Risk Stage Embedding: Unsupervised Representation Learning for Enhanced Heart Disease Prediction

Heart disease prediction is a critical task in clinical decision support. Existing methods typically rely on raw clinical features, overlooking the synergistic risk patterns among patient subgroups. We propose Risk Stage Embedding (RSE), a novel unsupervised representation learning approach that enhances predictive performance by capturing latent patient risk phenotypes. Specifically, RSE first clusters patients into risk stages using K-Means on clinical features, then generates low-dimensional stage embeddings via PCA, and finally fuses these embeddings with original features for prediction. Evaluated on the standard Heart Disease Dataset using 5-fold stratified cross-validation, RSE consistently improves AUC, precision, and recall across four mainstream models—Logistic Regression, SVM, Random Forest, and XGBoost—with XGBoost achieving an AUC of 0.921 (an absolute gain of + 0.018). Elbow method and silhouette analysis confirm that the optimal number of risk stages is K = 2, corresponding to clinically meaningful “low-risk” and “high-risk” patient groups. Our method requires no labels during representation learning, is model-agnostic, and provides interpretable risk stratification. This work demonstrates that unsupervised discovery of risk stages can significantly boost predictive accuracy in tabular medical data.