Comparison of Machine Learning Detection of Low Left Ventricular Ejection Fraction Using Individual ECG Leads

Abstract

The 12-lead electrocardiogram (ECG) is the most common front-line diagnosis tool for assessing cardiovascular health, yet traditional ECG analysis cannot detect many diseases. Machine learning (ML) techniques have emerged as a powerful set of techniques for producing automated and robust ECG analysis tools that can often predict diseases and conditions not detectable by traditional ECG analysis. Many contemporary ECG-ML studies have focused on utilizing the full 12-lead ECG; however, with the increased availability of single-lead ECG data from wearable devices, there is a clear motivation to explore the development of single-lead ECG-ML techniques. In this study we developed and applied a deep learning architecture for the detection of low left ventricular ejection fraction (LVEF), and compared the performance of this architecture when it was trained with individual leads of the 12-lead ECG to the performance when trained using the entire 12-lead ECG. We observed that single-lead-trained networks performed similarly to the full 12-lead-trained network. We also noted patterns of agreement and disagreement between network low LVEF predictions across the different lead-trained networks.

Figure 1.

ECG low LVEF detection architecture. This network consists of an input stage, temporal and spatial residual blocks, and an output stage. Each residual block consists of four layers of residual blocks, similar to the common resnet structure. In cases where the number of input channels is less than the output channels (layers 1 through 3), the input is re-sampled using a 1×1 convolutional layer. The spatial residual block uses 7×1 convolutional filters whereas the temporal uses 1×3 filters. The features from the two residual blocks are concatenated before the output stage. When single-leads are used, the spatial blocks instead use 1×1 convolutional filters.

Figure 2.

Average correlation of network outputs between each lead scenario. Each correlation in the larger heat map is an average of the correlations between the 5 instances of each lead-specific network. The identify correlations (along the diagonal for the same lead to same lead comparisons) were excluded from the average. The inset shows the individual per iteration comparisons for lead V4 compared to lead V4.