Diagnostic Accuracy of ChatGPT Model 5.1 in the Optical Characterization of Colorectal Lesions
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Background Optical characterization of colorectal lesions during colonoscopy is essential for determining the therapeutic strategy and predicting histology in real time. Artificial intelligence (AI) has emerged as a promising adjunctive tool to improve lesion characterization and reduce interobserver variability. Although computer-aided diagnosis systems have shown encouraging results, evidence regarding the performance of large language models in endoscopic lesion characterization remains limited. Aim The aim of this study is to evaluate the diagnostic accuracy of ChatGPT model 5.1 in differentiating adenomatous from non-adenomatous colorectal lesions during colonoscopy, using histopathology as the gold standard. Methods A retrospective diagnostic accuracy study was conducted including 93 colorectal lesions identified during colonoscopy. Optical descriptions of lesions were analyzed using ChatGPT 5.1, which classified lesions as adenomatous or non-adenomatous according to predefined criteria. Histopathological examination served as the reference standard. Diagnostic performance was evaluated through sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy using a 2×2 contingency matrix. Results Among 93 lesions, 35 (37.6%) were histologically confirmed adenomas. ChatGPT correctly identified 34 adenomas (true positives) and misclassified one lesion as non-adenomatous (false negative). Among non-adenomatous lesions, 46 we
Abstract
Background Optical characterization of colorectal lesions during colonoscopy is essential for determining the therapeutic strategy and predicting histology in real time. Artificial intelligence (AI) has emerged as a promising adjunctive tool to improve lesion characterization and reduce interobserver variability. Although computer-aided diagnosis systems have shown encouraging results, evidence regarding the performance of large language models in endoscopic lesion characterization remains limited. Aim The aim of this study is to evaluate the diagnostic accuracy of ChatGPT model 5.1 in differentiating adenomatous from non-adenomatous colorectal lesions during colonoscopy, using histopathology as the gold standard. Methods A retrospective diagnostic accuracy study was conducted including 93 colorectal lesions identified during colonoscopy. Optical descriptions of lesions were analyzed using ChatGPT 5.1, which classified lesions as adenomatous or non-adenomatous according to predefined criteria. Histopathological examination served as the reference standard. Diagnostic performance was evaluated through sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy using a 2×2 contingency matrix. Results Among 93 lesions, 35 (37.6%) were histologically confirmed adenomas. ChatGPT correctly identified 34 adenomas (true positives) and misclassified one lesion as non-adenomatous (false negative). Among non-adenomatous lesions, 46 were correctly classified (true negatives), while 12 were incorrectly categorized as adenomatous (false positives (FPs)). Sensitivity was 97.1%, specificity 79.3%, PPV 73.9%, NPV 97.9%, and overall diagnostic accuracy 86.0%. Most discrepancies occurred in lesions with inflammatory changes, superficial erosions, or mixed histologic patterns. Conclusions ChatGPT 5.1 demonstrated high sensitivity and excellent NPV for adenomatous lesion detection, suggesting potential utility as a supportive tool for optical diagnosis during colonoscopy. However, moderate specificity and the presence of FPs indicate that histopathological confirmation remains necessary before clinical implementation.