Deep Atrial Fibrillation Classification Based on Multi-modal Attention Network

Electrocardiography (ECG) is a popular technique for Atrial Fibrillation diagnosis. Due to the enormous variability of ECG waveforms, the precise detection of characteristic ECG points is a challenging task. Hence, there have no universal rules for determining the range of individual component waveforms. In this paper, we propose a multi-modal attention network named MMAN to improve the performance of ECG classification. Specifically, we design the MMAN based on a two-stream CNN and multi-modal attention module (MMAM). The two-steam CNN extracts the multi-modal patterns from the multi-level ECG features and the original ECG signal. Then, the MMAM is proposed to obtain the weighted multi-modal features. Benefiting from the multi-modal information and attention mechanism, the MMAN improves the performance of ECG classification. Experiment results show that the MMAM-based models perform well on the 2017 PhysioNet/CinC Challenge and MIT-BIH Arrhythmia datasets.