Leveraging Machine Learning to Improve PCOS Detection through Ovarian Analysis

Abstract: A complex endocrine condition, PCOS affects up to 10% of women who are of reproductive age. Infertility and metabolic problems can be avoided with early diagnosis. For PCOS prediction, this study suggests a machine learning (ML)-based diagnostic approach that makes use of clinical, biochemical, and lifestyle variables. Using the Kaggle PCOS dataset, which included 541 samples with 42 characteristics, models including Random Forest, Support Vector Machine (SVM), and Neural Networks were trained. Following 10-fold cross-validation and Randomized Search hyperparameter adjustment, the Random Forest classifier produced an accuracy of 87.73%, a macro F1-score of 0.88, and an AUC-ROC of 0.93. Weight gain, hair growth, and follicle count were found to be the most significant predictors by SHAP analysis. The clinical promise of ML-based screening for early, non-invasive, and interpretable PCOS identification is demonstrated by these findings.