Deep Learning Models To Predict Ankle Joint Reaction Force During Gait

Accurately quantifying ankle joint reaction forces (JRFs) is critical for understanding their role in joint health and disorders, such as idiopathic arthritis, which significantly impacts mobility and quality of life. Conventional methods, including motion capture systems and inverse dynamics analysis, provide reliable results but are limited by their cost, complexity, and impracticality for clinical application. This study introduces a deep learning-based framework employing Sequence-to-Sequence (Seq2Seq) models to predict ankle JRFs using kinematic data derived from marker-based motion capture. Model performance was evaluated using leave-one-subject-out cross-validation to ensure robust and unbiased results. The GRU-based models achieved the lowest Root Mean Square Errors (RMSEs) across all force components, demonstrating their superior predictive capability compared to other architectures. The proposed framework achieves high accuracy while circumventing the need for expensive and complex laboratory setups, paving the way for realtime estimation of JRFs using more accessible tools such as videobased pose estimation technologies. However, limitations include reliance on healthy subject data and inverse dynamics-derived forces, underscoring the need for future studies incorporating diverse populations and musculoskeletal modeling.