Automated Comprehensive Interpretation of 12-lead Electrocardiograms Using Pre-trained Exponentially Dilated Causal Convolutional Neural Networks

Abstract

Correct interpretation of the electrocardiogram (ECG) is critical for the diagnosis of many cardiac diseases, and current computerized algorithms are not accurate enough to provide automated comprehensive interpretation of the ECG. This study aimed to develop and validate the use of a pre-trained exponentially dilated causal convolutional neural network for interpretation of the ECG as part of the 2020 Physionet/Computing in Cardiology Challenge. The network was pre-trained on a physician-annotated dataset of 254,044 12-lead ECGs. The weights of the pre-trained network were partially frozen, and the others were finetuned on the challenge dataset of 42,511 ECGs. 10-fold cross-validation was applied and the best performing model in each fold was selected and used to construct an ensemble. The proposed method yielded a cross-validated area under the receiver operating curve (AU-ROC) of 0.939 ± 0.004 and a challenge score of 0.565 ± 0.005. Evaluation on the hidden test set resulted in a score of 0.417, placing us 7th out of 41 in the official ranking (team name UMCUVA). We demonstrated that an ensemble of exponentially dilated causal convolutional networks and pre-training on a large dataset of ECGs from a different country and device manufacturer performs excellent for interpretation of ECGs.