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

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.

Keywords: Data science; Machine learning; Musculoskeletal; Neuromuscular.

Fig. 1.

Study Characteristics. (A) The use of data science methods in human movement biomechanics studies that focus on neuromuscular and musculoskeletal pathologies has increased exponentially in recent years. (B) The mostly commonly used algorithms in the reviewed articles were support vector machines, neural networks, and generalized linear models, including linear and logistic regression. (C) The number of observations (subjects) in most studies was small to moderate. (D) Wearable sensors were the leading source of information, following by optical motion tracking. (E) The most commonly studied conditions were Parkinson’s, stroke, cerebral palsy, and arthritis.

Fig. 2.

Model Complexity and the Bias-Variance Trade-off. In supervised learning tasks, bias is error that results from incorrect assumptions (e.g., fitting a linear model when the underlying relationship is not linear), causing algorithms to miss important relationships between features and labels, while variance is error that results from sensitivity to small variations, causing algorithms to model noise in the training data. The bias-variance tradeoff refers to the process of minimizing bias and variance in order to train models that can generalize well to non-training data. Simple models may be characterized by high bias and low variance (underfit to the training data), while complex models may be characterized by low bias and high variance (overfit to the training data).

Fig. 3.

Framework for Building Robust Predictive Models. (A) When the data are high-dimensional, feature engineering is first used to derive lower-dimensional representations of the data. This can be accomplished through domain knowledge (i.e., extract important gait characteristics) or automated techniques, such as principal component analysis, fast Fourier transforms, and other approaches. (B) After this step, the data are split into training and test sets, with the test set aside for performance evaluation. The model selection and training step, which uses only training data, may focus solely on parameter learning (e.g., learn regression coefficients) or may include one or two additional steps: feature selection and hyper-parameter tuning. Feature selection is performed to ensure that only relevant features are included in the model. This often improves model performance. Hyper-parameter tuning is performed to select additional parameters in complex models (e.g., select the degree of the polynomial if not a linear model). To avoid overfitting, both feature selection and hyper-parameter tuning are carried out using training data alone, which are further split into training and validation sets. (C) Last, the trained model is tested on held-out test data and comprehensive performance metrics that are more meaningful than model accuracy are reported.