Intra-group and inter-group electrocardiograph coding image fusion and classification based on multi-scale group convolution feature fusion network

Abstract

Electrocardiograph (ECG) signals exhibit outstanding advantages in the diagnosis of cardiovascular diseases, but their classification is difficult to perform in present studies. The first lead of 12 lead ECG signal has important scientific significance and convenience on the classification of cardiovascular disease. This study proposes a classification model in which the time domain feature and independent component extracted from the first lead ECG signal were transformed into images, and a multi-scale group convolutional feature fusion (MSGCFF) was used to classify seven signal types. First, the time-domain features and principal component features of ECG signals are extracted and thereafter encoded into ECG images using the Gramian angular fields coding method. Next, the depth features of different output layers are extracted by training the ECG image data with the MSGCFF model, and the features are fused by the parallel method. Finally, the fused features were classified using the Softmax classifier. The experimental results were evaluated using the Chinese cardiovascular disease database dataset. The accuracy, precision, sensitivity, and F1 scores of the method were 93.0533%, 81.6354%, 94.8723, and 85.9987%, respectively. This method effectively reconstructs the ECG signal, which not only retain the time dependence, but also utilized the advantages of convolution operation to capture multi-scale features, providing a feasible method for the study of cardiovascular disease classification.