A Mixed-Domain Self-Attention Network for Multilabel Cardiac Irregularity Classification Using Reduced-Lead Electrocardiogram

TL;DR

This paper introduces a novel mixed-domain self-attention ResNet model designed to classify cardiac irregularities from reduced-lead ECGs, demonstrating its effectiveness across multiple recording sources in a challenge setting.

Contribution

The study presents a new mixed-domain self-attention network architecture that improves classification of cardiac irregularities using fewer ECG leads, addressing generalizability issues.

Findings

Achieved competitive scores across different reduced-lead ECG configurations.

Demonstrated the model's ability to generalize across multiple data sources.

Ranked within top 40 in the PhysioNet Challenge 2021.

Abstract

Electrocardiogram(ECG) is commonly used to detect cardiac irregularities such as atrial fibrillation, bradycardia, and other irregular complexes. While previous studies have achieved great accomplishment classifying these irregularities with standard 12-lead ECGs, there existed limited evidence demonstrating the utility of reduced-lead ECGs in capturing a wide-range of diagnostic information. In addition, classification model's generalizability across multiple recording sources also remained uncovered. As part of the PhysioNet Computing in Cardiology Challenge 2021, our team HaoWan AIeC, proposed Mixed-Domain Self-Attention Resnet (MDARsn) to identify cardiac abnormalities from reduced-lead ECG. Our classifiers received scores of 0.602, 0.593, 0.597, 0.591, and 0.589 (ranked 54th, 37th, 38th, 38th, and 39th) for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set with the evaluation metric defined by the challenge.