Background: Cardiovascular diseases represent an underlying cause of death worldwide as they are responsible for about one million deaths annually in the United States alone. Hence, contributing to the identification of different arrhythmias would help cardiologists better diagnose patients with cardiovascular diseases. For this reason, the current work aims to classify multichannel ECG signals into 9 heart rhythms using deep learning.
Materials and Methods: The study dataset includes a cohort of 6877 12-lead ECG recordings with various lengths. Two options were tested to standardize the recordings length: zero padding and signal repetition. We design a deep convolutional neural network (CNN) inspired by VGG16 architecture regularized with dropout. Several options were tested like adding residual layers and customizing classification thresholds. By benchmarking the classification performance of several architectures, we opt for a deep CNN for multilabel classification with binary cross entropy loss. Class weights were configured to ensure balance. Data preprocessing includes zero padding, considering only the first 18000 samples of the padded recordings, then applying first-order high-pass Butterworth filter.
Results: The training scores F2-score, G2-score and geometric mean (GM) after 5-fold cross validation are respectively 0.77, 0.58, 0.67 with zero padding and 0.74, 0.54, 0.63 with signal repetition. The scores’ standard deviations are below 0.02. By analyzing the confusion matrix, we notice that ST-segment elevation (STE) class has poor sensitivity (0.43). It can be explained by the low number of samples (220) compared to atrial fibrillation, for instance, that has 1221 samples and high sensitivity (0.93). Results on the test set are F2-score=0.77, G2-score=0.55 and GM=0.65.
Conclusions: We designed a deep learning solution to help cardiologists distinguish 8 cardiac arrhythmias and sinus rhythm. Further work should aim at enhancing the classifier’s performance while handling the highly imbalanced dataset ratio between different classes and the various length recordings.