900 Section - The Role of Generative AI and Large Language Models in Rehabilitation Treatment Planning: A Systematic Review

Wednesday, May 20, 2026

6:00 PM - 7:30 PM PT

Presenting Author/Autor expositor(s)

AD

Andrea Demeco, MD

Researcher
University of Catanzaro "Magna Graecia"
Catanzaro, Calabria, Italy

Co-Author/Coautor(s)

NM

Nicola Marotta, MD

Associate Professor
University of Catanzaro "magna Grecia"
Catanzaro, Calabria, Italy

Corresponding Author/Autor a quien dirigir la corr(s)

AR

Andrea Racinelli, MD

Doctor
University Magna Graecia, Department of Medical and Surgical Sciences
Catanzaro, Calabria, Italy

Co-Author/Coautor(s)

MM

Marco Megna, MD

Doctor
University Magna Graecia, Department of Medical and Surgical Sciences
Crotone, Calabria, Italy
BP

Bruno Pansera, MD

Doctor
University Magna Graecia, Department of Medical and Surgical Sciences
Melito di Porto Salvo, Calabria, Italy
Ad

Alessandro de Sire, MD

Associate Professor
University of Catanzaro "Magna Grecia"
Catanzaro, Calabria, Italy
AA

Antonio Ammendolia, MD

Full Professor
University of Catanzaro "magna Grecia"
Catanzaro, Calabria, Italy

Objectives : Individualized rehabilitation plans are essential to maximize clinical outcomes. With healthcare digitalization, generative artificial intelligence (AI), particularly large language models (LLMs), offers potential to streamline and enhance the planning of personalized rehabilitation interventions. This systematic review critically assessed the effectiveness, accuracy, and practical application of LLMs in rehabilitation treatment planning.

Design: Following PRISMA 2020 guidelines, a comprehensive search was conducted in PubMed, Web of Science, and Scopus until June 2025. The PICO framework included LLMs (P), their use in rehabilitation treatment planning (I), comparison with expert opinion or other AI systems (C), and the ability to generate valid rehabilitation plans (O). Studies involving non-generative AI or clinical decision systems were excluded. Methodological quality was assessed with the Joanna Briggs Institute tools. Data extraction was independently performed by three reviewers with a process to resolve conflicts.

Results: Eight studies met inclusion criteria. ChatGPT-4 and GPT-4o were most commonly used. Concordance between AI-generated content and expert opinion varied, reaching up to 98.75% after clinical revision. In musculoskeletal and cardiovascular rehabilitation, LLMs demonstrated high classification accuracy and consistent alignment with clinical guidelines. ChatGPT-4 outperformed prior versions and competitors in generating structured, readable plans. However, hallucinations and gaps in multimodal integration were observed, underscoring the importance of expert supervision.

Conclusion: LLMs show promising capabilities in supporting rehabilitation treatment planning, offering scalability and personalization. While initial results are encouraging, further high-quality studies are necessary to validate real-world effectiveness, mitigate hallucinations, and define clear clinical implementation strategies.