MSSCNN: An intelligent heart disease diagnosis method based on heart sound signals under limited data

Heart auscultation, a simple and effective diagnostic approach for heart diseases, heavily relies on the subjective judgment and clinical experience of physicians, necessitating the urgent development of objective and intelligent heart sound diagnostic methods. Deep learning technology provides a new solution for this purpose, but its effectiveness depends on sufficient high-quality training data. To address the challenges of scarce training data and the limitations of existing methods in handling data scarcity, a novel diagnostic approach named MSSCNN, based on Mel Spectrograms (MSs) and an improved Siamese convolutional neural network (SCNN), is proposed in this paper to enhance intelligent heart disease diagnosis under limited data conditions. Specifically, the method first converts the heart sound signals into MSs and randomly selects a small number of samples as training data. Next, pair these training samples pairwise within the group to increase the number of training samples, and feed these paired samples into the MSSCNN model for feature extraction and classification. Then, input the test set samples into the model for performance testing. Finally, two experimental cases confirm that MSSCNN can efficiently utilize limited training samples to tackle the scarce data issue and achieve better intelligent heart disease diagnosis. This study not only provides new methods to solve the problem of scarce training data, but also opens up new avenues for objective, accurate and timely diagnosis of cardiovascular diseases.