Heart Disease Prediction Using an Ensemble Machine Learning Approach

Abstract

The number of cardiovascular diseases (CVDs) is the cause of death in most parts of the world, with around 32 percent of all deaths around the world. Prediction of CVDs is essential and needs to be done at the earliest stage and with accuracy to succeed in intervening at the earliest stage and to help patients. The study under consideration provides a detailed comparative study of machine learning and deep learning methods to predict heart disease on the basis of the Behavioral Risk Factor Surveillance System (BRFSS) 2015 sample that encompasses 253,680 patient records that include 22 health indicators. In order to mitigate the class imbalance of the medical datasets, which are unavoidable, we use Adaptive Synthetic (ADASYN) oversampling methodology. We compare fifteen different models, such as Convolutional Neural Networks (LeNet), Recurrent Neural Network (GRU, LSTM, BiLSTM), hybrid systems (CNN+LSTM, EnsCVDDNet, BICVDD-Net), and classical machine learning algorithms (AdaBoost, KNN, SVM, XGBoost, Naive Bayes, Logistic Regression, Decision Tree). Results of the experiment show that Voting Classifier, an ensemble technique, which combines several algorithms, is more effective with high accuracy (91.7%), precision (92.0), recall (91.7), and F1-score (91.8). Our results indicate that ensemble methods are very effective in improving predictive accuracy on structured healthcare data, which provides a powerful framework of clinical decision support systems