A Deep Learning Approach for Prediction of Heart Disease using Ensemble Learning Techniques

Abstract

Heart Disease (HD) stands as a significant contributor to escalating mortality rates worldwide, impacting communities on a vast scale. While numerous Deep Learning (DL) methodologies have been proposed for HD detection and diagnosis, prior studies have encountered limitations, notably small dataset sizes leading to overfitting. To address these challenges, we present novel ensemble learning approaches aimed at enhancing the authenticity and accuracy of HD prediction. Our work introduces two ensemble techniques, blending and voting, leveraging a combination of Multi-Layer Perceptron, AlexNet, and Deep Belief Network models. Utilizing the heart-disease-health-indicators-BRFSS2015 dataset, characterized by significant imbalances, we employ Adaptive Synthetic oversampling technique to balance the data. Results demonstrate the superior performance of our ensemble models over individual models. The blending model achieves an accuracy of 90.0%, precision of 90.1%, recall of 90.0%, and an F1-score of 90.0%. The voting model attains an accuracy score of 89.3%, precision of 92.2%, recall of 85.9%, and an F1-score of 88.9%. The blending model exhibits superior classification ability compared to the voting model. To assess the generalizability and validate results, K-Fold Cross Validation is employed. For model interpretability and explainability, we utilize an eXplainable Artificial Intelligence technique, Shapley Additive Explanations.