Classification of 12-Lead Electrocardiograms Using Residual Neural Networks and Transfer Learning

Abstract

This article concerns the PhysioNet/Computing in Cardiology Challenge 2020 which focused on building computational methods to identify cardiac abnormalities from 12-lead ECGs. Our team, MCIRCC, utilized a large secondary dataset of 12-lead ECGs obtained from the Section of Electrophysiology at the University of Michigan, called the MUSE dataset, to pre-train multiple residual neural networks that were later retrained on the challenge dataset. To do so, the diagnosis statements that existed in our dataset were utilized to assign the same labels to our ECGs as the challenge data. After parameter optimization, we selected a subset of top performing models and created an ensemble model that achieved a challenge validation score of 0.616, and full test score of 0.141, placing us 27 th out of 41 teams in the official ranking.

Robustness of Residual Network in Predicting PR Interval Trained using Noisy Labels

“…The arrhythmia labels were generated using a combination of three approaches. The majority of semi-structured diagnosis statements, i.e fragment of text that contains a single or multiple diagnoses [2], embedded in the dataset were assigned to a Unified Medical Language System Concept Unique Identifiers (CUI) [3] and the corresponding Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) [2]. All the segments of diagnosis statements that did not have a corresponding CUI were then split into two groups sorted by frequency.…”

“…Building on the previous work of Ansari et al [2], a ResNet model was used to estimate the PR interval. The raw 12-lead ECG signals were fed to a single 1Dconvolutional layer with 64 filters and a kernel size of 15, followed by 9 residual blocks.…”