PCOS-XAI: An Interpretable Predictive Model for Polycystic Ovary Syndrome Using Feature Selection and Ensemble Classification

Abstract

Polycystic Ovary Syndrome (PCOS) is a disease that has spread across the globe and has become a significant health concern that mainly affects women of reproductive age. The detection, diagnosis, treatment, and management of the condition at an early stage are vital in order to lower the risk of long-term complications, primarily an elevated risk of type 2 diabetes and gestational diabetes. In line with advancements in computational methods, machine learning and ensemble learning techniques have drawn significant attention as a means of facilitating automated medical diagnosis. This paper aims to build a reliable, efficient, and interpretable PCOS diagnostic system that not only supports evidence-based decision-making but also provides global explanations of feature contributions. To identify the best-performing model and reduce the number of features, six machine learning algorithms-Logistic Regression, Random Forest, Decision Tree, Naive Bayes, Support Vector Machine, and k-Nearest Neighbors-have been employed, with Bayesian hyperparameter tuning applied for optimization. High-performing base learners, together with a meta-learner, were then stacked using an ensemble approach to further enhance predictive performance. Experiments were conducted on a publicly available PCOS dataset using two train-test split ratios (70:30 and 80:20). The results show that the stacking ensemble model along with optimized features demonstrated improved performance over individual baseline models.