An Ensemble Neural Network for Multi-label Classification of Electrocardiogram.

An electrocardiogram (ECG) record potentially contains multiple abnormalities concurrently, therefore multi-label classification of ECG is significant in clinical scenarios. In this paper, we propose an ensemble neural network to address the multi-label classification of 12-lead ECG. The proposed network contains two modules, which treat the multi-label task from two different perspectives. The first module deals with the task in a sequence-generation manner by a novel encoder-decoder structure. The second module treats the multi-label problem as multiple binary classification tasks, by employing two convolutional neural networks of different structure. Finally, the predictions of two modules are integrated as the final result. Our method is trained and evaluated on the dataset provided by the First China ECG Intelligent Competition, and yields a Macro- F 1 of 0.872 on the test set.