An Ensemble Approach to Classifying 12-lead ECGs

David Kaftan1 and Richard Povinelli2
1Marquette Energy Analytics, 2Marquette University


The aim of this work is to classify 12-lead ECGs into nine classes, including normal sinus rhythm, atrial fibrillation, left bundle branch block, and ST-segment depression and elevation. The training data consists of 6,877 recordings from six to 60 seconds in duration. We will take a Bayesian ensemble approach, combining a reconstructed phase space — Gaussian mixture model method, a convolutional deep neural network, and a long short-term memory network. In addition to ensembling at the model level, we will also ensemble at the lead level. A model is learned for each lead and these lead level models are combined using a Bayesian ensembler. This work is a submission to the PhysioNet/Computing in Cardiology Challenge 2020. As such there is a public training dataset and a private test set. During the unofficial phase of the competition on the test set, we achieved an F-Beta score of 0.746 and a G-Beta score of 0.489 using only a deep convolutional neural network. We attribute these low scores to the fact that we have not yet ensembled all the models, nor ensembled the models for each lead.