Comparative analysis and feature importance of machine learning and deep learning for heart disease prediction

Abstract

Cardiovascular disease (CVD) or heart disease is one of the main reasons for early death, even at young age and that too often sudden. If it is detected more accurately, much before it seriously affects the individual, life can be saved through proper medication and changes in lifestyles. In this work different machine learning classifiers and a deep learning algorithm multi-layer perceptron (MLP) were applied on two different datasets, Framingham heart study dataset and UCI heart disease dataset for prediction of heart disease. These algorithms were optimized using hyperparameter tuning and compared for their performance measures and prediction accuracies. For different features, feature importance scores were calculated using machine learning algorithms. The features were ranked according to their scores. Out of various classification algorithms, random forest algorithm has shown the best results with prediction accuracy of 97.13% for Framingham dataset. MLP has shown good performance for both the datasets.

Evaluation of feature scaling for improving the performance of supervised learning methods

“…Despite the wider application of various supervised learning methods such as SVM [9], [10], deep learning [11], RF, decision tree (DT) [12], MLP [13], and LR [14] for HF prediction, the effectiveness of these HF prediction methods have scope for improvement and requires much research effort. The literature review shows that most of the supervised methods are developed on the original dataset and the significance of pre-processing such as feature scaling, and dimensionality reduction for the linear model such as SVM is widely ignored [15]- [20].…”