Hybrid ensemble approaches for cardiovascular disease prediction: Leveraging interpretable AI for clinical insight

Cardiovascular diseases are one of the most common and serious conditions that affect people's health. The use of early-stage signs to diagnose cardiovascular disease is a major problem in the modern world. However, there is a need for less expensive tools that detect cardiovascular disease without overbearing procedures. In this study, we develop ensemble hybrid and blending techniques for the early detection of cardiovascular disease. Firstly, the proximity weighted random affine shadow sampling is utilized to reduce risk of bias model learning. Next, principal component analysis extracts most relevant features to improve accuracy and computational time. After that, we developed two novel ensemble models: hybrid and blending. A hybrid model combines convolutional neural networks and temporal convolutional networks, while blending model places convolutional neural networks and temporal convolutional networks, and a deep belief network on base layer and highway networks on meta layer to enhance cardiovascular disease detection performance. A combination of 10-fold cross-validation on clinical cardiovascular data showed both ensemble methods surpassing conventional models achieving peak results through the blending approach. The blending model reaches 0.914 accuracy and 0.9097 F1-score yet the hybrid model shows 0.8496 accuracy with 0.8496 F1-score. The project utilizes Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP) through explainable artificial intelligence (XAI) techniques for achieving dependability and transparency. The methods deliver explanations to clinicians by showing how clinical characteristics influence the predicted results. The proposed ensemble model demonstrates strong predictive capabilities combined with interpretability which allows its use in real-time operations for cardiovascular disease prediction. • Hybrid CNN-TCN + DBN-HN ensemble model improve cardiovascular disease prediction. • ProWRAS sampling and PCA enhance class balance and computational efficiency. • Blending approach yields 91.4 % accuracy and ROC-AUC of 0.967. • LIME and SHAP integration boosts model interpretability for clinicians. • Outperforms traditional risk scores like Framingham and ASCVD in reliability.