Automated Classification of Electrocardiograms Using Wavelet Analysis and Deep Learning

Andrew Demonbreun and Grace Mirsky
Benedictine University


Abstract

The standard 12-lead electrocardiogram (ECG) is a non-invasive diagnostic tool for measuring and recording the electrical activity of the heart. The ECG is commonly used in the diagnosis of cardiac arrhythmias and abnormalities; however, the accurate interpretation of the ECG requires highly skilled practitioners. Therefore, automated diagnostic classification of ECGs can greatly assist clinicians, particularly when a shortage of such specialized personnel exists. In recent years, there has been increased interest in this research topic; however, these studies tend to be limited in the number of samples and/or diversity of the datasets. The 2020 PhysioNet/Computing in Cardiology Challenge facilitates the development of robust classification algorithms over a large, diverse dataset in order to overcome limitations of previous studies.

Our approach is to use wavelet analysis to develop multiple deep learning models, creating a unique model for each arrhythmia and lead. This approach leverages the ability of different leads, based upon their anatomical position, to better observe different arrhythmias. A voting scheme is implemented amongst the leads, allowing for confirmation of arrhythmia diagnosis from multiple leads. By having separate models for each arrhythmia type, this approach also allows for the classification of multiple coexisting arrhythmias. We use a modified version of GoogLeNet for training. Since GoogLeNet is designed for image classification, we convert the ECG signals to scalograms to leverage transfer learning. Scalograms are time-frequency representations of the absolute value of the continuous wavelet transform coefficients plotted over time and frequency. In the Unofficial Phase, we were unable to build models for all leads due to time limitations; our best performing entry received F_beta score = 0.310 and G_beta score = 0.170. Validation accuracy ranged from 83-98% for the different arrhythmias during training. In the official phase, we plan to build the complete set of models, which should improve our results.