Phenotype-Specific Lifestyle Prediction for PCOS Using Machine Learning Multi-Class Classification and SHAP Explainability

Polycystic Ovary Syndrome (PCOS) is the most prevalent form of endocrine disorder that affects globally 5–13%

of women of reproductive age; nevertheless, its heterogenous phenotypic expressions are often overlooked in the

clinic. Thus, the regular binary diagnostic approach fails to identify some subtypes (HA+OD-, HA-OD+,

HA+OD+) who should be managed differently in terms of lifestyle and pharmacological approaches. This work

presents a multi-class phenotype labeling framework based on the Rotterdam criteria and describes a machine

learning pipeline that employs five classifiers to predict four PCOS phenotypes using non-invasive lifestyle and

symptom data. Applying Support Vector Machines (SVM), Extreme Gradient Boosting (XGBoost), Random

Forest (RF), Logistic Regression (LR) and K-Nearest Neighbours (KNN) via five-fold stratified cross-validation

to the IEEE DataPort PCOS dataset, which includes 15 features and 267 patients. All 5 models demonstrate

impressive discrimination (accuracy ≥ 98%); XGBoost and Random Forest separate perfectly. SHAP (Shapley

Additive explanations) analysis suggests cycle_length is the most impactful predictor for all model types,

reinforced by logistic regression coefficient plot and Random Forest impurity scores. These findings validate the

clinical relevance of cycle regularity as a biomarker for PCOS, and highlight that explainable machine learning

can enable large-scale phenotype-specific lifestyle recommendations.