Analysis and Selection of Neural Network Architectures for Recognizing Heart Diseases by Cardiogram

• Published in: 2020 International Conference on Engineering Management of Communication and Technology (EMCTECH) pp. 1 - 8
• Main Authors: Brus, V. R., Voronova, L. I.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 20-10-2020
• Subjects: Annotations; atrial extrasystole; Biological neural networks; cardiogram; cardiology; Convolution; convolutional neural network; disease recognition; diseases; ECG; electrocardiography; fully connected neural network; Heart; heart contraction; heart diseases; heart diseases recognition; medical signal processing; neural net architecture; neural network; PhysioNet database; residual network; Rhythm; signal classification; Training; ventricular extrasystole
• DOI: 10.1109/EMCTECH49634.2020.9261514

The article describes the development of a computer application with elements of artificial intelligence in the field of medicine. The authors solve the problem of recognizing heart diseases from cardiogram by using neural network modeling. The system recognizes 8 classes that include diseases of different severity and normal rhythm. Preparing of the training set based on data from the PhysioNet database is explained. Experiments on parameterization of neural network architectures to improve recognition accuracy are described. The results of training a fully connected neural network, a convolutional neural network, and a convolutional network with residual connections are presented. The fully connected network showed low disease recognition results. The worst accuracy was 39% when recognizing ventricular extrasystole. The convolutional network showed higher accuracy in recognizing most classes. The worst accuracy was 81% and was obtained when recognizing atrial extrasystole. As a result of testing the residual network, the same accuracy of recognition of atrial extrasystole was obtained, but the accuracy of recognition of other classes became more than 90%. Thus, it is proved that the best architecture for recognizing diseases from a cardiogram is the residual network.