| Background and objective: Congenital heart defect(CHD)is the most common birth defect,and its severity depends on the size and complexity of the deformity.Some children may have a series of irreversible complications due to delayed diagnosis and missed the best operation period,such as heart failure,severe pulmonary hypertension,Eisenmann’s syndrome,and eventually lead to death.Therefore,early diagnosis is of great significance for the treatment and prognosis of CHD.At present,CHD is mainly screened by heart auscultation and pulse oximetry(POX),but the effect depends on the level of doctors’ auscultation and the degree of hypoxia caused by CHD,which is easily missed.This study aims to explore new algorithms for screening CHD in children with the help of artificial intelligence(AI)technology and improve the screening detection rate of CHD in children.Methods:1.Data collection: Heart sounds of children admitted to the Department of Cardiology,Children ’s Hospital Affiliated to Chongqing Medical University from October 2019 to October 2020 were collected by electronic stethoscope.2.Model construction: Feature extraction combined with traditional machine learning method and classification algorithm based on deep learning are used to construct a binary classification model by comparing different types of CHD heart sounds with control heart sounds.3.Statistical analysis: The best classification model was selected and compared with the auscultation of cardiovascular physicians with more than 10 years of clinical experience.Chi-square test was performed on the count data,and P < 0.01 indicated that there was a significant statistical difference.The correlation between heart sounds and LVEF,pulmonary artery pressure and malformation size was analyzed.Pearson correlation coefficient range was-1 ～ + 1,and absolute value > 0.5 showed significant linear correlation.Results: A total of 930 heart sounds were collected,and 884 cases were included in the analysis,including 192 cases of atrial septal defect(ASD),98 cases of ventricular septal defect(VSD),95 cases of patent ductus arteriosus(PDA),90 cases of combined CHD and 409 cases of control group.The RCRnet classification model based on deep learning had the highest recognition accuracy and sensitivity among the four types of CHD.In the test set,the accuracy of ASD was 0.654,VSD was 1.000,PDA was 0.891,and combined CHD was 0.924.The best auscultation sites of ASD,VSD,PDA and combined CHD were shown as the 4th,1st,3rd and4 th auscultation areas,respectively.The sensitivity,specificity,precision and accuracy of this model for the diagnosis of four types of CHD were higher than those of expert auscultation,with the sensitivity of 0.932 –1.000,specificity of 0.944 – 0.997,precision of 0.888 – 0.997 and accuracy of 0.940 – 0.994.Pearson correlation coefficient absolute value < 0.5suggests that there is no linear correlation between the four types of CHD heart sounds and LVEF,pulmonary artery pressure,malformation size.Conclusion:1.The RCRnet model based on deep learning has the best classification performance for ASD,VSD,PDA and combined CHD,and the best auscultation sites are shown as the 4th,1st,3rd and 4th auscultation areas,respectively.2.The new model can preliminarily determine the type of CHD in children,providing a new choice for the algorithm of AI-assisted auscultation screening for CHD in children.3.The diagnostic effect of the new model on four types of CHD is better than that of cardiovascular physicians with more than 10 years of clinical experience.4.No linear correlation was found between the four types of CHD heart sounds and important indicators such as LVEF,pulmonary artery pressure and malformation size. |
PDF Full Text Request