Multi-scale attention convolutional neural network for noncontact atrial fibrillation detection using BCG

Atrial fibrillation (AF) is the most frequently occurring clinical arrhythmia. It is of great significance to develop an AF screening and early warning system applicable to daily life scenarios. In this manuscript, we proposed a noncontact AF detection method based on ballistocardiogram (BCG) and convolutional neural network. A BCG dataset consisting of 9405 nonoverlapping thirty-second segments was first constructed after wavelet denoising, root mean square (RMS) filtering and segmentation. Then, we proposed a multi-scale attention convolutional neural network (MSA-CNN) to automatically detect AF from BCG segments. The network allowed different input length ranging from 5s to 30s and used multi-scale convolution to capture the deep features of BCG at different scales, and built an attention module to learn feature weights automatically. The results showed that under the inter-patient evaluation, the proposed MSA-CNN achieved 95.0% AF detection sensitivity and 97.1% overall classification accuracy with 5-s BCG segments as input. The results indicated that the proposed method may lay foundations for the development of long-term home cardiac monitoring and AF screening system.