Machine learning based classification of breast cancer using lifestyle and clinical risk factors in Puerto Rican women

Globally, breast cancer incidence has increased by more than 20 percent and mortality has increased by 14 percent making it a significant public health challenge. Effective prevention and screening strategies require improved early risk assessment tools, especially for understudied populations. Current risk models are predominantly trained on European descent populations, resulting in suboptimal and uneven performance in admixed populations. Our goal was to compare how current machine learning (ML) methods perform to classify breast cancer status in a Puerto Rican population. We used a structured dataset of lifestyle, demographic, reproductive, and established epidemiological risk-factor variables from a Puerto Rican cohort of 1393 woman. Data was preprocessed using systematic data cleaning, removal of outcome-leaking features, normalization of categorical labels, imputation of missing values, and encoding of non-numeric variables. The final dataset was split using a stratified train-test strategy to preserve case-control balance, followed by model development using multiple supervised algorithms. Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost models were trained and compared using stratified 10-fold cross-validation and standardized performance metrics. After hyperparameter tuning with randomized search, the CatBoost classifier achieved the strongest performance on the held-out test set, with a ROC-AUC of 0.8009, PR-AUC of 0.8285, F1-score of 0.7239, and balanced accuracy of 0.7342, outperforming both traditional models and other boosting algorithms. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) and permutation-based feature importance, which consistently identified influential predictors including menopausal status, smoking history, polygenic risk score, age, and number of sisters with breast cancer, among other risk factors. These findings confirm that gradient-boosting approaches, particularly CatBoost, effectively capture nonlinear interactions in multidimensional health data while maintaining interpretability. Overall, this work demonstrates the feasibility and utility of integrating machine learning to predict breast cancer risk using non-genetic and genetic risk factors in a Puerto Rican population.