Explainable AI for Generalizable PCOS Diagnosis: A Geographically Validated Ensemble Learning Approach With Feature Selection

Abstract

Diagnosing Polycystic Ovary Syndrome (PCOS) is challenging due to its varied symptoms and the absence of a single definitive test. This study develops a robust and interpretable machine learning framework to enhance PCOS diagnosis and its applicability across diverse patient populations. From an initial set of 45 clinical features, 23 were selected for their strong statistical and biological relevance to established PCOS diagnostic criteria. Our novel approach combines these features within a weighted ensemble of classifiers, which significantly outperformed individual models. The final model achieved a 94.34% accuracy and a strong AUC of 93.38%, surpassing previous benchmarks. Critically, the model demonstrated consistent and reliable performance across distinct geographic cohorts, validating its generalizability. Furthermore, the use of explainable AI techniques ensures the model's decisions are transparent and clinically interpretable for healthcare providers. These findings confirm that this ensemble-driven tool can serve as a reliable, scalable, and practical aid for the early and accurate detection of PCOS in clinical settings.