Explainable Machine Learning for Non-invasive Screening of Polycystic Ovary Syndrome: Development, Internal Validation and Stepped-Care Evaluation.

Abstract

Polycystic ovary syndrome (PCOS) affects 10–13% of women of reproductive age, yet diagnosis is often delayed because confirmation requires hormonal assays and pelvic ultrasound. We developed and internally validated a machine learning model that estimates PCOS risk from 16 non-invasive history, symptom and examination features, and quantified what hormonal and ultrasound tests add. Using a public multicentre dataset of 541 women (177 with PCOS) from Kerala, India, logistic regression, support vector machine, random forest and XGBoost models were built in leakage-free pipelines, compared by nested cross-validation and evaluated once on a stratified 20% held-out test set with bootstrap confidence intervals, calibration and decision curve analysis (TRIPOD+AI).
The random forest, selected on nested cross-validation (AUC 0.891 ± 0.026), achieved a test AUC of 0.884 (95% CI 0.803–0.951), with sensitivity 77.8% and specificity 90.4% at a threshold fixed on training data; calibration-in-the-large was good (intercept 0.04; slope 1.41, 95% CI 0.97–2.27), and the model gave higher net benefit than referring all or no women. Logistic regression performed comparably (AUC 0.860; DeLong p = 0.220). Adding serum hormones gave no significant gain (AUC 0.893), whereas adding ultrasound raised it to 0.945. SHAP analysis identified excess hair growth, skin darkening, weight gain, irregular cycle and menstrual duration as the leading predictors. Non-invasive features can triage women for confirmatory testing; external validation is needed before clinical use.