An ECG Beat Classification Method using Multi-kernel ResNet with Transformer

In this work, we investigated the electrocardiogram (ECG) for cardiovascular disease diagnosis using the deep learning method. The diagnosis of a more complicated cardiovascular disease can be performed by analyzing beat-unit abnormalities. Beat classification plays the basic role of classifying arrhythmia in ECG. To classify the four beat classes of the MIT-BIH arrhythmia data base, we extracted features using the multi-kernel ResNet, trained the system to successfully learn the mutual information of continuous beats using the transformer encoder, and enabled the transformer to learn stably by applying position embedding based on the heart rate. We used the icentia11k data set, which is a big public data set, for the transfer-learning and could make a more general model with a variety of patient data. We obtained an average F1 score of 0.839, which is better than any existing score, for the four-class beat classification of the MIT-BIH arrhythmia database.