Skip to main content
ISPRM 2026 Main banner
Final Program


Final Congress Guide

Functioning and Disability

600 Section - LLM-based Analysis of Patient-Reported Responses for Assessing ICF-based Functioning in Musculoskeletal Rehabilitation: A Development Study

Tuesday, May 19, 2026
6:00 PM - 7:30 PM PT

    Presenter(s)

  • Anke Scheel-sailer, MD (she/her/hers)

    Chief Physician Musculoskeletal Rehabilitation / PD Dr. med.
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

    • Presenting Author/Autor expositor(s)

  • EH

    Eva Hefti, MS

    PhD Candidate
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • NA

    Noora E. Angelva, MS

    PhD Candidate
    University of Bern, ARTORG Center, Bern, Switzerland
    Bern, Bern, Switzerland

    • Co-Author/Coautor(s)

  • PG

    Patrizia Götschmann, MS

    Assistant Physician
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • VF

    Valerie Frischknecht, MS

    PhD Candidate
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • TM

    Thimo Marcin, PhD

    Head of Bern Reha Research
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • BS

    Branislav Savic, PhD

    Scientist
    Department of Neurology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • SI

    Silviya Ivanova, MD

    Senior Physician / Dr. med.
    Department of Orthopaedic Surgery and Traumatology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • JB

    Johannes D. Bastian, MD

    Chief Physician, Head of Orthogeriatrics, Co-Head of Hip and Pelvis / Prof. Dr. med.
    Department of Orthopaedic Surgery and Traumatology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • SJ

    Simon Jung, MD

    Chief Physician / Prof. Dr. med.
    Department of Neurology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

  • TN

    Tobias Nef, PhD

    Technical Group Head / Prof. Dr.
    University of Bern, ARTORG Center, Bern, Switzerland
    Bern, Bern, Switzerland

  • MW

    Matthias Wilhelm, MD

    Chief Physician and Medical Director / Prof. Dr. med.
    Center for Rehabilitation & Sports Medicine Inselspital and Berner Reha Zentrum, Bern University Hospital, University of Bern, Bern, Switzerland
    Bern, Bern, Switzerland

Objectives : Automated documentation of functioning across the care continuum remains challenging. Large Language Models (LLMs) have the ability to process narrative texts. We aim to develop an LLM-supported, patient-reported solution to comprehensively assess ICF-based functioning.

Design: Narrative responses from participants undergoing musculoskeletal rehabilitation, with and without geriatric specialization, were collected through interviews using an intuitive item-rating task and a visual analogue scale (Swiss Ethics Project-ID 2025-01071). Interview items were based on the Functional Independence Measure (FIM), Clinical Functioning Information Tool (ClinFIT), and ICF Minimal Environmental Core Set. Interviews were audio-recorded, transcribed with locally run customized Whisper model, and de-identified. By using prompt-engineered local (gemma-2, gemma-3, Llama-3.1, Swiss-Apertus-8B, gpt-oss-20B) and online LLMs (Gemini-2.5-Pro, Gemini-3, Swiss-Apertus-70B), responses were classified into 4 categories: positive, negative, mixed, and missing. For validation the reference was established by two raters.

Results: In total, 197 answers from 7 participants were classified into 60 positive, 28 negative, 105 mixed, and 4 missing responses. The agreement between the two raters was excellent with Cohen's Kappa of 0.875. Only the online LLMs were able to generate a reasonable output. Best performance was shown by Gemini-3 with an accuracy of 87.3% (95% CI 81.8% - 91.3%), followed by Gemini-2.5-Pro with 79.7% (95% CI 73.4% - 85.1%), and Swiss-Apertus-70B with 64.5% (95% CI 57.4% - 71.1%). For the three models, recall per-class was 95.0%, 91.7%, and 46.7% (positive), 75.0%, 89.3%, and 32.1% (negative), 85.7%, 69.5%, and 84.8% (mixed), 100.0%, 100.0%, and 25.0% (missing). Precision per-class was 85.1%, 76.4%, and 87.5% (positive), 80.8%, 64.1%, and 52.9% (negative), 90.0%, 91.3%, and 65.9% (mixed), 100.0%, 66.7%, and 7.7% (missing).

Conclusion: It was possible to analyze comprehensive ICF-based topics using LLMs. This analysis gives the basis for in-depth validation with 35 participants undergoing rehabilitation. We will further refine classification and improve LLMs’ performances including extending to other non-communicable diseases.