A Multi-Stage Hybrid Feature Selection and Robust Voting Ensemble Framework for High-Dimensional Cardiovascular Risk Stratification

Abstract

The global escalation of cardiovascular disorders demands highly accurate, objective, computerized diagnostic tools to mitigate manual clinical subjectivity. This study built a multi-layered hybrid computing system designed to handle high dimensionality and redundant features in intricate clinical profiles. Our approach deploys a cascaded three-tier feature selection engine comprising metaheuristic Particle Swarm Optimization (PSO), a Fast Correlation-Based Filter (FCBF), and a ReliefF algorithmic ranker. Random Forest (RF), Support Vector Machines (SVM), and Extreme Gradient Boosting (XGBoost) base learners are used in a consolidated majority-voting ensemble to compute the optimal attribute subset. When tested on a variety of patient clinical records in a Java-WEKA execution environment, the Random Forest engine produced an exceptional peak accuracy of 96.00% and a precision of 97.53%, whereas XGBoost showed better diagnostic sensitivity with a recall of 95.78%. The empirical findings demonstrate that sequential heuristic attribute pruning combined with parallel ensemble voting effectively mitigates learning model overfitting, rendering the architecture highly viable for deployment in real-time clinical decision support infrastructures.