A Study on the Effectiveness of the Comparative Neural Network Model for Abnormal Beat Detection in Electrocardiogram Signals

Abnormal beat detection is an important research field in electrocardiogram (ECG) signal analysis. However, because the shapes and characteristics of beats vary according to the individual, it is difficult to classify normal and abnormal beats. To imitate cardiologists' analysis scheme, deep learning based analysis is becoming active. In particular, cardiologists' abnormal beat detection techniques resemble comparative learning in that they use normal beats as a reference. In this paper, we examined a comparative learning method by acquiring a normal reference beat using a template cluster to imitate a cardiologists' scheme. To analyze a suitable model for the comparative learning of ECG signals, we tested our method using ResNet, GoogLeNet, and DarkNet, which are widely used models provided by MATLAB deepNetworkDesigner. Our experimental results indicate that GoogLeNet minimized non-detection, DarkNet minimized over-detection, and ResNet showed intermediate results. In ECG signals, it is important to minimize the non-detection of abnormal beats. Thus, we confirmed that GoogLeNet is effective for comparative learning.