Yuchen Jiang, Peirong Zheng (co-first author) and Dakun Lai, “A Semi-supervised Algorithm for Atrial Fibrillation Attack Prediction Using Convolution Auto-encoder of Time Series Signal.”

Download paper here

Abstract—At the beginning of the attack, atrial fibrillation (AF) usually presents as paroxysmal atrial fibrillation (PAF), and may be further transformed into persistent AF that may cause high-risk diseases such as ischemic stroke and heart failure. Under considerations of current machine learning algorithms for AF predictions that manure extraction of features and tag electrocardiograph (ECG) data are time consuming and labor-intensive processes, a novel two-stages of semi-supervised algorithm for AF attack prediction is proposed in this paper. With the input of the time series signal of RR interval, the first stage is designed as an unsupervised learning based on convolution auto-encoder (CAE) network and the second stage is a supervised learning based on Long short-term memory (LSTM) model. A total of 855,882 heartbeat activities including 30 segments of PAF and 30 segments of normal heart rate were collected so as to evaluate the performance of the CAE-LSTM combination model. The obtained results showed that the averaged accuracy and the root mean square error (RMSE) of ten-fold cross-validation are 90.16% and 0.0113, respectively. In summary, the preliminary results suggested that the combination of the unsupervised CAE model and of the supervised LSTM model could reduce the dimension of the input data meanwhile perform the subsequent classification with a small amount of labeled data as input. Moreover, the proposed algorithm could be useful to predict AF when the sample is scarce, and potentially prevent the occurrence of a paroxysmal atrial fibrillation.