Time adaptive ECG driven cardiovascular disease detector

Abstract

Electrocardiograms (ECGs) are widely used to detect cardiovascular disease (CVD). Deep learning is a topic of interest in healthcare, in which timely detection of ECG anomalies can play a vital role in patient monitoring. However, automatic detection of CVD via ECGs is a complex problem, where state-of-the-art performance is achieved typically by the rule-based diagnosis models, which are inefficient to deal with large amount of heterogeneous data and requires significant analysis and medical expertise to achieve adequate precision in CVD diagnosis.

In this paper, we propose a two-stage multiclass algorithm. The first stage performs ECG segmentation based on Convolutional Bidirectional Long Short-Term Memory neural networks with attention mechanism. A second stage is based on a time adaptive Convolutional Neural networks applied to ECG beats extracted from first stage for several time intervals. ECG beats are converted to 2D images using Short-Time Fourier Transform to automatically discriminate normal ECG from cardiac adverse events such as arrhythmia and congestive heart failure and predict sudden cardiac death. Model accuracy was compared across different time scales. Data used to train and test the models were extracted from MIT/BIH-PhysioNet databases.

By using 4 min ECG, we achieved an accuracy of 100% to automatically detect congestive heart failure events, 97.9% for arrhythmia events, and 100% to predict sudden cardiac deaths. This offers unprecedented results by supporting domain-experts work, by computing signal characteristics via an automated complete system for CVD diagnosis. The achieved results showed to be promising compared to state-of-the-art algorithms used for similar purposes.