Artificial Intelligence in Colposcopy: Evaluation of ChatGPT as a Diagnostic and Predictive Support Tool

BackgroundColposcopy remains a critical step in the diagnostic algorithm following a positive HPV test, but it suffers from limitations such as inter-operator variability and reliance on clinical expertise. The integration of artificial intelligence (AI) could enhance diagnostic consistency and support decision-making, especially in resource-limited settings. ObjectiveTo evaluate the performance of ChatGPT, a generative AI language model, in the diagnostic interpretation of colposcopic images, and to compare its clinical decisions and histological predictions with those of expert colposcopists and real pathology outcomes. MethodsA retrospective study was conducted using 146 colposcopic cases sourced from the publicly available IARC Colposcopy Atlas. For each case, ChatGPT was provided with static images, patient age, and HPV status and asked to classify the case (physiological vs. pathological), recommend a clinical action (biopsy or not), and predict the histological outcome. The results were compared with expert colposcopist decisions and pathology reports. ResultsThere was a high agreement between ChatGPT and colposcopists in diagnostic impression (92.5%, {kappa} = 0.83) and biopsy decision (82.9%, {kappa} = 0.58). In biopsied cases (n = 95), ChatGPT matched the histological diagnosis in 88.4% of cases. Sensitivity and positive predictive value for detecting CIN2+ were 97.1% and 94.4%, respectively. In logistic regression models predicting CIN2+, ChatGPT-based models achieved comparable AUCs to colposcopist-based models (0.947 vs. 0.966, p = 0.129). ConclusionChatGPT demonstrated high concordance with expert evaluations and robust diagnostic performance. Its multimodal reasoning capabilities and accessibility suggest that language-based AI models may serve as valuable support tools in colposcopy, particularly where expertise is limited.