Elsevier

Journal of Biomechanics

Volume 81, 16 November 2018, Pages 1-11
Journal of Biomechanics

Review
Machine learning in human movement biomechanics: Best practices, common pitfalls, and new opportunities

https://doi.org/10.1016/j.jbiomech.2018.09.009Get rights and content

Abstract

Traditional laboratory experiments, rehabilitation clinics, and wearable sensors offer biomechanists a wealth of data on healthy and pathological movement. To harness the power of these data and make research more efficient, modern machine learning techniques are starting to complement traditional statistical tools. This survey summarizes the current usage of machine learning methods in human movement biomechanics and highlights best practices that will enable critical evaluation of the literature. We carried out a PubMed/Medline database search for original research articles that used machine learning to study movement biomechanics in patients with musculoskeletal and neuromuscular diseases. Most studies that met our inclusion criteria focused on classifying pathological movement, predicting risk of developing a disease, estimating the effect of an intervention, or automatically recognizing activities to facilitate out-of-clinic patient monitoring. We found that research studies build and evaluate models inconsistently, which motivated our discussion of best practices. We provide recommendations for training and evaluating machine learning models and discuss the potential of several underutilized approaches, such as deep learning, to generate new knowledge about human movement. We believe that cross-training biomechanists in data science and a cultural shift toward sharing of data and tools are essential to maximize the impact of biomechanics research.

Introduction

For most of the 20th century, inference in biomedical research was predominantly based on hypothesis testing using parametric tests, such as the Student’s t test. The current surge of data, however, presents new challenges and opportunities that are shifting the data analytics landscape in many biomedical disciplines, including human movement biomechanics. Data characterizing human movement are high-dimensional, heterogeneous, and growing in volume with wearable sensing; often, they do not satisfy assumptions associated with parametric tests. Advanced analytical techniques to extract informative features from these data and model underlying relationships that cannot be modeled with traditional statistical tools could transform biomechanics research, as they have autonomous driving, speech recognition, and automated cancer detection.

Efforts to modernize biomechanical data analysis are exemplified by the use of feature extraction algorithms such as principal component analysis (PCA). The literature reflects an evolving awareness about the drawbacks of using only summary metrics (e.g., mean acceleration) or salient features (e.g., the peak knee adduction moment) to describe gait data, as summary metrics are not always the most informative with respect to outcomes of interest (e.g., disease status). PCA, which preserves the variability of multivariate datasets while reducing dimensionality to make analyses more tractable, has been used as an alternative (Deluzio and Astephen, 2007, Donoghue et al., 2008, Duhamel et al., 2006, Ryan et al., 2006). Although most biomechanics studies that employ these methods for dimensionality reduction continue to analyze the reduced data with traditional statistical tools, biomechanists are now also considering new problem formulations in which features extracted using PCA are used as inputs in machine learning models.

Two machine learning approaches, predictive modeling and data mining, serve different purposes than traditional inferential statistics. Predictive modeling is concerned with finding a function that optimally maps input data (e.g., kinematic waveforms) to a given output (e.g., disease status) with the goal of making accurate predictions in the future. One example of predictive modeling in biomechanics is myoelectric control of prostheses, where models are trained to recognize an individual’s intention based on myoelectric signals and the predicted intention is used to control the prosthesis (Oskoei and Hu, 2008). More recent efforts have centered around diagnostic and prognostic predictive models for neuromuscular and musculoskeletal pathologies (e.g., Schwartz et al., 2013), fall prediction (e.g., Wei et al., 2017), activity recognition to facilitate out-of-clinic patient monitoring (e.g., Biswas et al., 2015), and event detection to guide interventions such as deep brain stimulation (e.g., Pérez-López et al., 2016). The goal of data mining, on the other hand, is to discover new patterns in the data. Using clustering methods to identify subpopulations that exhibit different types of pathological gaits is an example of data mining (e.g., Rozumalski and Schwartz, 2009).

While applications of machine learning methods are expanding in movement biomechanics, critical evaluation of studies that apply them remains difficult. Machine learning approaches differ from the traditional statistical tools that biomechanists are trained to apply and interpret based on established reporting standards (e.g., p value for statistical significance). As the field becomes more data-intense and the use of machine learning continues to increase, good practices for conducting and reporting research at the intersection of biomechanics and machine learning are needed to ensure that conclusions are valid and reproducible. A discussion of this topic will also enable researchers to develop an intuition for the types of problems that machine learning can address more successfully than traditional statistics. Accordingly, the goal of this survey is to make machine learning efforts more visible and propose standards to increase the quality and impact of future research in this exciting area. To achieve this goal, we first review applications of machine learning that focus on neuromuscular and musculoskeletal diseases. We outline best practices for reporting the results of these analyses and common pitfalls we encountered in the literature. Finally, we offer suggestions for overcoming some of the challenges facing biomechanical data analytics and highlight opportunities where emerging techniques are likely to have great impact in upcoming years. Key terms are defined in Appendix A and our most important recommendations are summarized in the Conclusions section.

Section snippets

Literature search approach

We carried out a search for original research articles published up to December 31, 2017 using the PubMed/Medline database (1946-). Our search identified articles that used machine learning methods to study human movement biomechanics and was limited to studies of common musculoskeletal and neuromuscular diseases affecting mobility. We used search terms from three different categories to identify relevant studies: (1) movement biomechanics terms, such as gait, kinematics, and kinetics; (2)

Results

Our search yielded 3193 research articles, out of which 129, dating from 1996 to 2017, satisfied the inclusion criteria (Fig. 1A; Supplementary Table 1). The majority of studies focused on predictive tasks—classification (80.6%) and regression (11.6%)—while a few focused on data mining, in particular clustering tasks (7.8%). The most used algorithms were support vector machines, artificial neural networks, and generalized linear models (linear or logistic regression) for predictive modeling and

Discussion

The use of machine learning methods in movement biomechanics research is on the rise (Fig. 1A). From passively monitoring post-stroke patients with wearable devices to predicting outcomes of interventions in children with cerebral palsy, the range of applications where advanced analytics can improve rehabilitation research will continue to expand, particularly as wearable sensing generates vast amounts of data. The aim of this review is to bring to light machine learning efforts in movement

Acknowledgements

This work was funded by the National Institutes of Health (NIH) Grant U54EB020405. The authors would like to thank Jessica Selinger, Rachel Jackson, Łukasz Kidziński, Wolf Thomsen, and Jennifer Yong for their insightful feedback.

References (100)

  • K. Kaczmarczyk et al.

    Associations between gait patterns, brain lesion factors and functional recovery in stroke patients

    Gait Post.

    (2012)
  • N.L.W. Keijsers et al.

    Movement parameters that distinguish between voluntary movements and levodopa-induced dyskinesia in Parkinson’s disease

    Hum. Mov. Sci.

    (2003)
  • G. Kuntze et al.

    Multi-muscle activation strategies during walking in female post-operative total joint replacement patients

    J. Electromyogr. Kinesiol. Off. J. Int. Soc. Electrophysiol. Kinesiol.

    (2015)
  • D. Laroche et al.

    A classification study of kinematic gait trajectories in hip osteoarthritis

    Comput. Biol. Med.

    (2014)
  • A. Laudanski et al.

    Activity classification in persons with stroke based on frequency features

    Med. Eng. Phys.

    (2015)
  • P. Levinger et al.

    The application of support vector machines for detecting recovery from knee replacement surgery using spatio-temporal gait parameters

    Gait Post.

    (2009)
  • J. Liu et al.

    EMG feature assessment for myoelectric pattern recognition and channel selection: a study with incomplete spinal cord injury

    Med. Eng. Phys.

    (2014)
  • C.M.B. Magalhães et al.

    Increased hip internal abduction moment and reduced speed are the gait strategies used by women with knee osteoarthritis

    J. Electromyogr. Kinesiol. Off. J. Int. Soc. Electrophysiol. Kinesiol.

    (2013)
  • N. Mezghani et al.

    Mechanical biomarkers of medial compartment knee osteoarthritis diagnosis and severity grading: discovery phase

    J. Biomech.

    (2017)
  • S.P. Moustakidis et al.

    A fuzzy decision tree-based SVM classifier for assessing osteoarthritis severity using ground reaction force measurements

    Med. Eng. Phys.

    (2010)
  • C. Nüesch et al.

    Gait patterns of asymmetric ankle osteoarthritis patients

    Clin. Biomech. Bristol Avon

    (2012)
  • D. Ormoneit et al.

    Representing cyclic human motion using functional analysis

    Image Vis. Comput.

    (2005)
  • C. Pérez-López et al.

    Dopaminergic-induced dyskinesia assessment based on a single belt-worn accelerometer

    Artif. Intell. Med.

    (2016)
  • A.J. Ries et al.

    A data driven model for optimal orthosis selection in children with cerebral palsy

    Gait Post.

    (2014)
  • P.J. Rousseeuw

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis

    J. Comput. Appl. Math.

    (1987)
  • A. Rozumalski et al.

    Crouch gait patterns defined using k-means cluster analysis are related to underlying clinical pathology

    Gait Post.

    (2009)
  • Y. Sagawa et al.

    Are clinical measurements linked to the gait deviation index in cerebral palsy patients?

    Gait Post.

    (2013)
  • A. Samà et al.

    Estimating bradykinesia severity in Parkinson’s disease by analysing gait through a waist-worn sensor

    Comput. Biol. Med.

    (2017)
  • M.H. Schwartz et al.

    Femoral derotational osteotomy: surgical indications and outcomes in children with cerebral palsy

    Gait Post.

    (2014)
  • M.H. Schwartz et al.

    Predicting the outcome of intramuscular psoas lengthening in children with cerebral palsy using preoperative gait data and the random forest algorithm

    Gait Post.

    (2013)
  • L. Yu et al.

    A remote quantitative Fugl-Meyer assessment framework for stroke patients based on wearable sensor networks

    Comput. Methods Prog. Biomed.

    (2016)
  • E.B. Zwick et al.

    Classification of equinus in ambulatory children with cerebral palsy-discrimination between dynamic tightness and fixed contracture

    Gait Post.

    (2004)
  • C. Ahlrichs et al.

    Detecting freezing of gait with a tri-axial accelerometer in Parkinson’s disease patients

    Med. Biol. Eng. Comput.

    (2016)
  • H. Akaike

    A new look at the statistical model identification

    IEEE Trans. Autom. Control

    (1974)
  • M.N. Alam et al.

    Vertical ground reaction force marker for Parkinson’s disease

    PLoS ONE

    (2017)
  • L.K. Bachrach et al.

    Bone mineral acquisition in healthy Asian, Hispanic, black, and Caucasian youth: a longitudinal study

    J. Clin. Endocrinol. Metab.

    (1999)
  • A.S. Buchman et al.

    Associations between quantitative mobility measures derived from components of conventional mobility testing and Parkinsonian gait in older adults

    PloS One

    (2014)
  • N.A. Capela et al.

    Evaluation of a smartphone human activity recognition application with able-bodied and stroke participants

    J. NeuroEng. Rehabil. JNER

    (2016)
  • O. Chapelle et al.

    Semi-Supervised Learning

    (2010)
  • J.B. Copas

    Regression, prediction and shrinkage

    J. R. Stat. Soc. Ser. B Methodol.

    (1983)
  • S.L. Delp et al.

    OpenSim: open-source software to create and analyze dynamic simulations of movement

    IEEE Trans. Biomed. Eng.

    (2007)
  • O.A. Donoghue et al.

    Functional data analysis of running kinematics in chronic Achilles tendon injury

    Med. Sci. Sports Exerc.

    (2008)
  • Doshi-Velez, F., Kim, B., 2017. Towards a rigorous science of interpretable machine learning. ArXiv Prepr....
  • P. Drotár et al.

    Decision support framework for Parkinson’s disease based on novel handwriting markers

    IEEE Trans. Neural Syst. Rehabil. Eng. Publ. IEEE Eng. Med. Biol. Soc.

    (2015)
  • A. Duhamel et al.

    Functional data analysis for gait curves study in Parkinson’s disease

    Stud. Health Technol. Inform.

    (2006)
  • J.C. Dunn

    A fuzzy relative of the ISODATA process and its use in detecting compact well-separated clusters

    J. Cybern.

    (1973)
  • G.D. Fulk et al.

    Identifying activity levels and steps of people with stroke using a novel shoe-based sensor

    J. Neurol. Phys. Ther.

    (2012)
  • G.D. Fulk et al.

    Using sensors to measure activity in people with stroke

    Top. Stroke Rehabil.

    (2011)
  • Hastie, T.J. 2016. StatLearning - SELF PACED [WWW Document]. Lagunita. URL...
  • F.C. Huang et al.

    Movement distributions of stroke survivors exhibit distinct patterns that evolve with training

    J. Neuroeng. Rehabil.

    (2016)
  • Cited by (261)

    View all citing articles on Scopus
    View full text