Machine Learning Based Early Prediction of Polycystic Ovary Syndrome (PCOS) Using Lifestyle and Symptom Features

The diagnosis of the Polycystic Ovary Syndrome (PCOS) is usually made on fewer than three million women. Symptoms and impacts of this syndrome are experienced by about $\mathbf{8}-\mathbf{1 3}$ percent of the world women in their ways of reproduction: its positive identification is still hard due to the invasive tests as well as clinical examination of hormones. The paper will suggest a new machine learning algorithm to predict PCOS early when only non-invasive lifestyle and symptom characteristics low- end diagnostic tools are available to fill this incredibly significant gap in the available diagnostic methods with this underrepresented group. All the data was collected to form a comprehensive dataset that consisted of 541 records, and even the nine most important variables of age, weight, BMI, menstrual irregularity, hirsutism, darkening of skin, acne, sleep schedule and fast-food consumption, exercise, and stress were included. The comparison and implementation of five algorithms of machine learning included the following algorithms: Logistic Regression, Random Forest, XGBoost, LightGBM, and Support Vector Machine. Out of these, XGBoost model showed the best results with an accuracy of 94.5 %, precision 93.2 %, recall 95.8 % and F1-score 94.5 with an AUC-ROC value of 0.967, strongly exceeding any previously analyzed ones on the basis of invasive clinical tests. The most predictive features in SHAP analyses were menstrual irregularity, BMI and hirsutism; they explained 86 and two-thirds percent of the overall predictive power. This is an unobtrusive screening method that suggests the potential of an early diagnosis whereby health providers can counsel them early enough in resource-constrained environments and potentially benefit a couple of million women across the world with regards to reproductive health.