Back to Projects Projects / Healthcare AI
Coronary Heart Disease Risk Predictor Preview
Healthcare AI 2025 · Clinical ML

Coronary Heart Disease Risk Predictor

A hybrid ensemble predictive modeling pipeline engineered to estimate 10-year risk of coronary heart disease with automated hyperparameter tuning and cross-validation.

Project Overview

Coronary Heart Disease (CHD) remains one of the world's leading causes of mortality. Early identification of vulnerable patients can significantly alter clinical trajectories through proactive lifestyle interventions and targeted therapies.

This system implements an end-to-end clinical decision support pipeline. Leveraging patient demographic markers, clinical biomarkers (systolic/diastolic blood pressure, glucose, total cholesterol), and behavioral history from longitudinal cardiovascular datasets, the model calculates personalized 10-year risk probabilities with high sensitivity and calibration.

Key Features

Hybrid Ensemble Architecture

Combines Gradient Boosting, Random Forests, and penalized Logistic Regression via soft-voting to optimize ROC-AUC and prevent single-estimator bias.

Automated Tuning & Validation

Utilizes stratified k-fold cross-validation paired with Bayesian hyperparameter optimization to reliably calibrate predicted risk probabilities.

Robust Imputation Pipeline

Handles missing clinical values through iterative multivariate feature imputation, safeguarding patient sample distributions without data loss.

Risk Stratification Output

Translates continuous model confidence scores into standardized low, intermediate, and high risk strata aligned with clinical practice guidelines.

Technical Challenges Faced

Severe Class Imbalance: The incidence of positive 10-year cardiac events in screening cohorts was under 15%. Applied SMOTE-Tomek synthetic oversampling in conjunction with class-weighted loss functions to elevate recall without inflating false-alarm rates.
Multicollinearity Among Hemodynamic Features: Systolic and diastolic blood pressure exhibited tight collinearity. Evaluated Variance Inflation Factor (VIF) scores and engineered composite pulse pressure indicators to stabilize regression weights.
Model Calibration: Raw ensemble margins tended towards overconfidence. Implemented isotonic regression probability calibration, ensuring predicted percentages accurately reflect empirical cardiovascular event rates.

Improvements & Future Roadmap

SHAP & LIME Feature Attribution: Integrating local model explanation dashboards so clinicians can review which patient metrics contributed most significantly to risk scores.
Interactive Web Diagnostic UI: Packaging the pipeline into a lightweight web dashboard allowing medical practitioners to input patient records and receive immediate prognostic reports.