Learning ECG Representations for Multi-Label Classification of Cardiac Abnormalities

Jangwon Suh1, Jimyeong Kim1, Eunjung Lee1, Jaeill Kim1, Duhun Hwang1, Jungwon Park1, Junghoon Lee1, Jaeseung Park1, Seo-Yoon Moon1, Yeonsu Kim1, Min Kang1, Soonil Kwon2, Eue-Keun Choi3, Wonjong Rhee1
1Seoul National University, 2Seoul National University Hospital, 3Seoul National University College of Medicine


The goal of PhysioNet/Computing in Cardiology Challenge 2021 was to identify clinical diagnoses from 12-lead and reduced-lead ECG recordings, including 6-lead, 4-lead, 3-lead, and 2-lead recordings. Our team, snu_adsl, have used EfficientNet-B3 as the base deep learning model and have investigated methods including data augmentation, self-supervised learning as pre-training, label masking that deals with multiple data sources, threshold optimization, and feature extraction. Self-supervised learning showed promising results when the size of labeled dataset was limited, but the competition's dataset turned out to be large enough that the actual gain was marginal. In consequence, we did not include self-supervised pre-training in our final entry. Our classifiers received scores of 0.626, 0.610, 0.612, 0.611, and 0.610 for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set with the Challenge evaluation metric.