K-B2S+: A one-dimensional CNN model for AF detection with short single-lead ECG waves from wearable devices

Abstract

Wearable signal analysis is an important technology for monitoring physiological signals without interfering with an individual's daily behavior. As detecting cardiovascular diseases can dramatically reduce mortality, arrhythmia recognition using ECG signals has attracted much attention. In this paper, we propose a single-channel convolutional neural network to detect Atrial Fibrillation (AF) based on ECG signals collected by wearable devices. It contains 3 primary modules. All recordings are firstly uniformly sized, normalized, and Butterworth low-pass filtered for noise removal. Then the preprocessed ECG signals are fed into convolutional layers for feature extraction. In the classification module, the preprocessed signals are fed into convolutional layers containing large kernels for feature extraction, and the fully connected layer provides probabilities. During the training process, the output of the previous pooling layer is concatenated with the vectors of the convolutional layer as a new feature map to reduce feature loss. Numerous comparison and ablation experiments are performed on the 2017 PhysioNet/CinC Challenge dataset, demonstrating the superiority of the proposed method.