Heart Sound Signal Processing Methods Based On HSMM And CNN

In recent years,the prevalence rate of cardiovascular disease in China is continuously increase.The urgent task is to take effective measures actively to reduce the prevalence rate.Heart sound signals could reflect situations of heart tissues,therefore processing of heart sound signals is particularly important.Heart sound analysis and processing are usually consist of the following aspects: denoising,envelope extraction,segmentation,feature extraction and classification.This research focuses on the segmentation and classification methods of heart sound signal processing.(1)A new algorithm proposed to solve burrs of the signal’s envelope originated from the Hilbert transform in the Hidden Semi-Markov model(HSMM)based on Logistic Regression(LR).Namely,a heart sound segmentation algorithm for HSMM based on Support Vector Machine(SVM)and Shannon energy.Firstly,the wavelet denoising method is used to denoise signals.According to R peak and T wave,the heart sound signals are labeled.Then envelope of Shannon energy and other feature sequences are extracted.Next,the parameters of LR-HSMM are adjusted and a Viterbi algorithm is applied to infer the most probable states.Finally,the first heart sound and second heart sound are identified by the SVM.The algorithm has not set a hard threshold and suppresses noise more effectively that is benefit for extracting the envelope.(2)Traditional classification methods have been widely used in heart sound recognition.Due to the wide variety of cardiovascular diseases in clinical practice,the simple two-class classification of heart sounds can’t meet with practical applications.Aiming at the above situation,a Convolutional Neural Networks(CNN)algorithm based on Mel Frequency Cepstral Coefficient for multi-category heart sounds recognition is proposed.Firstly,use wavelet transform to denoise the original heart sound signal,then extract the MFCC parameters of the heart sound.Next,build a CNN model and a Long Short-Term Memory(LSTM)model,adjust the relevant parameters continuously.Finally,the respective trained models are used to recognize the heart sound signals.In this paper,the heart sound segmentation algorithm based on HSMM combined with Shannon energy is proposed.The envelope of extracted heart sound is smoother and has good anti noise performance.The segmentation accuracy is significantly improved compared with the reference algorithm.The proposed heart sound classification algorithm based on CNN uses pooling layer and weight sharing to extract more useful information,which greatly reduces the training time and improves the recognition accuracy compared with the other two reference models.In a word,the research method of heart sound segmentation and classification proposed in this paper plays an important role in the efficient analysis and processing of heart sound.