Utility of Artificial Intelligence in Diagnosing Image-Based Vaginal Pathology.
文献摘要
Importance Artificial intelligence (AI) tools are increasingly used by patients for health-related concerns before formal medical evaluation. However, data evaluating the diagnostic performance of these AI models for vaginal pathology are limited. Objectives The objectives of this study were to evaluate the diagnostic accuracy of publicly available AI models for image-based identification of vaginal pathology and compare their performance with expert physician review. Study design This diagnostic‑accuracy study assessed 25 deidentified clinical images representing various vaginal conditions. Four publicly accessible AI models (ChatGPT, Gemini, Claude, and Grok) and 2 board-certified urogynecology and reconstructive pelvic surgery physicians independently provided top 3 and top 1 diagnoses, malignancy identification, diagnostic confidence (0-100%), and triage recommendations using a standardized prompt. Results Experts achieved top 3 diagnostic accuracy of 84% and top 1 accuracy of 66%. AI systems demonstrated variable performance, with mean top 3 accuracy of 51.8% and a mean top 1 accuracy of 32.5%. ChatGPT and Gemini achieved the highest AI performance (68% and 68% top 3 accuracy; 44% and 52% top 1 accuracy, respectively). In malignant cases (n=3), the correct diagnosis was included in the top 3 in 83.5% of evaluations across both groups, with AI slightly more likely to rank malignancy first. AI systems recommended urgent evaluation more frequently than experts (77% vs 22%). Conclusions Publicly available AI models demonstrate moderate accuracy in identifying vaginal pathology from images but perform below expert clinicians. While some models approach expert performance in detecting malignancy, they tend to over-recommend urgent care, highlighting a limited but evolving role for patient-facing AI in image-based triage.