Related Experiment Video
Updated: Mar 31, 2026
Standardized Histomorphometric Evaluation of Osteoarthritis in a Surgical Mouse Model
Published on: May 6, 2020
Evaluating Large Language Model Adherence to AAOS Knee Osteoarthritis Guidelines: A Comparative Study of ChatGPT and
Cemil Yıldız1, Mehmet Yiğit Gökmen2, Çetin Utlu3
1Department of Orthopaedics and Traumatology, Gulhane Faculty of Medicine, University of Health Sciences, Ankara, Türkiye.
Purpose:
This study evaluated how closely large language models (LLMs), specifically ChatGPT (OpenAI) and NotebookLM (Google), adhere to orthopedic guidelines. The objective was to determine whether AI-generated reasoning aligns with the 2021-2022 American Academy of Orthopaedic Surgeons (AAOS) clinical practice guidelines for knee osteoarthritis (OA).
Methods:
A mixed-methods design combined quantitative concordance scoring with qualitative content analysis. Thirty-three decision points covering non-arthroplasty and surgical management were derived from AAOS guidelines. Structured Population-Intervention-Comparison-Outcome (PICO) prompts were presented to each model. Two orthopedic surgeons independently rated all outputs using a four-domain rubric assessing accuracy, evidence reasoning, additional information, and knowledge integration (0-4 scale). Concordance was classified as full (4), partial (3), or discordant (≤ 2), with disagreements resolved through consensus. Inter-rater reliability was almost perfect (weighted κ = 0.87).
Results:
ChatGPT achieved a mean composite score of 3.67 ± 0.92, and NotebookLM 3.55 ± 0.87, with no significant difference between models (p = 0.18). Full concordance with AAOS recommendations occurred in 84.8% of ChatGPT responses and 75.8% of NotebookLM responses. Both models performed consistently in high-evidence domains such as NSAID therapy, tranexamic acid use, and weight-loss counseling. Variability increased in limited-evidence or technology-driven areas. Partial concordance reflected the omission of evidence qualifiers, while discordant responses involved overstated or speculative interpretations.
Conclusion:
Both LLMs demonstrated strong alignment with evidence-based orthopedic reasoning. ChatGPT showed slightly higher fidelity to recommendation strength, whereas NotebookLM provided broader contextual interpretation. Structured, guideline-oriented prompting may enhance AI reasoning consistency and support the potential role of LLMs as complementary tools for evidence translation and orthopedic education.
Supplementary Information:
The online version contains supplementary material available at 10.1007/s43465-025-01653-6.
