CNN Based Atrial Fibrillation Diagnosis with ECG Signals

Abstract

Background/Objectives: Atrial Fibrillation (AFib), one of the arrhythmias, causes the atrium to beat irregularly. AFib can be diagnosed by observing electrocardiogram (ECG) signals. However, the degree of irregular running depends on the patient, and it is difficult to detect AFib in early patients. As a result, it is important to accurately judge Sinus Rhythm (SR) and AFib, and only doctors with extensive experience in cardiology are known to judge accurately. In this paper, we propose a convolutional neural network (CNN) to perform accurate AFib diagnosis from ECG signals.

Methods/Statistical analysis: The proposed AFib diagnostic technique has the following characteristics. The proposed artificial neural network consists of three layers CNNs and a fully connected layer. The final decision uses an ensemble of five models to reduce the deviation from the decision results and increase the decision accuracy. The design and performance verification of the proposed CNN diagnostic technique is based on Python TensorFlow 2.0.

Findings: Diagnostic performance is evaluated using ECG signals obtained from real-world normal people and patients with AFib. The evaluation results show an accuracy of approximately 96.25%, 98.72% sensitivity, and 93.39% specificity. Furthermore, we analyze using Gradient-Class Activation Mapping (Grad-CAM) what part of ECG signals the proposed CNN distinguishes AFib rhythm from SR.

Improvements/Applications: The proposed technique can be used as a method to accurately diagnose patients with AFib only with ECG signals without the help of a doctor. It could also be applied to a variety of wearable healthcare devices to diagnose AFib around the clock.