Single-lead ECG recordings modeling for end-to-end recognition of atrial fibrillation with dual-path RNN

Abstract

Atrial fibrillation (AF) is the most common type of sustained cardiac arrhythmia, and is associated with stroke, coronary artery disease and mortality. Thus, early detection is crucial to avoid serious complications. Existing methods require specialized equipment and technical expertise, and accurate machine learning diagnosis of AF remains a dream. In this paper, we propose an end-to-end AF recognition method with dual-path recurrent neural network (DPRNN) from single-lead ECG. The model takes the whole ECG as input, and DPRNN splits the ECG into shorter segments and models the sequence between intra- and inter-segment iteratively. A mix-up operation is used for data augmentation, which overcomes the issue of limited data. We evaluated our method on the dataset from PhysioNet Challenge 2017. Experimental results shows that the proposed method can both effectively recognize AF with ECG signal without any human expertise, and outperforms state-of-the-art baseline methods. This demonstrates that dual-path model is effective for ECG analysis. We postulate that this framework can be generalized for other medical sequence signal.