A Systematic Review of Heart Sound Detection Algorithms: Experimental Results and Insights

Automatic heart sound detection plays a vital role in the early detection of cardiovascular diseases (CVDs). In recent years, many heart sound detection algorithms have been proposed, becoming a popular research topic in medical diagnosis. In this article, our goal is to answer which heart sound detection algorithms and heart sound features perform best in which situations and to explore the problems in existing research. This work will provide ideas for subsequent research in this area. We achieve this aim through a standard process that includes preprocessing, segmentation, feature extraction, and classification. We discuss the length and overlap rate in segmentation and analyze the classification methods, especially the nine most salient heart sound features. The performance of the existing techniques is evaluated in different datasets and at the same level of comparison. The experiments show that the segmentation length includes at least one complete period, and we should use a large overlap rate in the segmentation phase. For feature extraction, time-domain feature (TIME) and fast Fourier transform (FFT) features based on single independent variables perform better in deep models and traditional classifiers. Short-time Fourier transform (STFT), continuous wavelet transform (CWT), and S-transform (ST) features based on double independent variables can perform well in all classification models; Mel spectrum/Mel frequency cepstrum coefficient (MFCC) is only better on deep neural networks (DNNs). The best performance is achieved by combining TIME or MFCC with DNNs.