Explainable artificial intelligence for gait analysis: advances, pitfalls, and challenges - a systematic review.

Area of Science:

• Biomechanics
• Artificial Intelligence
• Clinical Informatics

Background:

• Machine learning (ML) models excel at gait data analysis but often lack transparency.
• Explainable artificial intelligence (XAI) offers a solution to enhance ML model interpretability for clinical settings.

Purpose of the Study:

• To systematically review the application of XAI in gait analysis.
• To assess XAI methods, performance, and clinical utility in diverse patient populations.

Main Methods:

• Systematic review of 31 studies identified through a comprehensive database search (PROSPERO registration: CRD42024622752).
• Categorization of XAI approaches into model-agnostic, model-specific, and hybrid methods.
• Analysis of interpretation techniques including SHAP, LIME, Grad-CAM, and attention mechanisms.

Main Results:

• XAI successfully identified biomechanically relevant gait features (e.g., stride length, joint angles) differentiating pathological gaits.
• Model-agnostic methods (SHAP, LIME) were most frequently applied.
• Studies covered various clinical populations, including Parkinson's disease, stroke, and sarcopenia.

Conclusions:

• XAI demonstrates potential to bridge the gap between ML predictive performance and clinical interpretability in gait analysis.
• Standardization, validation, and balancing accuracy with transparency are crucial for widespread clinical adoption.
• Further research is needed to refine XAI frameworks and evaluate their real-world applicability across different gait disorders.