Investigation of the usability of artificial intelligence large language models in physiotherapy and rehabilitation of knee osteoarthritis

This study aims to examine the potential usability of artificial intelligence large language models (LLMs) in knee osteoarthritis (OA) physiotherapy and rehabilitation. Forty knee osteoarthritis patients were included in the study. This study compared physiotherapy programs for knee OA developed by three experienced physiotherapists "program recommended by physiotherapists (PRP)" and those generated by the LLMs ChatGPT-4 and Gemini Advanced. Forty patients were assessed using OARSI-recommended criteria. Rehabilitation programs were designed based on this data, with PRP programs developed through collaboration among the physiotherapists and individualized programs generated by each LLMs. The effectiveness of these programs was evaluated by examining the alignment between programs, using chi-square tests to determine statistical significance at p < 0.05. ChatGPT-4o and Gemini Advanced recommended ultrasound more frequently than PRP in Phase 1 (p<0.001), while PRP emphasized hip mobilization exercises (p<0.001), hip abduction, and hamstring curls (p<0.001, p<0.05). Phase 2 highlighted PRP favoring TENS, and hip stabilization exercises over ChatGPT-4o and Gemini Advanced (p<0.001). In Phase 3, ChatGPT-4o recommended NMES and RUS currents significantly more (p<0.001), while dynamic balance exercises aligned with PRP but were lower in Gemini Advanced (p<0.001). ChatGPT-4o exhibited discrepancies with the Consensus in 13 of the 50 parameters, achieving an agreement rate of 74%. In contrast, Gemini Advanced showed discrepancies in 15 parameters, corresponding to a 70% agreement rate. The average percentages of recommendations for each parameter were calculated as follows: In Phase 1, recommendations were 82.5% for PRP, 75% for ChatGPT4o, and 76.11% for Gemini Advanced; in Phase 2, 90.66% for PRP, 72.24% for ChatGPT4o, and 62.5% for Gemini Advanced; and in Phase 3, 82.15% for PRP, 84.42% for ChatGPT4o, and 68.08% for Gemini Advanced. This study highlights ChatGPT-4o and Gemini Advanced's potential to assist clinicians in creating patient-specific knee OA rehabilitation programs, while emphasizing the need for improved guideline adherence and data accuracy for full clinical integration.