A 12-lead ECG correlation network model exploring the inter-lead relationships

Abstract

The 12-lead electrocardiogram (ECG) is widely used for automatic diagnosis of arrhythmia based on deep neural networks (DNN). In this paper, we use the 12-lead ECG dataset provided by the China Physiological Signal Challenge 2018 (CPSC2018), which contains 6877 samples and each sample contains 12-lead ECG records and corresponding reference labels. First, by statistical analysis of the results of 90 DNN models with F 1 > 0.6 published by CPSC2018, we found that almost all DNN models had high accuracy in identifying the left bundle branch block (LBBB) even when the number of training samples for LBBB is severely insufficient. Second, through ablation studies, we found that the absence of the 7th lead V1 severely affected the diagnostic accuracy of many DNN models, where ablation studies were used to quantify the effect of the disappearance of a single lead on the F 1 of the DNN model. We aim to explain the above two special phenomena using complex network theory. A 12-lead ECG correlation network based on the inter-lead Pearson correlation coefficient is proposed, which allows us to observe the correlation between a single lead and others, and quantify the correlation strength of each lead through a projection process. We used the covariance method to quantify the consistency of the change trend for the average correlation strength of 12 leads between any two categories, and found that the mean values of the covariance for LBBB under the positive and negative 12-lead ECG correlation network were 0.01 and 0.07, respectively, much smaller than other categories. This uniqueness may explain from the perspective of complex networks why LBBB can be diagnosed accurately by almost all DNN models when its number of samples used for training is severely insufficient. Furthermore, we found that the correlation between the lead V1 and other leads was close to 0. This low correlation may make the information of the lead V1 significantly different from other leads, resulting in its important role in the automatic diagnosis of arrhythmia.