The study, published in the Journal of Ovarian Research, took the form of a multicenter retrospective diagnostic prediction model study. The development cohort included 5,829 women, of whom 964 had PMOS. Validation was performed in two independent datasets: a temporal cohort of 1,269 women (169 with PMOS) drawn from the same centers at a later time, and a fully external cohort of 1,962 women (218 with PMOS). The investigators evaluated discrimination, calibration and threshold-based triage performance, applying thresholds derived in the development cohort—set to achieve sensitivities of at least 90% and 95%—unchanged in the validation cohorts.

Discrimination, the model’s ability to separate women with PMOS from those without, was strong across all datasets. The area under the receiver operating characteristic curve (AUC) was 0.889 (95% confidence interval, 0.878–0.900) in the development cohort, rising to 0.907 (95% CI, 0.882–0.932) in temporal validation and 0.939 (95% CI, 0.923–0.955) in external validation. These figures indicate that the three-variable combination captures much of the signal that distinguishes PMOS in this pre-ART population.

At the sensitivity-prioritized thresholds, performance in validation was consistent with the design targets. Observed sensitivities ranged from 92.9% to 93.6% at the 90% target and from 95.9% to 96.8% at the 95% target, meaning that only 3.2% to 7.1% of PMOS cases went unflagged. The trade-off came in specificity, which ranged from 66.1% to 78.4% at the 90% target and fell to 49.2% to 64.3% at the 95% target—so a substantial number of women without PMOS would still receive a flag and undergo assessment.

The comparison with AMH alone is where the added value of the model becomes concrete. At matched sensitivity targets, the full AMH–TyG–BMI model generated 52.5 to 89.0 fewer false-positive flags per 1,000 women screened at the 90% sensitivity target, and 23.4 to 80.4 fewer per 1,000 at the 95% target. In practical terms, adding the metabolic markers to AMH reduces the burden of unnecessary downstream assessments, although the authors characterize this incremental benefit as modest rather than transformative.

Calibration—the agreement between predicted probabilities and observed PMOS frequency—told a more cautionary story. Both validation cohorts showed average overprediction, with calibration-in-the-large values of −0.413 and −0.182, and calibration slopes greater than 1 (1.190 and 1.330). Brier scores were 0.068 and 0.055. The authors are explicit that these deviations mean the high AUCs alone do not justify implementation: a model can rank patients well while systematically misstating absolute risk, which matters when a flag is meant to guide who gets reviewed.

The study’s authors also draw a firm boundary around what the model cannot do. It cannot replace the Rotterdam criteria, the standard diagnostic framework for PCOS/PMOS, and it cannot independently diagnose or exclude the condition. Its intended role is narrower: to prompt a structured clinician assessment during pre-ART evaluation so that cases are less likely to be missed before fertility treatment begins. Whether an electronic flag actually improves the timing or completeness of existing ART assessments remains unknown and was not tested in this retrospective design.

As with any retrospective diagnostic model study, several limitations shape the interpretation. The cohorts were drawn from women already undergoing pre-ART evaluation, so performance in other populations or care settings is not established. The calibration deviations observed in validation would need to be addressed before probabilities from the model could be trusted as absolute risks. The authors state that implementation would require local calibration assessment, prospective validation, and evaluation of how the flag fits into clinical workflow.

For now, the AMH–TyG–BMI model stands as a carefully validated triage instrument rather than a ready-made diagnostic test. Its strengths are real: strong and reproducible discrimination across development, temporal and external cohorts, high observed sensitivity at fixed thresholds, and a measurable reduction in false-positive flags compared with AMH alone. Its weaknesses are equally clear: limited specificity, substantial residual false positives, calibration drift, and an unproven effect on actual clinical outcomes. The next step for this line of research is prospective testing in live ART clinics, where the true value of an automated flag will be judged not by AUCs but by whether women with PMOS receive timely, complete assessment before treatment.

Subject of Research: Development and external validation of an AMH–TyG–BMI prediction model to triage women for structured PMOS assessment before assisted reproductive technology

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

Article References: Hu, J., Zhu, Q., Shi, J., Zhang, Y., Li, J., & Zhou, L. (2026). Development and external validation of an AMH–TyG–BMI model to prompt structured clinician assessment for PMOS during pre-ART outpatient evaluation. Journal of Ovarian Research. https://doi.org/10.1186/s13048-026-02321-5

Image Credits: AI Generated

DOI: 10.1186/s13048-026-02321-5

Keywords: PMOS, PCOS, anti-Müllerian hormone, triglyceride-glucose index, body mass index, prediction model, assisted reproductive technology, diagnostic validation, triage, calibration, Rotterdam criteria, Development