An Enhanced PCOS-MLPro Diagnosis Model Using Multi-Dimensional Clinical Indicators

PCOS-MLPro (Polycystic Ovary Syndrome - ML Predictor) advances from previous PCOS prediction works by suggesting a state-of-the-art and clinically-oriented machine learning pipeline that leverages high-dimensional structured data combined with optimized preprocessing toward more accurate diagnosis. In the current study, 920 de-identified patient records from 15 multi-specialty hospitals in South India with 55 clinical, biochemical, symptomatic and metabolic features that are relevant for assessing PCOS are integrated. To overcome minority under-representation, the pre-processing pipeline was enhanced with adaptive KNN-based imputation, interquartile-range outlier handling, min-max normalization and ADASYN-driven class balancing. Five supervised models, which included Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), Light GBM (LGBM) and an optimized shallow MLP, were evaluated using stratified 10-fold crossvalidation with Bayesian hyperparameter optimization. The best among them for this task is the optimized XGBoost classifier achieving 96.1 % accuracy, 0.952 F1-score and an AUROC of 0.965, a substantial improvement over the previously reported baselines. Global interpretability based on SHAP and local explanation via LIME both converge on AMH, LH/FSH ratio, BMI, the severity of acne and cycle irregularity as being dominant predictors pointing to their clinical significance. More sanity checks were enforced to detect conflicting inputs before inference, improving robustness in practical scenarios. PCOS-MLPro shows that expanded clinical parameters, stable preprocessing, and lightweight model enhancements can make a drastic contribution to improving interpretability as well as convenience on implementation for PCOS testing methods.