Cardiac Disease Prediction Using Artificial Intelligence and Deep Learning: A Comprehensive Multi-Model Clinical Evaluation with Recall Optimization

Abstract

Cardiovascular diseases (CVDs) are still the leading cause of death in the world and are responsible for killing about 17.9 million people every year. Identification of high-risk individuals is a key factor in lowering the death rate, but this is difficult because cardiac risk is multifactorial and nonlinear. Using a dataset of 10,585 patient records with 26 clinical variables, this paper presents a systematic five-phase experimental evaluation of diverse machine learning (ML) and deep learning (DL) models and configurations for cardiac disease prediction. The work documents the incremental contribution of each methodological step, starting from baseline classifiers and then moving to ensemble methods, neural networks, automated hyperparameter tuning, probability calibration and a clinically motivated recall- . Four key contributions are combined: (i) three engineered composite features based on the knowledge of the cardiovascular domain and validated using mutual information analysis; (ii) a four-class-imbalance correction strategy, ranging from standard SMOTE to an aggressive approach based on 3× oversampling and asymmetric cost weighting; (iii) a validated three-class patient risk stratification where the high-risk class shows a 94.2% observed cardiac event rate; and (iv) a threshold-adjusted Random Forest with a 99.21% sensitivity with only 6 of 755 confirmed cardiac events not detected. How data preparation, particularly clinical feature engineering and class-imbalance correction, affects predictive performance is shown, highlighting the importance of these steps over the choice of model architecture; and it illustrates a reproducible workflow from generic ML classifiers to a clinically g rounded, patient-safety regulated screening tool.