A Machine Learning Approach for Non-Invasive PCOS Diagnosis from Ultrasound and Clinical Features

This study explores the use of machine learning algorithms for diagnosing polycystic ovary syndrome (PCOS) non-invasively, focusing on predictive performance and clinical applicability. Among the evaluated algorithms, XGBoost outperformed others and was selected for final development. The study structured data into clinical, biochemical, and ultrasound categories, identifying the top 10 predictive features using SelectKBest and validating them through XGBoost's feature importance and expert assessment. The XGBoost model showed robust performance with high accuracy and precision across different feature combinations, with the most influential features including follicle count, weight gain, AMH, hair growth, menstrual irregularity, and hair loss. External validation using a dataset of 320 instances confirmed the model's perfect performance, though further validation is needed to ensure no overfitting. The findings suggest that combining clinical and ultrasound features allows for accurate, non-invasive, and cost-effective PCOS diagnosis, highlighting the potential of ML tools to enhance clinical workflows and support early intervention in women's health.