Diagnostic performance of ChatGPT in tibial plateau fracture in knee X-ray

Mohammadreza Mohammadi; Sara Parviz; Parinaz Parvaz; Mohammad Mahdi Pirmoradi; Mohammad Afzalimoghaddam; Hadi Mirfazaelian

doi:10.1007/s10140-024-02298-y

Diagnostic performance of ChatGPT in tibial plateau fracture in knee X-ray

Emerg Radiol. 2024 Nov 30. doi: 10.1007/s10140-024-02298-y. Online ahead of print.

Authors

Mohammadreza Mohammadi¹, Sara Parviz², Parinaz Parvaz³, Mohammad Mahdi Pirmoradi¹, Mohammad Afzalimoghaddam^{1

4}, Hadi Mirfazaelian⁵

Affiliations

¹ Emergency Medicine Department, Tehran University of Medical Sciences, Tehran, Iran.
² Musculoskeletal Imaging Research Center (MIRC), Tehran University of Medical Sciences, Tehran, Iran.
³ Radiology Department, Tehran University of Medical Sciences, Tehran, Iran.
⁴ Prehospital and Hospital Emergency Research Center, Tehran University of Medical Sciences, Tehran, Iran.
⁵ Prehospital and Hospital Emergency Research Center, Tehran University of Medical Sciences, Tehran, Iran. H-Mirfazaelian@sina.tums.ac.ir.

PMID: 39613920
DOI: 10.1007/s10140-024-02298-y

Abstract

Purpose: Tibial plateau fractures are relatively common and require accurate diagnosis. Chat Generative Pre-Trained Transformer (ChatGPT) has emerged as a tool to improve medical diagnosis. This study aims to investigate the accuracy of this tool in diagnosing tibial plateau fractures.

Methods: A secondary analysis was performed on 111 knee radiographs from emergency department patients, with 29 confirmed fractures by computed tomography (CT) imaging. The X-rays were reviewed by a board-certified emergency physician (EP) and radiologist and then analyzed by ChatGPT-4 and ChatGPT-4o. The diagnostic performances were compared using the area under the receiver operating characteristic curve (AUC). Sensitivity, specificity, and likelihood ratios were also calculated.

Results: The results indicated a sensitivity and negative likelihood ratio of 58.6% (95% CI: 38.9 - 76.4%) and 0.4 (95% CI: 0.3-0.7) for the EP, 72.4% (95% CI: 52.7 - 87.2%) and 0.3 (95% CI: 0.2-0.6) for the radiologist, 27.5% (95% CI: 12.7 - 47.2%) and 0.7 (95% CI: 0.6-0.9) for ChatGPT-4, and 55.1% (95% CI: 35.6 - 73.5%) and 0.4 (95% CI: 0.3-0.7) for ChatGPT4o. The specificity and positive likelihood ratio were 85.3% (95% CI: 75.8 - 92.2%) and 4.0 (95% CI: 2.1-7.3) for the EP, 76.8% (95% CI: 66.2 - 85.4%) and 3.1 (95% CI: 1.9-4.9) for the radiologist, 95.1% (95% CI: 87.9 - 98.6%) and 5.6 (95% CI: 1.8-17.3) for ChatGPT-4, and 93.9% (95% CI: 86.3 - 97.9%) and 9.0 (95% CI: 3.6-22.4) for ChatGPT4o. The area under the receiver operating characteristic curve (AUC) was 0.72 (95% CI: 0.6-0.8) for the EP, 0.75 (95% CI: 0.6-0.8) for the radiologist, 0.61 (95% CI: 0.4-0.7) for ChatGPT-4, and 0.74 (95% CI: 0.6-0.8) for ChatGPT4-o. The EP and radiologist significantly outperformed ChatGPT-4 (P value = 0.02 and 0.01, respectively), whereas there was no significant difference between the EP, ChatGPT-4o, and radiologist.

Conclusion: ChatGPT-4o matched the physicians' performance and also had the highest specificity. Similar to the physicians, ChatGPT chatbots were not suitable for ruling out the fracture.

Keywords: Artificial intelligence; ChatGPT; Diagnosis; Emergency medicine; Radiology; Tibial plateau fracture.