Explainable AI-Based PCOS Detection Using SMOTE-Balanced Clinical Markers

Abstract

Polycystic Ovary Syndrome (PCOS) is a disease that plagues 8 to 13% of women of child bearing age; though only about 70 percent of the cases are diagnosed. Available machine-learning (ML) studies on PCOS detection rarely consider issues of multicollinearity in sets of hormonal features, fail to explicitly rectify the imbalance of classes and lack clinically interpretable prediction. The current research proposes a single end-to-end and reproducible ML pipeline, which simultaneously addresses all these three shortcomings. On a combined dataset (n=541 patients in Kerala, India) of merged patient data, we use iterative Variance Inflation Factor (VIF) elimination to reduce 41 raw features to 19 predictors that do not exhibit multicollinearity (all VIF< 5). Synthetic Minority Oversampling Technique (SMOTE) is then used to correct a 2:1 imbalance in training set classes. There are four variants of XGBoost, two variants of CatBoost, and K-Nearest Neighbours whose classifiers are benchmarked under the same experimental conditions on a held-out test set of 162 records. Accuracy of the SMOTE-tuned XGBoost model is 90.12%, with precision of 84 percent in PCOS-positive, recall of 84 percent, and an F1-score of 0.84, which is better than CatBoost (88.27 per cent) and KNN (86.42 per cent). SHAP (SHapley Additive exPlanations) analysis supports the assertion that the quantity of follicles, FSH/LH ratio, and the level of LH should be considered the most prevalent predictors, and that the evidence provided by the analysis can be interpreted by clinicians and corresponds to the Rotterdam diagnostic criteria. Such a pipeline provides a proven and multicollinearity-controlled foundation that supports reproducible PCOS screening studies.