A Step Toward the Use of AI for Polyendocrine Metabolic Ovarian Syndrome (PMOS): Cost-Aware Risk Stratification with UncertaintyAware Triage and Conformal Prediction
Polyendocrine metabolic ovarian syndrome (PMOS) previously known as Polycystic Ovary Syndrome (PCOS) is a common endocrine disorder affecting women of reproductive age, yet its diagnosis remains challenging because accurate assessment often depends on investigations that are costly, less accessible, and difficult to scale in resource-constrained clinical settings. We develop and assess a cost-aware modelling framework for PCOS risk stratification, comparing an inexpensive-feature model based on demographic, symptom, anthropometric, and routine clinical variables with an augmented model that additionally incorporates laboratory and ultrasound features when fuller diagnostic work-up is available. Logistic Regression (LR) and Random Forest (RF) models are evaluated using discrimination and calibration metrics, including AUC, accuracy, F1, precision, recall, Brier score, and expected calibration error (ECE), alongside clinical utility through decision curve analysis (DCA). To support safer decision-making under constrained capacity, we further integrate conformal prediction to provide finite-sample coverage guarantees and controlled abstention. Across train-test and out-of-fold evaluations, the augmented model showed consistent performance gains over the inexpensive-feature model, with AUC increasing by 6.9% for LR and 7.4% for RF. LR exhibited more favourable calibration in most settings, while RF achieved higher precision. Decision curve analysis showed higher net benefit for the augmented model across clinically relevant threshold regions, although feature sensitivity analysis indicated that a compact subset of inexpensive predictors preserved over 80% of maximal AUC, supporting the practical value of lower-burden screening. Conformal prediction achieved 94.5% overall coverage with a 41.3% abstention rate, maintaining near-nominal validity across age and BMI subgroups. Under constrained referral capacity, prioritising highest-risk cases maximised net benefit at lower capacity levels, while uncertainty-aware deferral became more compatible with decision utility as available capacity increased. These findings support a proof-of-concept framework in which inexpensive information provides meaningful early risk stratification, while added laboratory and ultrasound inputs offer incremental value when more resource-intensive assessment is feasible.