Autoencoder-Based Feature Learning for Predicting Cardiovascular Disease

Abstract

Cardiovascular disease is the leading cause of death in the world. Several studies have used machine learning methods to predict cardiovascular disease based on medical records. However, due to highly the correlation between data in medical records, much needs to be done in the field. Here, we propose to use Autoencoder based feature learning to predict cardiovascular. Autoencoder is a neural network trained to learn the data themselves by finding the underlying latent variables that produce them. This is done by designing the network with reduced number of nodes in the middle of the architecture (bottleneck). Thus, we expect autoencoder could learn the nonlinear and complex relationships between data. We varied the depth of autoencoder in this paper from 3 to 7 layers, and the depth of layer was varied to several neurons at the bottleneck. The output are later used as inputs to various classifiers: Logistic Regression, Naive Bayes, SVM, KNN, Decision Tree, XGBoost, Random Forest and Neural Network. Our experiments show that used of autoencoder-based features learning can improves the performance of the classifier. However, we notice that the depth of layer does not necessarily improves performance and needs to be determined empirically