An End-to-end Deep Learning Scheme for Atrial Fibrillation Detection

The aim of this study was the detection of atrial fibrillation (AF) from continuous ECG analysis. In this study, an end-to-end deep learning scheme was proposed. When the scheme was applied, 30-second multi-lead ECG data segments with an overlapping window of 5 seconds were preprocessed and sequentially fed into a multi-layer residual convolutional neural network (CNN) to extract ECG's multi-scale local morphological (spatial) features, The generated local spatial features were then processed by the following two bidirectional long short-term memory (LSTM) layers, and the output sequences of the LSTM layers were then weighted by an attention module and processed by a following dense network to complete AF detection. Finally, the sequential detection results were further processed to improve accuracy. To demonstrate its effectiveness, the proposed scheme was trained and tested on multiple ECG databases annotated by cardiologists. Episode and duration accuracies were calculated according to the performance evaluation method of atrial fibrillation detection defined in the EC57 standard [1]. An episode F1 score of 85.7% and a duration F1 score of 95.5% were achieved on the independent testing dataset.