Research on Heart Sound Recognition by Multi-feature Combination Based on Mel Frequency and Linear Prediction Coefficient

Due to the problem that the accuracy of the existing algorithms was not good enough, a novel method of improved heart sound recognition by multi-feature combination based on Mel frequency and linear prediction coefficient was proposed in this study. Firstly, the non-linear transformation was performed on the linear prediction coefficients according to the Mel scale to obtain the spectral characteristic parameters of the heart sound signal. Secondly, the feature parameters were processed by differential processing and weighted reorganization to complete a 2-D feature map. Finally, the deep learning models, such as Inception_v4 was used to perform classification of heart sound signal. The experimental results showed that the algorithm performs better than the several existing algorithms. An accuracy of 93.7
^%
, a sensitivity of 95.6%, and a specificity of 89.5 % were achieved using the proposed method in this work. It is expected to be used for machine-assisted diagnosis of the initial diagnosis of congenital heart disease.