An Improved Random Forest Model for Detecting Heart Disease
Avijit Kumar Chaudhuri, Sulekha Das, Arkadip Ray
Abstract
Diagnosing cardiovascular disease (CVD) is a crucial issue in healthcare and research on machine learning. Machine-learning techniques can predict risk at an early stage of CVD based on the features of regular lifestyles and results of a few medical tests. The Framingham Heart Study dataset has 15.2% of patients with CVD, which increases the likelihood of classifying CVD patients as healthy. We create approximately equal instances of each class by over-sampling. We evaluate: (i) no over-sampling, (ii) random over-sampling of the training dataset, and (iii) over-sampling before splitting the dataset. We apply 50–50%, 66–34%, and 80–20% train-test splits and 10-fold cross-validation. We compare logistic regression (LR), Naive-Bayes (NB), support vector machine (SVM), decision tree (DT), and random forest (RF) classifiers. The comparison based on accuracy, sensitivity, specificity, area under the curve (AUC), and Kappa statistics highlights the importance of the AUC score. The best performers in Case (i) are NB and DT; in Case (ii) are LR, SVM, and NB (due to the reduction in class overlap); and in Case (iii) are DT and RF (due to incorrect over-sampling). RF, with more scope for improvement, is used as a meta-classifier with LR, SVM, and NB as the base classifiers in a stacking-ensemble technique. It records a 20(10)% improvement in AUC score over RF and DT (NB) in Case (ii). We can match the reported predictive performances on the same dataset, without using feature engineering and an optimal value of hyper-parameters, in Case (ii). We could exceed it only in Case (iii).