Classification of 12-lead ECG with an Ensemble Machine Learning Approach

Abstract

The PhysioNet 2020 Challenge focused on the automatic classification of 27 cardiac abnormalities (CAs) from 12-lead ECG signals. We investigated on a hybrid approach, combining average-template-based algorithms with deep neural networks (DNNs), to build an ensemble classification model. We calibrated the model on the available 40,000+ ECGs, while organizers tested the model on a private test set. Standard ECG preprocessing was applied. For ECGs related to CAs altering the ECG morphology, multi-lead average P, QRS, and T segments were computed. For signals associated with irregular rhythms, time dependent features were computed. The ensemble model comprised of: i) three DNNs to classify morphology-related CAs. ii) a fully connected neural network to classify irregular rhythm; and iii) a threshold-based classifier for premature ventricular beat detection. The organizers designed a score for ranking the models. The ensemble model proposed by our team 'BiSP Lab' reached the 40th position, and obtained a score of -0.179 on the private test set. Despite the low performance obtained on the private test set, our ensemble model showed potential for classification of CAs from ECGs.