Joint angle prediction and joint-type classification in human gait analysis using explainable deep reinforcement learning

Abstract

Human gait analysis is a key component of rehabilitation, prosthetics, and sports science, especially for clinical evaluation and the development of adaptive assistive technologies. Accurate joint-angle estimation and dependable joint-type classification remain difficult because of the complex temporal behavior of gait signals and the limited interpretability of many deep learning (DL) approaches. While recent techniques have enhanced predictive accuracy, their clinical applicability is often limited by insufficient transparency and adaptability in learning mechanisms. To overcome these limitations, this work proposes an integrated framework that unifies DL, reinforcement learning (RL), and explainable artificial intelligence (XAI). Stochastic depth neural networks (SDNN) are applied for joint-angle regression, whereas deep feature factorization networks (DFFN) are used for multi-class joint-type classification. Optimization is achieved using Q-learning (QL) and mutual information maximization (MIM), ensuring stable convergence and improved learning efficiency. To improve interpretability, the framework incorporates counterfactual and contrastive explanations, feature ablation studies, and prediction probability analysis. Experimental findings show that the SDNN MIM model attains an R² score of 0.9881, with RL rewards increasing from 0.997 to 0.999 during regression training. For joint-type classification, the DFFN MIM model achieves an accuracy of 0.95, with reward values improving from 0.90 to 0.98. These results demonstrate the effectiveness of the proposed framework in delivering accurate and interpretable gait predictions, supporting its relevance to biomechanics, healthcare, personalized rehabilitation, and intelligent assistive systems.