Computer Aided Diagnosis Algorithm Of Heart Sound Signals In Congenital Heart Disease

Heart sound is an important physiological signal of the human body,which can be used as a key basis for the diagnosis of congenital heart disease.The traditional diagnosis method of congenital heart disease is for doctors to go to the countryside for auscultation screening.Due to the limited human resources,this method is expensive and inefficient,so it is difficult to spread to all areas.With the help of digital information processing technology,this thesis presents a computer-aided diagnosis algorithm for heart sound signals of congenital heart disease,to provide a new technical means for congenital heart disease screening.Heart sound is a weak low-frequency signal,so it is easily affected by the clinical environment in the process of acquisition.Therefore,it is a great challenge to classify heart sounds effectively and accurately in the clinical environment.Starting from the practical application,the three stages of heart sound preprocessing,heart sound feature extraction,and heart sound classification was redesigned in this thesis.A set of feasible algorithm frameworks for computer-aided diagnosis of heart sound signals was constructed,which explores a new technical route for the computer-aided auscultation system in the clinical environment.In the Heart sound preprocessing step,the wavelet algorithm was used to denoise the heart sound.To segment heart sound accurately without using ECG,a duration method is introduced as the probability transfer condition of the hidden Markov model in this thesis.A duration hidden Markov model（DHMM）is constructed for heart sound segmentation,and it is stored as a heart sound sample based on the heart cycle.An average evaluation accuracy score（F₁）of 0.933,an average sensitivity of 0.930,and an average precision of 0.936 were achieved on a large representative dataset by this work.The algorithm has good robustness and anti-noise performance,which is the key step to automate the machine-aided diagnosis algorithm proposed in this thesis.To solve the phenomenon of cardiac cycle inequality caused by heart rate variability,the method of dynamic frame length was used to extract the Mel-frequency spectral coefficients（MFSC）of the heart sound.The feature extraction method can transform the cardiac cycle signal of any duration to the same plane dimension.At the same time,the relative reference position information of each heart sound state is retained.The MFSC feature can retain the time domain information,frequency domain information,and energy information of heart sound at the same time,which makes an outstanding contribution to improving the accuracy of the classifier.The convolution neural network（CNN）was used to classify the heart sound according to MFSC features in this thesis.Considering that the heart sound recording of each contains multiple cardiac cycles.The majority voting algorithm is used to optimize the classification results,and multiple heart cycles are mapped to individuals to improve the accuracy of heart sound classification.Finally,on the two-classification problem,the accuracy,sensitivity,and specificity were 0.939,0.928,and 0.950 respectively.In multi-classification,the accuracy was 0.863.The above steps together constitute a computer-aided diagnosis algorithm for heart sound signals of congenital heart disease.Because each step is designed from the point of view of clinical application,the algorithm has high accuracy,good robustness,a large data set,and strong universality.It opens up a new technical means for the screening of congenital heart disease in the clinical environment.