Multiclassification for heart sound signals under multiple networks and multi-view feature

Abstract

Heart sound signals play a crucial role in assessing cardiovascular health. Building upon this, we have designed five networks to effectively classify these signals. Initially, it is essential to standardize them. Subsequently, we generate four types of two-dimensional (2D) heart sound feature images, namely Spectrogram, Mel-spectrogram, Bispectral analysis, and Phonocardiogram (PCG). To enhance performance, we refine the block in ConvNext and compare two one-dimensional (1D) ConvNext networks. After obtaining a superior network, we merge it with a pre-trained 2D ConvNext network on ImageNet. Then, we introduce the Encoder block from the Transformer network and combine it with the improved 1D ConvNext network. Among them, the Conv-Encoder1 network achieved classification accuracies of 98.7%, 91.9%, and 97.9% on the three datasets, respectively. Furthermore, by observing all the test results, it can be concluded that utilizing 1D heart sound signals for automatic classification offers significant advantages compared to using 2D feature images.