Impact of Feature Selection Technique to Classify Early Polycystic Ovary Syndrome

Abstract

Polycystic ovarian syndrome (PCOS) is a multifaceted endocrine condition prevalent among women of reproductive age. Recent studies suggest that 1 in 10 women is diagnosed with PCOS. In this study, an openly usable dataset was collected from the Kaggle repository and preprocessed to use ML techniques. 7 different ML models, including 3 customized models, were used to predict PCOS. The importance of feature selection techniques is shown in this study. We have used two very famous feature selection methods to show how effective they are. We conducted the analysis in three different orders. The outcome indicates that the Adaptive boosting algorithm showed the best accuracy result of 88.99% without any feature selection. After performing the LASSO regression feature selection method, the Adaptive boosting model's performance jumped, and it got a 90.83% accuracy score. However, all the ML models' performance was boosted after using the top features of the Pearson correlation coefficient (PCC) feature selection method. The Logistic regression model attained the highest $\mathbf{9 5. 4 1} \boldsymbol{\%}$ accuracy and $\mathbf{0. 9 2}$ AUC-ROC score with the PCC method. The comparison stated that the ML models with the PCC feature selection technique performed best in detecting early PCOS disorder. We hope this study will help to build a better computer-aided predictive model to predict early PCOS.