DeepAF: A Multi-task Deep Learning Model for Arrhythmias Detection at Resource-Constrained Mobile Device

Atrial fibrillation (AF) is the most commonly treated arrhythmia and is associated with the risk of stroke and heart failure. As its diagnosis is often based on the analysis of a Holter electrocardiogram (ECG), the use of automated detection methods is common. Recent AF detection methods typically use 1D convolutional neural network (CNN) architectures. Although these models achieve relatively good performance, they are still complex, resulting in high implementation costs on mobile phones. In this paper, we propose a multitask CNN architecture to improve detection accuracy and reduce network complexity. The model takes three inputs (ECG, signal quality features and arrhythmia features) and returns two outputs (signal quality and arrhythmia type). For model training and evaluation, we built a dataset with the PhysioNet Challenge 2017, Brno Quality, PhysioNet Challenge 2011 and TeleECG databases. Experimental results and analysis show that our method achieves an overall accuracy of 96.98% and 78.3% for noise and arrhythmia classification, respectively. Compared to existing deep learning-based methods, the proposed network shows acceptable detection accuracy and lower network complexity, making it suitable for mobile applications.