Decoding Gait Signatures: Exploring Individual Patterns in Pathological Gait Using Explainable AI

Abstract

This study explores the application of machine learning (ML) to derive and analyze individual gait patterns (i.e., gait signatures) from ground reaction force data. This study leverages three datasets containing 2,092 individuals, including 1,283 cases with pathological gait, and addresses three key objectives: (1) Demonstrating the uniqueness of gait signatures in a large-scale dataset with heterogeneity introduced by patient data and various conditions. (2) Characterizing gait signatures using explainable artificial intelligence (XAI) to highlight specific features that contribute to their uniqueness. (3) Evaluating the reliability of gait signatures and their characterizations across different numbers of individuals and training samples per individual. The results show that ML can accurately differentiate unique gait patterns across healthy individuals and patients with pathological gait patterns, highlighting the importance of considering individual gait signatures in clinical gait analysis. The high reliability of a person’s unique gait signature may provide the basis for more personalized treatment decisions and rehabilitation programs, with XAI methods providing valuable insights into the key features that characterize individual gait. These results indicate that even more refined and personalized approaches are possible, extending beyond the conventional categories of pathology, age, and sex. This study provides a foundation for exploring the practical impact of gait signatures on rehabilitation, clinical diagnosis, and personalized treatment strategies.

Keywords

• Biomechanics
• gait analysis
• gait recognition
• ground reaction forces
• interpretability
• layer-wise relevance propagation