Electrocardiographic Classification using Deep Learning with Lead Switching

The classification algorithms of rhythm and morphology abnormalities in electrocardiogram (ECG) signals have been widely studied. However, the existing study uses ECGs with fixed leads. We propose a neural network-based method to improve the ECG classification performance by switching the observing lead, where a single-lead ECG signal is observed at a time. The proposed model takes a sequence of ECGs with different leads as input, and outputs its abnormal probability. Using the trained model, an optimal lead order is determined based on the areas under the receiver-operating characteristic curve (AUC). The experimental results using 6,877 ECG recordings with nine diagnoses show that the proposed lead switching approach achieves significantly better AUCs than the fixed single-lead approaches, and comparable performance with 12-lead ECGs for classification of several diagnoses.