A Transfer Learning Approach to Detect Paroxysmal Atrial Fibrillation Automatically Based on Ballistocardiogram Signal

Atrial fibrillation (AF) is a common arrhythmia disease, which is divided into persistent and paroxysmal. Due to the uncertainty for the onset of paroxysmal atrial fibrillation (PAF), it is crucial to detect AF timely in daily monitoring for the early diagnosis and treatment of the cardiac disease. The popular method to detect AF is based on electrocardiogram (ECG) signal, and the gold standards of diagnose are the disappearance of the P-wave and the replacement by the f-wave of varying size and irregular shape. However, it isn't suitable for PAF monitoring in home environment because of the electrodes fixed on the body surface. Therefore, this paper proposes a transfer learning method to detect the occurrence of PAF automatically with ballistocardiogram (BCG) signal, which non-invasive records the micro-movements produce by the recoil forces to maintain the overall momentum. In this study, the ECG.data from database and BCG data from acquisition equipment were preprocessed and segmented as 24 s frame firstly. And then, a 17-layer one-dimensional convolution neural networks (CNN) was designed and pre-trained by ECG data, which is transferred to the BCG data to classify the AF or non-atrial fibrillation (NAF). The knowledge of ECG data with large amount contributes to the PAF detection based on the BCG data with small amount. It is beneficial to improve the classification performance due to the correlation between ECG and BCG. A total of 10000 frames of ECG data from database were applied to pre-train CNN and 1200 frames of BCG data from acquisition equipment were used to fine-tune the network and achieve the classification results. Consequently, the classification performance including the accuracy, sensitivity, specificity, and precision are 95.8%, 98.3%, 93.3%, and 93.7% respectively. For comparison, the CNN directly, support vector machine (SVM), random forests (RF) approaches were implemented with the same batch of small volume BCG data. This paper proposes a transfer learning method to detect AF automatically from ECG to BCG, which has the advantages of better performance and higher speed, compared with the state-of-art methods. Therefore, it is suitable for real-time and unobtrusive PAF monitoring daily at home.