Exploring Large Language Models for Automated Gait Analysis

Biomechanical gait analysis is a critical tool in orthopedic diagnosis and rehabilitation, particularly for patients undergoing total knee arthroplasty. Traditional assessments, however, are time-intensive, subjective, and reliant on expert interpretation. In this study, we investigate the use of large language models (LLMs), specifically GPT-4o, to generate clinically relevant gait assessments based on spatiotemporal motion data. We collected gait recordings from 11 preoperative knee arthroplasty patients and obtained expert annotations from physiotherapists. Using a structured prompt engineering approach, we enabled GPT-4o to produce full-body gait descriptions, which were then evaluated in a blinded study by physiotherapy students and experts teaching gait analysis. Our results show that the LLM achieves comparable levels of correctness and clarity to human-generated assessment. Nonetheless, limitations such as the absence of pathological context, and evaluator variability highlight the need for further refinement. This work presents a proof of concept for integrating generative AI into clinical gait analysis and underscores its potential as an assistive tool in physiotherapy and orthopedic diagnostics. Additionally we made the data publicly available.