Stability-aware hybrid optimisation-based feature selection and stacking classification for PCOS prediction

Polycystic Ovary Syndrome (PCOS) is one of the most common endocrine diseases in women of reproductive age all over the world. It is estimated that 6–13% of women are affected worldwide, and a significant percentage of them go undiagnosed because of their heterogeneity. The increased rate of PCOS is directly associated with lifestyle transition, metabolic stress, and hormonal imbalance. PCOS is linked to infertility, metabolism, and cardiovascular risks, which are long-term. Machine learning provides data-driven approaches for identifying patterns associated with PCOS from clinical and related patient characteristics. Such approaches may support predictive modelling by identifying informative features and combining them to distinguish between PCOS and non-PCOS cases. Feature selection is therefore used to identify a compact and stable subset of input variables, with the aim of improving predictive performance and model reliability. This study presents a hybrid optimisation approach for selecting a compact and stable subset of features for PCOS prediction. First, an ensemble wrapper-based approach is implemented, and then a combination of hybrid Optimisation strategies is examined with several combinations and ends with an ACOBO (Ant Colony Optimisation - Butterfly Optimisation) technique. Additionally, robustness of feature selection methods, a most common issue when algorithms run on the same dataset multiple times or with small modification has been checked with the help of rank-based Spearman and Kendall correlations, and distribution-based Jensen Shannon divergence using 10-fold cross-validation. Additionally, SMOTE (Synthetic Minority Oversampling Technique) has been used to address the class imbalance issue. Finally, multiple stacking configurations were evaluated with SMOTE by varying the meta-classifier while using the remaining candidate classifiers as base learners. Among the evaluated configurations, the stacking model with AdaBoost as the meta-classifier achieved the highest accuracy of 95.64%.