327 Diagnostic Ability of ChatGPT Using Microscopic Descriptions in Pathology Reports

Abstract Introduction/Objective Artificial intelligence (AI) and large language models like ChatGPT are reshaping how clinicians and patients interact with pathology reports. Our goal was to evaluate ChatGPT’s diagnostic accuracy using WHO criteria and microscopic descriptions. Methods/Case Report Microscopic descriptions for gynecologic entities were obtained and input into ChatGPT-4o along with the WHO criteria for the corresponding entities. The gynecologic entities were selected based on frequency of diagnosis and input of microscopic description. Results ChatGPT correctly diagnosed 100% of cases based on microscopic descriptions: endometrioid carcinoma (32/32), serous carcinoma (18/18), clear cell carcinoma (5/5), CIN 1 & 3 (50/50, 50/50), mature teratoma (10/10), and carcinosarcoma (20/20). WHO-based accuracy was minimally lower with CIN 1 being misclassified as CIN 2. Conclusion ChatGPT showed excellent accuracy with pathologist-generated descriptions but was less accurate with WHO criteria alone. This may reflect concise and distinguishable clues within pathology reports. WHO’s extensive CIN 1 description may have altered ChatGPT’s interpretation and altered the classification. This highlights the ever evolving nature of large language models and the ability to interpret diagnostic descriptions.