An atrial fibrillation detection system based on machine learning algorithm with mix-domain features and hardware acceleration.

Abstract

This paper presents a real-time electrocardiogram (ECG) analysis system that can detect atrial fibrillation (AF) using machine learning algorithms without a cloud server. The system takes advantage of the heterogeneous structure of the Zynq system-on-chip (SoC) to optimize the tasks of local implementation of AF detection. The features extraction is based on multi-domain features including entropy features and RR interval features, which is conducted using the embedded micro controller to generate significant features for AF detection. An AF classifier based on artificial neural network (ANN) algorithm is then implemented in the programmable logic of the SoC for acceleration. The validation of the proposed system is performed by using the real-world ECG data from MIT-BIH database and CPSC 2018 database. The experimental results show an accuracy 93.60% and 97.78% when tested on these two databases respectively. The AF detection performance of the embedded algorithm is majorly identical to that of the PC-based algorithm, indicating a robust performance of hardware implementation of the AF detection.