Gürses Ömer Alperen, Özüdoğru Anıl, Tuncay Figen, Kararti Caner
School of Physical Therapy and Rehabilitation, Department of Physiotherapy and Rehabilitation, Kırşehir Ahi Evran University, Merkez, Kırşehir, 40100, Türkiye.
Faculty of Medicine, Department of Physical Medicine and Rehabilitation, Kırşehir Ahi Evran University, Merkez, Kırşehir, 40100, Türkiye.
J Med Syst. 2025 Jun 3;49(1):73. doi: 10.1007/s10916-025-02207-x.
Large language models (LLMs) can contribute to treatment options and outcomes by assisting physiotherapists for conditions like osteoarthritis.
The objective of this early-stage cross-sectional study is to assess the alignment of large language models with physiotherapists in designing physiotherapy and rehabilitation programs for knee osteoarthritis.
Forty patients diagnosed with knee osteoarthritis were assessed using standardized clinical criteria. For each patient, individualized rehabilitation programs were created by three physiotherapists and by ChatGPT-4o and Gemini Advanced using structured prompts. The presence or absence of 50 clinically relevant rehabilitation parameters was recorded for each program. Chi-square tests were used to evaluate agreement rates between the LLMs and the physiotherapist-generated Consensus programs.
ChatGPT-4o achieved a 74% agreement rate with the physiotherapists' Consensus programs, while Gemini Advanced achieved 70%. Although both models showed high compatibility with general rehabilitation components, they demonstrated notable limitations in exercise specificity, including frequency, sets, and progression criteria. ChatGPT-4o performed as well as or better than Gemini in most phases, particularly in Phase 3, while Gemini showed lower consistency in balance and stabilization parameters.
ChatGPT-4o and Gemini Advanced demonstrate promising potential in generating personalized rehabilitation programs for knee osteoarthritis. While their outputs generally align with expert recommendations, notable gaps remain in clinical reasoning and the provision of detailed exercise parameters. These findings underscore the importance of ongoing model refinement and the necessity of expert supervision for safe and effective clinical integration.
大语言模型（LLMs）可以通过协助物理治疗师治疗骨关节炎等病症，为治疗方案和结果做出贡献。
这项早期横断面研究的目的是评估大语言模型在为膝关节骨关节炎设计物理治疗和康复计划方面与物理治疗师的契合度。
使用标准化临床标准对40名被诊断为膝关节骨关节炎的患者进行评估。对于每位患者，由三名物理治疗师以及ChatGPT-4o和Gemini Advanced使用结构化提示创建个性化康复计划。记录每个计划中50个临床相关康复参数的有无。使用卡方检验评估大语言模型与物理治疗师生成的共识计划之间的一致率。
ChatGPT-4o与物理治疗师的共识计划达成了74%的一致率，而Gemini Advanced达成了70%。尽管两个模型在一般康复组成部分方面都显示出高度兼容性，但它们在运动特异性方面表现出明显局限性，包括频率、组数和进展标准。ChatGPT-4o在大多数阶段的表现与Gemini相当或更好，特别是在第3阶段，而Gemini在平衡和稳定参数方面的一致性较低。
ChatGPT-4o和Gemini Advanced在为膝关节骨关节炎生成个性化康复计划方面显示出有前景的潜力。虽然它们的输出总体上与专家建议一致，但在临床推理和提供详细运动参数方面仍存在明显差距。这些发现强调了持续改进模型的重要性以及专家监督对于安全有效的临床整合的必要性。
