| The prevalence of congenital heart disease among newborns in China is about 8.98‰,which seriously threatens the health of children and causes a heavy economic burden to the families of affected children.Although many screening efforts for congenital heart disease have been carried out in China,there are serious cases of underdiagnosis due to uneven medical development and imperfect screening mechanisms.The home use of electronic stethoscopes provides the conditions for the development of intelligent auscultation screening for neonatal precocious heart disease.Both traditional heart sound classification algorithms and deep learning classification models trained based on large-scale data need to undergo pre-processing such as signal interception and noise reduction for subsequent studies.Various types of noise,especially the crying noise,often occur in children’s phonocardiograms(PCGs),resulting in poor quality of the collected signal.Currently,manual interception of signal segments is time-consuming and subjective,and the noise reduction methods used fail to suppress these noises well or lose murmurs that reflect pathological information when removing the noise.In this paper,we used a balanced dataset consisting of 124 PCGs collected by the home electronic stethoscope and 16 randomly selected normal PCGs from Physio Net dataset.We conducted the research based on the dataset,which mainly included the automatic interception of the best signal segment from the original acquisition signal,the noise reduction for children’s PCGs,and the establishment of intelligent screening systems for the children with congenital heart diseases.1.A method for automatic interception of the optimal PCG segment based on power spectral density was developed.After the localization of the first heart sound(S1)and the second heart sound(S2)according to the child’s age,the similarity of power spectral density of three adjacent heartbeat cycle sequences starting and ending at the peak points is calculated and taken as the quality factor.To ensure that the three complete cycle signals were retained,the cycle sequence with the largest quality factor was intercepted,and the final intercepted signal was regarded as the best PCG segment.The results of automatic interception of 140 children’s PCGs showed that the method could effectively reduce noise interference in the signal segments used for later researches,and the success rate of automatic selection was high enough to meet the requirements of intelligent auscultation.2.A novel denoising method based on combined variational modal decomposition(VMD)and wavelet soft threshold algorithm(WST)was proposed in this paper,and a mechanism for screening modes based on permutation entropies and correlation coefficients was established.After the 6-layer decomposition of PCG by VMD,the screening index of each mode was evaluated for the screening of modes,and the WST was applied to further address the retained modes.The proposed method was compared with the conventional VMD denoising method and the only WST method for denoising experiments on children’s PCGs,and the results show that the new combined noise reduction method could suppress the noise(especially the crying noise)in children’s PCGs without losing murmurs.The best denoising evaluation metrics were obtained by the combination of applying WST to each modality first and then modal reconstruction.3.Intelligent diagnosis systems were established for congenital heart diseases.The 140 automatically intercepted PCGs were denoised by the combined VMD-WST method and the only VWG method,and the features were subsequently calculated.Among them,MelFrequency Cepstral Coefficients(MFCCs)were employed as inputs to the convolutional neural network(CNN),and 10 time and frequency domain features were used as inputs to the BP neural network and support vector machine(SVM).The results show that the noise impact on the features could be reduced by improving the noise reduction method,thus improving the performance of the classification.The best performance of the screening system based on the BP network model with the VMD-WST noise reduction method was achieved with the accuracy,sensitivity,specificity and Auc of 96.19%,96.39%,95.68% and 0.9707,respectively.The innovation of the current work is that the difference between children’s heart rates at different ages was considered for the localization of S1 and S2,and the quality of the PCG is evaluated from the frequency perspective,which is the basis for the proposed method to quickly intercept the best PCG segment.The proposed denoising method is highly adaptive and makes up for the deficiencies of suppressing strong Gaussian noise by VMD and abrupt noise by WST.In addition,the importance of the denoising method for classification studies is demonstrated by the cross-sectional comparison of multiple classification models.This paper provides both an objective and convenient sample interception method for the studies of heart sounds from children with congenital heart diseases and an effective combined denoising method for the classification of PCG,which is conducive to improving signal quality and promoting the development of intelligent screening systems for congenital heart disease for home use and productization. |
PDF Full Text Request