Development and external validation of an AMH–TyG–BMI model to prompt structured clinician assessment for PMOS during pre-ART outpatient evaluation

Assessment of polyendocrine metabolic ovarian syndrome (PMOS), formerly polycystic ovary syndrome (PCOS), before assisted reproductive technology (ART) requires integrated clinical, biochemical, and ultrasonographic review. We developed and validated a model using anti-Müllerian hormone (AMH), triglyceride-glucose index (TyG), and body mass index (BMI). For future electronic health record integration, a pre-review triage flag would prompt structured clinician assessment for PMOS without assigning a diagnosis. This multicenter retrospective diagnostic prediction model study used a development cohort of 5,829 women, including 964 with PMOS. Validation included a same-center temporal cohort of 1,269 women (169 with PMOS) and an external cohort of 1,962 women (218 with PMOS). We assessed discrimination, calibration, and threshold-based triage performance. Development-derived thresholds targeting sensitivities of at least 90% and 95% were applied unchanged in validation cohorts. Incremental value was compared with AMH alone at matched sensitivity targets. The development-cohort area under the receiver operating characteristic curve (AUC) was 0.889 (95% confidence interval [CI], 0.878–0.900). AUCs were 0.907 (95% CI, 0.882–0.932) in temporal validation and 0.939 (95% CI, 0.923–0.955) in external validation. At the 90% and 95% targets, validation sensitivities were 92.9%–93.6% and 95.9%–96.8%, leaving 6.4%–7.1% and 3.2%–4.1% of PMOS cases not flagged. Specificities were 66.1%–78.4% and 49.2%–64.3%, respectively. Compared with AMH alone, the model generated 52.5–89.0 fewer false-positive flags per 1,000 women at the 90% target. At the 95% target, the reduction was 23.4–80.4 per 1,000 women. Both validation cohorts showed average overprediction (calibration-in-the-large, − 0.413 and − 0.182), slopes greater than 1 (1.190 and 1.330), and Brier scores of 0.068 and 0.055. The AMH–TyG–BMI model discriminated well, with modest incremental value beyond AMH alone. Fixed sensitivity-prioritized thresholds yielded high observed validation sensitivity, but specificity remained limited and false-positive flags were substantial. Calibration deviations mean that high AUCs alone do not justify implementation. Whether an electronic flag improves the timing or completeness of existing ART assessments remains unknown. The model cannot replace Rotterdam criteria or independently diagnose or exclude PMOS. Implementation requires local calibration assessment, prospective validation, and workflow evaluation.