Improved Detection of Paroxysmal Atrial Fibrillation Using an ECG-based Semisupervised Model

Paroxysmal atrial fibrillation (PAF) is a common intermittent supraventricular arrhythmia whos identification can be challenging.The 12-lead electrocardiogram (ECG) is a cheap and non-invasive tool, ideal for diagnosing this condition.In recent years, neural networks (NNs) have been used in cardiology for the prediction of cardiac risk, however, their use is still limited due to the scarcity of wellcurated labelled data.The objective of this project is to check if pretraining enhances the performance of NNs in a supervised learning task for the diagnosis of PAF, given limited training data with an unbalanced distribution of cases.First, ECG datasets publicly available were downloaded and preprocessed, resulting in a total of 1,602,185 ECG signals.Then, a multilayer convolutional neural network (CNN) was implemented for the diagnosis of PAF.Next, we developed and pretrained, in a self-supervised manner, a NN based on contrastive predictive coding.Both supervised and semisupervised models showed a good performance.By using a simplified CNN together with a selfsupervised pretrained NN, accuracy improved from 0.565 to 0.680, the area under the curve rose from 0.692 to 0.714, and specificity at a sensitivity of 0.750 grew from 0.534 to 0.579.Our findings support that the semisupervised learning method improved the performance and simplified the supervised model architecture and training.