Can Chatgpt Predict Development of Ectopic Pregnancy? A Retrospective Cohort Analysis in Patients With Pregnancy of Unknown Location

Objective: To evaluate ChatGPT’s ability to predict ectopic pregnancy (EP) in patients with pregnancy of unknown location (PUL) using structured clinical data. Design: Retrospective cohort study with predictive modeling. Materials and Methods: Data were collected from 162 patients with PUL in 2024 and known outcomes (intrauterine pregnancy [IUP], spontaneous abortion [SAB], EP, or lost to follow-up [LTFU]) at a tertiary care center. Each case included demographics, symptoms, hCG levels and trends, ultrasound findings, and medical history. Structured case summaries were input into ChatGPT, prompting a prediction. Model outputs were compared to actual diagnoses. Sensitivity, specificity, and accuracy were calculated. Fisher’s exact and Mann-Whitney U tests were used to compare features between correctly and incorrectly predicted EPs. p<0.05 was considered significant. Results: ChatGPT correctly predicted 25.9% of final diagnoses, with best performance in identifying EPs (n=14; sensitivity 64.3%, specificity 31%). However, it overpredicted EP, assigning it to 111 cases. IUPs (n=29) and LTFUs (n=43) were never correctly predicted (sensitivity 0%). SABs (n=76) had a sensitivity of 30% and specificity of 67%. Of the 14 EP cases, 9 were correctly predicted and 5 misclassified. No significant differences were found between correctly and incorrectly predicted EPs in adnexal mass (p=0.51), free fluid (p=1.00), corpus luteum (p=0.58), prior miscarriage (p=1.00), hCG (p=0.28), or infertility history (p=0.36). Conclusions: ChatGPT demonstrated moderate sensitivity in identifying EPs but performed poorly for other outcomes. Its tendency to overpredict EP highlights limitations in diagnostic reasoning under uncertainty. Correct EP predictions likely reflect integration of multifactorial clinical cues rather than singular risk factors.