MSCMN : Multi-scale convolutional modulation network for atrial fibrillation detection from a single-lead electrocardiogram

Abstract The development of wearable sensor technology has made the acquisition of continuous single-lead electrocardiogram (ECG) data more convenient and accessible. Atrial fibrillation (AF) detection in a single-lead ECG wave is becoming a popular research topic. However, it is challenging due to the following issues: 1) the AF signal is weak and frequently disturbed by nonstationary noises; 2) both AF and non-AF rhythms have various morphologies; and 3) a single-lead ECG lacks spatial correlation information on the electrical movement of the heart. To solve these issues, we propose a multi-scale convolutional modulation network (MSCMN) composed of stacked modulators and converters. The modulator achieves an efficient self-attention mechanism that adaptively suppresses the noises by computing the Hadamard product between the multi-scale nonlinear convolutional features and the linear channel mixing features. The converter, which consists of two partial convolutional layers and one linear layer, is used to capture ECG temporal dependency and cross-channel correlation among the modulated features. Furthermore, because AF rhythms exhibit irregular R-R intervals, we employ an early fusion of R-peak position encoding and ECG waveforms to enhance feature extraction. We conducted extensive experiments on three public datasets. The results consistently validate the effectiveness of the proposed MSCMN in detecting AF from a single-lead ECG.