Leveraging Artificial Intelligence for Personalized Rehabilitation Programs for Head and Neck Surgery Patients

Abstract

Background: Artificial intelligence (AI) and large language models (LLMs) are increasingly used in healthcare, with applications in clinical decision-making and workflow optimization. In head and neck surgery, postoperative rehabilitation is a complex, multidisciplinary process requiring personalized care. This study evaluates the feasibility of using LLMs to generate tailored rehabilitation programs for patients undergoing major head and neck surgical procedures. Methods: Ten hypothetical head and neck surgical clinical scenarios were developed, representing oncologic resections with complex reconstructions. Four LLMs, ChatGPT-4o, DeepSeek V3, Gemini 2, and Copilot, were prompted with identical queries to generate rehabilitation plans. Three senior clinicians independently assessed their quality, accuracy, and clinical relevance using a five-point Likert scale. Readability and quality metrics, including the DISCERN score, Flesch Reading Ease, Flesch–Kincaid Grade Level, and Coleman–Liau Index, were applied. Results: ChatGPT-4o achieved the highest clinical relevance (Likert mean of 4.90 ± 0.32), followed by DeepSeek V3 (4.00 ± 0.82) and Gemini 2 (3.90 ± 0.74), while Copilot underperformed (2.70 ± 0.82). Gemini 2 produced the most readable content. A statistical analysis confirmed significant differences across the models (p < 0.001). Conclusions: LLMs can generate rehabilitation programs with varying quality and readability. ChatGPT-4o produced the most clinically relevant plans, while Gemini 2 generated more readable content. AI-generated rehabilitation plans may complement existing protocols, but further clinical validation is necessary to assess their impact on patient outcomes.