Auto-Encoder based ECG Signal Feature Extraction for Real-Time Detection of Cardiac Arrhythmias

Abstract

In this paper, we proposed to use an auto-encoder for ECG signal feature extraction and a QRS complex shape prediction error as an estimate of signal shape changes, related to arrhythmia. In order to estimate the efficiency of auto-encoder based feature extraction, we performed an experimental investigation using Ventricular Tachycardia (VT) indicated incidents’ records from Physionet database. The experimental investigation has shown, that the application of auto-encoder makes possible to detect VT incidents in signals, where the features based on signal frame averaging fails. An additional investigation is needed in order to create a robust algorithm for decision threshold selection, to indicate the MSE value changes, at which the VT incident related alarm should be activated.