Machine Learning-Based Early Detection of Heart Disease Risk Using Self-Reported Non-Clinical Variables

Heart disease remains one of the leading causes of mortality worldwide, highlighting the need for accessible tools for early risk assessment. This study investigates the use of established machine learning (ML) techniques to predict heart disease risk from self-reported information that can be collected through telephone or online health questionnaires. The analysis is based on the publicly available Heart-2020 dataset, derived from the U.S. Centers for Disease Control and Prevention (CDC) Behavioral Risk Factor Surveillance System (BRFSS). Four widely used ML models were trained and evaluated: Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Bagged Trees (BT), and Extreme Gradient Boosting (XGBoost). Model performance was assessed using precision, recall, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC). To further improve predictive performance and robustness, an ensemble architecture based on a second-stage MLP was implemented to combine the outputs of the individual classifiers. In addition, several feature engineering techniques, including Pearson correlation analysis, Analysis of Variance (ANOVA), and Principal Component Analysis (PCA), were investigated. The proposed approaches achieved F1-scores exceeding 0.80, demonstrating strong predictive capability using only non-clinical, self-reported information. These findings suggest that simple questionnaire-based data can support automated early-warning systems capable of identifying individuals who may benefit from further medical evaluation. Beyond individual risk assessment, the proposed methodology could also facilitate large-scale population health monitoring, contributing to preventive healthcare strategies and informed public health policy development.