A Deep Ensemble Model with an Efficient Feature for Multi-class Arrhythmia Classification Utilizing 12-Lead ECG Signal

Abstract

Arrhythmia classification using an electrocardiogram (ECG) signal has obtained a wide range of attention as arrhythmia is a potentially fatal heart disease that necessitates immediate medical attention. Automated arrhythmia identification and categorization utilizing computational techniques can fasten the proper treatment. In this paper, a novel deep ensemble model is proposed where both time-domain and frequency-domain characteristics of ECG signals are explored for the purpose of automatic arrhythmia classification. Moreover, an efficient feature called Time Multiplexed Fast Fourier Transform (TMFFT) is extracted that provides useful information for categorization in the frequency domain. In addition, an end-to-end deep neural architecture namely LACRNN (Lead-wise Attention with CNN and RNN) is also proposed to classify arrhythmia from the raw ECG signal while the extracted TMFFT feature is applied to a simple 2D CNN. Very satisfactory performances are obtained for each class of arrhythmia where the experimentation is carried out on a publicly available dataset namely the China Physiological Signal Challenge (CPSC) dataset. The average F1 score obtained in the 9-class problem is 90%.