the outcome of differ-ent tissue-engineered peripheral nerve grafts used in research applications.17 Using over 30 inde-pendent variables to describe tissue-engineering materials, artificial neural networks were trained to predict the success of various grafts.17 The suc-cess of the grafts, placed in a rat model, was quan-tified using the critical regeneration length, along with a unitless parameter, the ratio of the actual length to the critical length.17 After application of the validation data, the predicative accuracy of artificial neural networks was 92.59 percent and 90.85 percent for the ratio of the actual length to the critical length and the critical regeneration length, respectively.17 Although preliminary, the results of this investigation highlight the poten-tial role for machine learning in the analysis and development of tissue-engineering strategies for peripheral nerve repair.

The biomechanics of the upper extremities are particularly complex, involving multivari-ate nonlinear relationships that are theoretically amenable to modeling by machine learning. In light of this, artificial neural networks have been used to develop automated controllers for a vari-ety of neuroprostheses, including those that are used to restore hand grasp and wrist control along with more proximal upper extremity func-tion in patients with C5/C6 spinal cord injury.18,19 Although the results of these initial investigations were mixed, they highlight the potential for arti-ficial neural networks, along with other machine learning techniques, in the development of neu-roprosthetic controllers for the hand and wrist.

Along with the examples presented above, machine learning has the potential to provide additional innovation in hand and nerve surgery.

For example, using information from previous cases and radiographic images of fractures, an algorithm could be developed to anticipate the positioning of Kirschner wires, plates, and screws in preoperative planning for hand surgery. Fur-thermore, using databases containing informa-tion derived from sensorial mapping following peripheral nerve repair, patterns of regrowth could be used to develop an algorithm capable of prognosticating the degree of sensory and motor restoration based on location, mechanism of nerve injury, and physical examination findings.

Aesthetic SurgeryMachine learning also has potential applica-

tions in more subjective areas of plastic surgery, such as aesthetics. Using a form of supervised learning, an automated classifier for facial beauty was trained using extracted facial features from 165 images of attractive female faces that were also independently graded by human referees.20 Decision tree algorithms assess a set of descrip-tive attributes, which in this particular investiga-tion included different facial ratios, and attempt to determine attractive facial features most closely related to postoperative target variables—the human classification of facial beauty.20 When subjected to the testing set of images, the auto-mated classifier was shown to have a high accu-racy at approximating human referee scores.20 In light of these results, this application may serve as a predictive tool for estimating a patient’s per-ceived beauty following aesthetic surgery, provid-ing a quantitative measure to set expectations and possibly discourage patients from undergo-ing procedures that offer marginal improvement.

In conjunction with optical head-mounted display technology, machine learning also has the potential to facilitate intraoperative visual-ization of surgical outcomes. One such applica-tion could be in reconstructive breast surgery, whereby machine learning software incorporated into a head mount display could predict what a breast might appear like in three-dimensional space based on potential changes in implant posi-tion. The system could be trained to identify fea-tures of breast aesthetics such as symmetry, nipple position, superior pole fullness, and degree of ptosis, optimizing aesthetic results while mini-mizing trauma and operating time. Although these potential tools may not replace the trained human eye, they would aid the plastic surgeon by providing a higher degree of objectivity to aes-thetic surgery.


896e


Resident Evaluation and TeachingCurrently, competency in plastic surgery train-

ing is assessed through written and oral exami-nations coupled with case logs and subjective evaluation by attending physicians and higher level residents.21 However, more objective methods for assessing “surgical competency” are required, which may be facilitated by machine learning. Recently, with the development of head-mounted cameras for resident teaching and wearable tech-nology, such as Google Glass (Google, Inc., Moun-tain View, Calif.), trainees are able record their cases and compare their performance to previous recordings of themselves and expert attending surgeons.21 By tracking various metrics, trainees can be graded on their surgical skills and realize methods for improvement. However, this tech-nology could be enhanced through the addition of sensors for precisely tracking eye movement, hand motion, and effective use of instruments and assistants. Machine learning could be used to identify salient aspects from the expert recordings that were absent or present in that of the train-ees, allowing for systematic and objective assess-ment of areas that need improvement in real time (Fig. 5). In addition, a set of objective milestones can be developed for systematic grading of surgi-cal competency.

Postoperative follow-up is also of paramount importance for an outcomes-based specialty such as plastic surgery and is an aspect of surgi-cal residency training that needs improvement.

By integrating intraoperative recording of the steps and techniques used during surgery with postoperative findings and images, machine learning could be used to identify the surgical techniques that lead to a particular outcome. This information would ultimately be linked to the resident involved in the procedure, ensuring appropriate long-term follow-up and facilitat-ing targeted feedback for future cases, some-thing that is difficult with the current pedagogic model, where residents rotate to different sites and services.

LIMITATIONSAlthough the potential applications and ben-

efits of machine learning for the field of plastic surgery are evident, to safely apply the findings obtained using this technology, clinicians must be aware of its limitations. Most importantly, machine learning has been criticized for exhibiting “black box” characteristics, with algorithms that pro-vide little or no justification for the outputs they provide.22 Furthermore, the learning behavior of some machines has been shown to be difficult to reproduce when similar training data are applied to different machine learning algorithms.23 To overcome these challenges, it has been sug-gested that future machine learning algorithms could be programmed to include justifications for their decisions.24 However, certain measures can be taken to demonstrate the diagnostic valid-ity of currently available algorithms. As demon-strated by many of the applications presented here, results obtained using machine learning can be compared to those derived from current gold standards for diagnosis. Another interesting approach is the emerging trend of “crowdsourc-ing analytics.”24 Through the use of multiple algo-rithms, or “crowdsourcing,” to address a specific problem, more accurate models of data can be obtained and simultaneously allow for the evalua-tion of the strengths and weaknesses of the differ-ent algorithms being used.24

Another concern is that some investigators may apply machine learning without the expertise to critically assess their results.23 To overcome this, investigators who wish to use machine learning to tackle complex problems should work closely with data scientists who are capable of accurately evaluating the validity of the outputs obtained.26 Ultimately, this would ensure that the results obtained using machine learning technology are interpreted correctly and are being applied in a safe and clinically relevant manner.

Fig. 5. a mock recording of a resident carrying out a carpal tun-nel release using a wearable technological device. The recording has been optimized using unsupervised learning approaches to identify salient features from expert recordings that are either absent or present in that of the trainees, allowing for direction and correction in real time.



897e

CONCLUSIONSThis introductory review of machine learning

highlights the potential this technology has for catalyzing a paradigm shift in research and clini-cal practice in plastic surgery. Machine learning has already demonstrated great success in a vari-ety of fields, including several medical disciplines. In plastic surgery, we have demonstrated that machine learning has the potential to become a powerful tool, allowing surgeons to harness complex clinical data to help guide key clinical decision-making. Although a potentially power-ful tool, computer-generated algorithms will not replace the trained human eye. However, these are tools that may help us not only in the deci-sion-making process but also in finding patterns that might not be evident in analysis of smaller data sets or anecdotal experience. By embracing machine learning, modern plastic surgeons may be able to redefine the specialty while solidifying their role as leaders at the forefront of scientific advancement in surgery.

Mirko Gilardino, M.D., M.Sc.Montreal Children’s Hospital

2300 Tupper Street, Room C-1135Montreal, Quebec H3H 1P3, Canada

[email protected]

Samuel Lin, M.D.Beth Israel Deaconess Medical Center

110 Francis Street, Suite 5ABoston, Mass. 02215

[email protected]

REFERENCES 1. Murdoch TB, Detsky AS. The inevitable application of big

data to health care. JAMA 2013;309:1351–1352. 2. Cleophas TJ, Zwinderman AH. Machine Learning in Medicine:

Cookbook. New York: Springer; 2014. 3. Bose I, Mahapatra RK. Business data mining: A machine

learning perspective. Inform Manage. 2001;39:211–225. 4. Malin JL. Envisioning Watson as a rapid-learning system for

oncology. J Oncol Pract. 2013;9:155–157. 5. Oncologists partner with Watson on genomics. Cancer Discov.

2015;5:788. 6. Furnkranz J, Gamberger D, Lavrac N. Foundations of Rule

Learning. New York: Springer; 2012. 7. Meyfroidt G, Güiza F, Ramon J, Bruynooghe M. Machine

learning techniques to examine large patient databases. Best Pract Res Clin Anaesthesiol. 2009;23:127–143.

8. Van Esbroeck A, Rubinfeld I, Hall B, Syed Z. Quantifying surgical complexity with machine learning: Looking beyond patient factors to improve surgical models. Surgery 2014;156:1097–1105.

9. Hudson Legal. Unsupervised learning. Available at: http://us.hudson.com/portals/US/images/blogs/legal/wp/2011/09/Unsupervised-Learning1.jpg. Accessed October 6, 2014.

10. Ebert BL, Golub TR. Genomic approaches to hematologic malignancies. Blood 2004;104:923–932.

11. Yeong EK, Hsiao TC, Chiang HK, Lin CW. Prediction of burn healing time using artificial neural networks and reflectance spectrometer. Burns 2005;31:415–420.

12. Kiranantawat K, Sitpahul N, Taeprasartsit P, et al. The first Smartphone application for microsurgery monitoring: SilpaRamanitor. Plast Reconstr Surg. 2014;134:130–139.

13. Suguna N, Thanushkodi K. An improved k-nearest neigh-bor classification using genetic algorithm. Int J Comput Sci. 2010;7:18–21.

14. Mendoza CS, Safdar N, Okada K, Myers E, Rogers GF, Linguraru MG. Personalized assessment of craniosyn-ostosis via statistical shape modeling. Med Image Anal. 2014;18:635–646.

15. Yang W, Wu H. Regularized complete linear discriminant analysis. Neurocomputing 2014;137:185–191.

16. Dixon MJ, Marazita ML, Beaty TH, Murray JC. Cleft lip and palate: Understanding genetic and environmental influ-ences. Nat Rev Genet. 2011;12:167–178.

17. Conforth M, Meng Y, Valmikinathan C, Xiaojun Y. Nerve graft selection for peripheral nerve regeneration using neural networks trained by a hybrid ACO/PSO method. Paper pre-sented at: 6th Annual IEEE Symposium on Computational Intelligence in Bioinformatics and Computational Biology; March 30–April 2, 2009; Nashville, Tenn.

18. Hincapie JG, Kirsch RF. Feasibility of EMG-based neural network controller for an upper extremity neuroprosthesis. IEEE Trans Neural Syst Rehabil Eng. 2009;17:80–90.

19. Luján JL, Crago PE. Computer-based test-bed for clinical assessment of hand/wrist feed-forward neuroprosthetic con-trollers using artificial neural networks. Med Biol Eng Comput. 2004;42:754–761.

20. Gunes H, Piccardi M. Assessing facial beauty through pro-portion analysis by image processing and supervised learn-ing. Int J Hum-Comput St. 2006;64:1184–1199.

21. Berger AJ, Gaster RS, Lee GK. Development of an afford-able system for personalized video-documented surgical skill analysis for surgical residency training. Ann Plast Surg. 2013;70:442–446.

22. Foster KR, Koprowski R, Skufca JD. Machine learning, medical diagnosis, and biomedical engineering research: Commentary. Biomed Eng Online 2014;13:1–11.

23. Imhoff M, Kuhls S. Alarm algorithms in critical care moni-toring. Anesth Analg. 2006;102:1525–1537.

24. Nesta. Machines that learn in the wild: Machine learn-ing capabilities, limitations and implications. Available at: http://www.nesta.org.uk/sites/default/files/machines_that_learn_in_the_wild.pdf. Accessed June 30, 2015.

mailto:[email protected]

mailto:[email protected]



http://www.nesta.org.uk/sites/default/files/machines_that_learn_in_the_wild.pdf

http://www.nesta.org.uk/sites/default/files/machines_that_learn_in_the_wild.pdf

big_data_and_machine_learning_in_plastic_surgery__.45 (5)

Documents