Optimized XGBRF-CatBoost model for accurate polycystic ovary syndrome prediction using ultrasound imaging

Abstract

Polycystic ovary syndrome (PCOS) is a multifactorial endocrine disorder characterized by hyperandrogenism, anovulation, oligomenorrhea, and ovarian microcysts, often resulting in infertility, obesity, and dermatological issues. This study proposes a hybrid machine learning (ML) framework for accurate PCOS prediction using ovarian ultrasound imaging and clinical parameters. A gradient regression-based multilayer perceptron neural network (GRMPNN) is employed for feature selection, followed by a stacked ensemble classifier combining extreme gradient boosted random forest (XGBRF) and CatBoost for final diagnosis. The dataset comprises 541 anonymized patient records from Ghosh Dastidar Institute for Fertility Research (GDIFR), incorporating 45 clinical, hormonal, and imaging features. Preprocessing includes normalization, noise reduction, and random oversampling to address class imbalance. Feature selection using univariate statistical testing and chi-square ranking identified 13 key attributes. The proposed XGBRF–CatBoost model achieved accuracy, precision, recall, and F1-score exceeding 98% across both benchmark datasets, outperforming principal component analysis (PCA) and neural fuzzy rough subset evaluating (NFRSE)-based models. This framework enhances diagnostic precision, reduces computational complexity, and supports scalable integration into clinical workflows. The findings underscore the potential of artificial intelligence (AI)-assisted tools in reproductive medicine and present a reproducible, interpretable approach for early PCOS detection.