A Study on Detecting Atrial Fibrillation by Identifying Its ECG Waveform Features with CNN

Abstract

Atrial fibrillation (AF) is one of the arrhythmias that causes a stroke, and early detection is required. In many ECG inspections, irregular RR-intervals (RRI) are first detected. Therefore, rare AF with regular RRI could be overlooked. As the other features, an ECG waveform of AF has several small waves (f-wave) and the absence of P-waves. In this study, a novel AF detection method by identifying these two features with a convolutional neural network (CNN) is proposed. In preprocessing, the artifacts on the ECG waveform are removed by a bandpass filter. First, the R-waves are detected using a variable threshold. Next, the waveform in 0.4 s before each R-wave is extracted. Then the f-wave and P-wave are identified using CNN with the LeNet structure, and AF is detected. Evaluating the performance with untrained ECG waveforms of 10 subjects, the accuracy was 84.0%. This result indicates the feasibility of this method.