DIAGNOSIS AND MANAGEMENT DECISIONS OF LARGE LANGUAGE MODELS IN PEDIATRIC SUPRACONDYLAR HUMERUS FRACTURES: AN AAOS GUIDELINE-BASED COMPARATIVE STUDY
Bozok Tıp Dergisi · 15.06.2026 · 16 (2) · 303-311
DOI: 10.16919/bozoktip.1909287 ↗Kurz gefasst
Leitlinie in der Zeitschrift Bozok Tıp Dergisi, online erschienen am 15.06.2026. Verfasst von Abdullah Alper Şahin und 1 weiteren Autor:innen. Die Arbeit nennt 13 Literaturangaben.
Abstract (Originalsprache)
Objective
This study aimed to evaluate the diagnostic performance and clinical safety of large language models (ChatGPT 5.2 Thinking, Gemini 3.1 Pro, and Claude Sonnet 4.6) in detecting pediatric supracondylar humerus (SCH) fractures, classifying them according to the Gartland system, and providing treatment recommendations based on American Academy of Orthopedic Surgeons (AAOS) clinical practice guidelines.Material and Methods: In this retrospective observational study, radiographs and clinical data from 128 pediatric patients presenting with elbow trauma (84 fractures and 44 non-fractures) were analyzed. The reference standard was established through consensus evaluation by orthopedic specialists according to AAOS guidelines. Anonymized AP and lateral elbow radiographs, along with age, sex, and brief clinical history, were provided to each model. Diagnostic performance was assessed using sensitivity, specificity, and overall accuracy. Agreement in Gartland classification was evaluated using Cohen’s kappa (κ). Clinical safety was examined by analyzing under-treatment and over-treatment risks in surgical decision recommendations.
Results
Gemini 3.1 demonstrated the highest sensitivity for fracture detection (96%) but had low specificity (41%). ChatGPT 5.2 showed a more balanced profile with both sensitivity and specificity of 64%, whereas Claude 4.6 demonstrated the lowest overall diagnostic accuracy (55%). In Gartland classification, Gemini 3.1 achieved the highest agreement (κ=0.60), while Claude 4.6 showed the lowest (κ=0.33). Regarding treatment recommendations, Gemini 3.1 avoided under-treatment of surgical cases but produced substantial over-treatment, whereas ChatGPT 5.2 missed 40% of surgically indicated cases.
Conclusion
Although LLMs may offer supportive insights in radiographic interpretation, their current performance demonstrates considerable risks of false-positive and false-negative decisions. These findings indicate that LLMs are not yet sufficiently reliable for independent clinical decision-making in pediatric SCH fractures, and expert oversight remains essential.
Zitieren
APA
Şahin, A. A., & Alparslan, M. (2026). DIAGNOSIS AND MANAGEMENT DECISIONS OF LARGE LANGUAGE MODELS IN PEDIATRIC SUPRACONDYLAR HUMERUS FRACTURES: AN AAOS GUIDELINE-BASED COMPARATIVE STUDY. Bozok Tıp Dergisi, 16(2), 303-311. https://doi.org/10.16919/bozoktip.1909287
Vancouver
Şahin AA, Alparslan M. DIAGNOSIS AND MANAGEMENT DECISIONS OF LARGE LANGUAGE MODELS IN PEDIATRIC SUPRACONDYLAR HUMERUS FRACTURES: AN AAOS GUIDELINE-BASED COMPARATIVE STUDY. Bozok Tıp Dergisi. 2026;16(2):303-311. doi:10.16919/bozoktip.1909287
BibTeX
@article{ahin2026,
author = {Şahin, Abdullah Alper and Alparslan, Murat},
title = {DIAGNOSIS AND MANAGEMENT DECISIONS OF LARGE LANGUAGE MODELS IN PEDIATRIC SUPRACONDYLAR HUMERUS FRACTURES: AN AAOS GUIDELINE-BASED COMPARATIVE STUDY},
journal = {Bozok Tıp Dergisi},
year = {2026},
volume = {16},
number = {2},
pages = {303--311},
doi = {10.16919/bozoktip.1909287}
}
EUVIMED