Segment, Perceive and Classify - Multitask Learning of the Electrocardiogram in a Single Neural Network

Abstract

As part of the Physionet 2021 Challenge, “Will Two Do? Varying Dimensions in Electrocardiography: The PhysioNet/Computing in Cardiology Challenge 2021”, we have developed a neural network to classify pathologies and changes in the ECG. Our team HeartlyAI has developed a novel multitask learning based network that combines classification with segmentation and extrasystole detection. To obtain segmentation annotations, we developed an annotation tool in Angular and have manually annotated 1,789 ECGs from all challenge data sources for a gold standard of P wave, QRS, and T wave segments. Each extrasystole was annotated as supraventricular or ventricular. In the first step of our classification workflow, the ECG is segmented using a U-Net. This segmentation is used to calculate within-net features such as the PQ, QTc time, and Q-Q interval. The bottleneck layer of the U-Net is used along with the computed features as an embedding for the classification. We have used the recent Perceiver architecture to perform the classification of the ECG. Our classifiers received scores of 0.40, 0.31, 0.34, 0.34, and 0.25 (ranked 18th, 24th, 23rd, 23rd, and27th) for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set with the Challenge evaluation metric.