Noninvasive Prediction of Atrial Fibrillation Recurrence Based on a Deep Learning Algorithm

Atrial fibrillation (AF) is an abnormal heart rhythm. The goal of radiofrequency ablation for AF is to regain a normal heart rhythm. Presently, the commonly used algorithms for prediction of AF recurrence face constraints such as no flexible feature selection and optimization. Deep learning algorithm, such as convolutional neural networks (CNN), can overcome these drawbacks by calculating features automatically. This study presents a noninvasive ECG-based approach with deep learning algorithm to predict AF recurrence according to preoperative AF signals of 14 patients. The features of cardiac activity can be extracted from the convolution neural network and then fed to multilayer perceptron (MLP) for classification. We have evaluated the result by measuring accuracy (ACC), sensitivity (SE) and specificity (SP). In the meantime, the optimal CNN combination of parameters are showed in a table. Our proposed method has shown good merits with ACC of 93.14%, SE of 83.5% and SP of 95.99%. It can be concluded that BSPM would contain more potential information because of its wide range of spatial coverage compared with traditional signal acquisition system, and the deep learning algorithm of CNN has more advantages in feature extraction for prediction of AF recurrence.