Machine Learning in Biomechanics: Key Applications and Limitations in Walking, Running, and Sports Movements

TL;DR

This chapter surveys how ML can transform gait and sports biomechanics by enabling pose estimation, feature estimation, event detection, data exploration, and automated classification across walking, running, and sports movements. It details biomechanics and ML workflows, identifies central limitations—data/annotation availability and explainability—and advocates physics-informed ML and cross-disciplinary collaboration to improve robustness, interpretability, and real-world applicability. Markerless pose estimation, field-based data collection, and time-series analysis emerge as key avenues, but current accuracy gaps and limited benchmark datasets temper immediate clinical deployment. Overall, the work highlights practical pathways to accelerate biomechanical analysis while outlining the infrastructural and methodological shifts needed for reliable, generalizable ML-enabled biomechanics.

Abstract

This chapter provides an overview of recent and promising Machine Learning applications, i.e. pose estimation, feature estimation, event detection, data exploration & clustering, and automated classification, in gait (walking and running) and sports biomechanics. It explores the potential of Machine Learning methods to address challenges in biomechanical workflows, highlights central limitations, i.e. data and annotation availability and explainability, that need to be addressed, and emphasises the importance of interdisciplinary approaches for fully harnessing the potential of Machine Learning in gait and sports biomechanics.

Figures (5)

• Figure 1: Overview of three commonly applied biomechanical testing approaches: i) biomechanical testing using marker-based infrared camera systems together with force platforms and electromyography in laboratories can be considered the gold standard, where achieving the highest level of precision and accuracy are prioritised over material and personal resources; ii) biomechanical testing using wearable sensors, e.g. IMUs in natural field-based settings, offering both long-term recordings and real-time feedback; and iii) biomechanical testing utilising portable sensors, e.g. RGB video cameras is well-suited for conducting assessments in natural field-based settings, where the emphasis is on minimising constraints, even at the expense of spatial and temporal accuracy. The number of squares next to a measurement approaches describe its prevalence for the biomechanical testing of a movement task (walking, running, sports movements). Copyright (left photo): Helene Sorger / St.Pölten University of Applied Sciences.
• Figure 2: In the realm of biomechanics, typical ML workflows are characterised by the following phases: data preprocessing, optionally feature engineering, ML model training (which can incorporate feature learning), and ML model evaluation.
• Figure 3: Overview of the key ML applications (green boxes), i.e. pose estimation, feature estimation, event detection, data exploration & clustering, and automated classification, and their integration into the phases of the biomechanical workflow. The key limitations (orange boxes) associated with the use of ML in biomechanics include data and annotation availability and explainability.
• Figure 4: Examples of input and output pairs for the key ML applications, i.e. pose estimation, feature estimation, event detection, data exploration & clustering, as well as automated classification.
• Figure 5: Overview of different data exploration approaches for biomechanical data. (1) t-Distributed Stochastic Neighbour Embedding visualisations of principal components (left) and embeddings obtained via metric learning using a Siamese Neural Network (right) for various walking patterns associated with cerebral palsy krondorfer2021deep. (2) t-Distributed Stochastic Neighbour Embedding latent space visualisation of a Variational Autoencoder trained on spinal posture data (black circles correspond to training data, blue to testing data, and orange to sampled latent vectors from the latent space of the trained Variational Autoencoder for data generation Dindorf.2022vertebrae).