Heart Sound Classification Algorithm Based on Sub-band Statistics and Time-frequency Fusion Features

Abstract

The clinically acquired heart sound signals always have inevitable noise, and the statistical features of these noises are different from heart sounds, so a heart sound classification algorithm based on sub-band statistics and time-frequency fusion features is proposed. Firstly, the statistical moments (mean, variance, skewness and kurtosis), normalized correlation coefficients between sub-band and sub-band modulation spectrum are extracted from each sub-band envelope of the heart sound signal, and these three features are fused into fusion features by Z-score normalization method. Finally, a convolutional neural network classification model is constructed, which are used for training and testing. The experimental results showed that the accuracy, sensitivity, specificity and F1 score of the algorithm were 95.12%, 92.27%, 97.93% and 94.95%, respectively. It has great potential in machine-aided diagnosis of precordial diseases.