A HYBRID MACHINE LEARNING AND FEATURE OPTIMIZATION APPROACH FOR CARDIAC RISK PREDICTION

Abstract

Cardiovascular disease (CVD) remains one of the major causes of mortality globally, emphasizing the importance of early prediction and effective risk detection. Recent breakthroughs in Machine Learning (ML) techniques have been increasingly applied to medical datasets to assist healthcare practitioners in clinical decision-making. Several studies have developed predictive architectures utilizing classification algorithms, feature selection techniques, hybrid learning models, and comparative analyses of multiple algorithms to identify bottlenecks and improve prediction capability. Despite these advancements, prior research often depends on relatively small or spatially specific datasets, lacks effective feature optimization, exhibits increased computational complexity, and provides reduced interpretability for clinicians. These limitations emphasize the need for a proficient integrated framework that productively reduces feature redundancy, enhances model generalization, and improves prediction accuracy. The proposed framework aims to develop an interpretable and accurate CVD prediction system by combining feature optimization techniques with ML classification models. The developed architecture focuses on identifying the most prominent clinical attributes, reducing redundant and correlated variables, and utilizing optimized features to train robust predictive models for early CVD detection. Primarily, data preprocessing and normalization are performed on the dataset. Dimensionality is subsequently reduced by applying Principal Component Analysis (PCA), followed by the Improved Dragonfly Optimization Algorithm (IDFOA) to select the most relevant features. The optimized features are then utilized to train deep learning-based predictive models and Gradient Boosting Machine (GBM) classifiers for disease prediction. Experimental analysis illustrates that the integrated optimization approach significantly enhances CVD prediction performance compared with baseline models. The combination of PCA and IDFOA effectively reduces redundant features while improving sensitivity and the Gini coefficient. Random Forest (RF) achieves the best predictive performance. Overall, the designed framework provides a proficient and reliable method for CVD prediction and supports intelligent clinical decision-support systems for early CVD risk assessment and improved healthcare outcomes.