Objective: This study compared guideline-concordant decision-making performance between obstetrics and gynecology physicians and large language models (LLMs) in the management of cervical intraepithelial neoplasia (CIN). Additionally, internal consistency and error direction were evaluated from a patient safety perspective. Methods: Ten colposcopy-based clinical vignettes constructed according to the 2019 ASCCP risk-based management guidelines were administered to 161 physicians (residents and specialists) practicing in Türkiye and to multiple LLM configurations. Models were evaluated in their default settings without external browsing and tested in two independent runs under both standard and advanced reasoning (deep search/expert) modes. The primary outcome was the rate of guideline-concordant responses. Internal consistency was assessed using Cohen’s kappa, and inter-physician agreement was evaluated using Fleiss’ kappa. Responses were ordinally classified according to intervention intensity and categorized as over-triage or under-triage. Results: LLMs demonstrated higher mean accuracy compared with physicians (8.4 ± 0.8 vs. 6.4 ± 1.6; p < 0.001). Inter-physician agreement was low to moderate, and physician responses showed a tendency toward over-triage. Although most models exhibited high internal consistency, response variability was observed particularly in certain advanced reasoning modes, and not all models demonstrated equivalent stability. Advanced reasoning modes improved accuracy in some configurations; however, these gains were not consistently accompanied by increased stability. While under-triage rates were generally low among LLMs, marked performance declines were observed in rare and risk-threshold–sensitive scenarios. Conclusions: LLMs demonstrated high accuracy and relative decision stability in CIN management; however, model- and scenario-specific variability persists. These findings support positioning LLMs as human-supervised decision-support components rather than autonomous clinical decision-makers.

Background/Objectives: Borderline ovarian tumors can overlap with benign and malignant adnexal masses in their preoperative features. The hemoglobin–albumin–lymphocyte–platelet (HALP) score reflects inflammatory and immunonutritional status. We compared HALP profiles across benign, borderline, and malignant epithelial ovarian tumors and secondarily explored their contribution beyond the risk of malignancy index (RMI). Methods: This retrospective single-center study included 464 women treated surgically for serous or mucinous epithelial ovarian tumors between 2015 and 2025: 164 benign cystadenomas, 78 borderline tumors, and 222 invasive carcinomas. HALP was compared among groups with additional adjustment for age and menopausal status. Regression and model-performance analyses were performed only in the borderline and malignant groups (n = 300). Results: Median HALP was lower in malignant tumors [28.6 (IQR 22.4)] than in benign [38.3 (23.8)] and borderline tumors [40.5 (21.4)] (both p < 0.001), whereas benign and borderline tumors were similar (p = 0.948). After adjustment, HALP remained lower in malignant tumors than in benign (adjusted median difference, −10.49; 95% CI, −15.47 to −4.55) and borderline tumors (−11.86; 95% CI, −18.63 to −5.84). In exploratory malignant-versus-borderline modeling, HALP was associated with malignant histology alone, but this association was attenuated after accounting for age and log-transformed RMI; adding HALP did not significantly improve model fit or discrimination. Conclusions: Preoperative HALP was consistently lower in malignant tumors, while borderline tumors resembled benign lesions. HALP may reflect systemic immunonutritional differences across pathological groups. Its incremental predictive role beyond established preoperative measures remains uncertain and requires external validation.