Impact of Neural Architecture Design on Cardiac Abnormality Classification Using 12-lead ECG Signals

Abstract

). In this paper, as part of the PhysioNet Challenge 2020, we propose two cardiac abnormality detection and classification neural models, using 12-lead ECG signals. Our ECU team proposes a hand-designed Recurrent Convolutional Neural Network (RCNN), consisting of 49 one-dimensional convolutional layers, 16 skip connections, and one Bi-Directional LSTM layer. This model, without relying on any pre-processing or manual feature engineering, achieved a Challenge validation score of 62.3% and a full test score of 38.2%, ranking us 9 th out of 41 teams in the official ranking. Our second neural model, designed through neural architecture search, did not score on the full test dataset nor on the validation dataset; however, we optimistically expect performance improvement compared to our handdesigned RCNN model. This model scored 64.4% using 10fold cross-validation on the training dataset, which is 2.5% higher than the training score of our RCNN model, using 10-fold cross-validation. 1.