FGSCare: A Feature-driven Grid Search-based Machine Learning Framework for Coronary Heart Disease Prediction

Although machine learning has become a powerful tool for coronary heart disease (CHD) prediction, its effectiveness is often hindered by the complexity and nonlinear interactions among medical risk factors. A major challenge lies in feature selection, where the absence of systematic strategies may lead to information loss, overfitting, or the inclusion of irrelevant variables, ultimately degrading predictive performance. Additionally, different ML models exhibit varying predictive capabilities depending on the selected features. However, many studies fail to systematically evaluate how feature selection influences the performance of traditional and deep learning approaches, limiting the understanding of optimal feature selection strategies and their impact on improving CHD prediction. To address these limitations, we propose a Feature-driven Grid Search-based Machine Learning Framework (FGSCare) for CHD prediction. FGSCare systematically filters and retains critical features through a data-driven selection process, enhancing model interpretability and generalization. We assess the impact of feature selection by comparing the performance of traditional machine learning classifiers. We assess the impact of feature selection by comparing the performance of traditional machine learning classifiers (e.g., k-Nearest Neighbors, ElasticNet, and Decision Tree) and deep learning models such as Transformers before and after feature selection. Experimental results demonstrate that feature selection significantly influences model performance across various evaluation metrics, including accuracy, precision, recall, F1-score, and AUC. Our findings provide valuable insights into the trade-offs between traditional ML and deep learning methods in CHD prediction, contributing to the development of more robust, data-driven healthcare applications. Our implementation is available at: https://github.com/zl3508/Heart.