Machine Learning Approach for Cardiovascular Disease Risk Stratification Using Ensemble Classifiers and Dimensionality Reduction

Cardiovascular disease (CVD) remains a leading cause of morbidity and mortality worldwide, motivating data-driven tools that complement established clinical risk criteria. We present a two-pathway machine learning framework for CVD risk assessment. Pathway I is a brief rule-learnability verification, a four-level risk label is generated by a deterministic MEWS/Framingham-inspired scoring rule, and because the label is a deterministic function of the inputs, its near-perfect recovery by the classifiers is only a consistency check, not predictive evidence. Pathway II, the focus of the study and the main contribution of this work, is a hybrid pipeline that combines unsupervised preprocessing (K-Nearest Neighbors imputation, PCA/UMAP dimensionality reduction, K-Means phenotyping) with supervised ensemble classifiers (Random Forest, XGBoost, LightGBM), trained on an ≈70,000-record Kaggle cohort and externally validated on an independent 219-patient cohort using the same prevalent-CVD definition. On the external cohort the six pipelines performed comparably (accuracy 70–74%, external-cohort AUC 0.82–0.85); paired McNemar tests showed no significant difference between UMAP- and PCA-based pipelines (all p>0.5), and a power analysis indicates the cohort is too small to distinguish them, so no configuration is claimed to be superior. An unsupervised phenotype analysis is included as a descriptive, hypothesis-generating step; its association with the outcome is statistically significant but of negligible effect size. The framework offers a reproducible, clinically interpretable methodology for CVD risk categorization.