Methods for estimating lower-limb joint kinematics from six stride descriptors

Abstract

The biomechanical analysis of running provides insights into an athlete health status and running style by measuring joint angular kinematics over multiple gait cycles. However, obtaining these measurements typically requires expensive laboratory-based motion capture systems, limiting accessibility in real-world settings. This study aims to develop a framework, named Stride2Kinematics, to estimate joint kinematics from stride descriptors, enabling gait analysis outside the laboratory. We propose the use of both traditional and deep-learning approaches to estimate angular gait variables from a set of six stride descriptors. Specifically, we employed Nadaraya-Watson estimator and trained a multilayer perceptron and different autoencoder architectures to reduce the dimensionality of joint kinematics and generate latent spaces that align with the stride descriptors. Our results show that the Nadaraya-Watson estimator and a variational autoencoder achieved the best performance (mean absolute errors of 4.7932∘ and 4.7462∘, and correlations of 0.9704 and 0.9712, respectively). Latent spaces analysis revealed distinct running patterns, and explainable artificial intelligence techniques identified key gait cycle segments influencing the estimation of joint kinematics. This study demonstrates that angular gait variables can be effectively estimated using wearable-compatible stride descriptors, offering a cost-effective alternative for injury prevention, performance optimization, and real-world gait assessments.