Heart disease prediction by tabular modeling with deep learning network and interpretability

Abstract Cardiovascular disease remains one of the most prevalent causes of mortality globally, underscoring the critical necessity for precise predictive models for early detection and intervention. Despite the proven potential of machine learning (ML), including deep learning (DL) models and convolutional neural networks (CNNs), in healthcare, their application to structured tabular datasets for heart disease prediction remains limited. Existing models often lack interpretability, which hinders their application in clinical settings, where understanding the decision-making process is vital. In response to this, our study proposes an innovative CNN-based predictive model for heart disease utilizing tabular modeling techniques and incorporating an interpretability tool such as SHapley Additive exPlanations (SHAP). The aim was to enhance both predictive accuracy and model transparency, empowering clinicians to comprehend and trust the model’s decisions. The methodology included data preprocessing, designing a CNN architecture tailored for tabular data, and integrating SHAP. The results showed superior predictive performance compared with the baseline models, with 98.54% accuracy, 97.14% sensitivity, 100% specificity, and SHAP, providing valuable insights into feature importance. This research advances heart disease prediction by harnessing the adaptability of CNNs to structured tabular datasets, while addressing the critical need for model interpretability in healthcare applications.