Auto-Encoder Based Nonlinear Dimensionality Reduction of ECG data and Classification of Cardiac Arrhythmia Groups Using Deep Neural Network

Abstract

In this paper, we propose a neural network based dimensionality reduction approach to classify the four different categories of cardiac arrhythmias such as 'Normal', 'Bradycardia', 'Tachycardia' and 'Block' using MIT-BIH Arrhythmia database. We designed a nonlinear auto-encoder with three hidden layers and applied it to a dataset of ECG signal. It was found that our method is able to compress the data with less reconstruction error than that of linear transformation such as Principal component analysis. Using a deep neural network, we get the best accuracy (92.1%) for classifying the cardiac arrhythmia groups into four categories compared to Ensemble of Binary Support Vector Machine Decision Tree and Multilayer Perceptron (MLP) feed-forward Neural Network with back-propagation. The classifier performance is evaluated using Positive predictive Value (Precision), False discovery Rate, True Positive Rate (Recall) and False Negative Rate. This method has a great importance for researcher to predict the potential cardiac disease before it is too late.