DeepTherapy: A mobile platform for osteoarthritis rehabilitation utilizing chain-of-thought reasoning and deep learning

Abstract

Objectives: To develop and evaluate an AI-driven mobile platform that integrates deep learning-based exercise analysis with large language model (LLM) feedback for enhancing osteoarthritis (OA) rehabilitation accessibility and effectiveness. Methods: A deep learning framework was developed using Long Short-Term Memory (LSTM) architecture to classify exercise phases from video data of 10 rehabilitation exercises. The dataset consisted of approximately 800,000 frames collected from 20 healthy volunteers. A feedback system utilizing chain-of-thought reasoning in LLMs (GPT-4o and Claude 3.5 Sonnet) was implemented to generate targeted corrective feedback. Evaluation was conducted with OA patients (n=2) and physiotherapists (n=7) using the Intraclass Correlation Coefficient (ICC) and Likert scales. Results: The developed LSTM models achieved 97.8% accuracy in exercise phase classification. Strong agreement between system-generated scores and expert evaluations was demonstrated (ICC=0.85). Physiotherapists slightly preferred Claude's outputs (52.4% vs 47.6%) but rated GPT-4o higher on clinical relevance (4.57/5 vs 4.13/5), clarity (4.71/5 vs 4.38/5), and helpfulness (4.50/5 vs 4.29/5). Conclusions: DeepTherapy effectively addresses critical limitations in rehabilitation monitoring by providing qualitative movement assessment, identifying incorrect movements, and offering detailed guidance on technique improvement, potentially increasing rehabilitation accessibility while maintaining quality of care.