| Objective:To investigate the role of lower digestive tract angiography combined with machine learning in the diagnosis of neonatal Short-segment Hirschsprung’s disease(S-HSCR),so as to provide certain reference for related clinical studies and early diagnosis of neonatal S-HSCR.Method:The retrospective analysis method was used to collect patients hospitalized with symptoms such as abdominal distension,constipation,vomiting and intestinal obstruction in Children’s Hospital of Soochow University from July 2015 to October 2022 and Xuzhou Children’s Hospital from October 2017 to October 2022.In addition,patients aged less than 28 days at the time of lower digestive tract angiography were screened,classified and sorted out.Two radiologists read the lower gastrointestinal angiography images in a double-blind way to obtain clinical diagnosis,and labeled the Region of Interest(ROI)in the lower gastrointestinal angiography images.Morphological quantitative feature extraction of ROI was performed to screen out clinical signs that could be used for modeling.Combined with the results of clinical diagnosis conducted by radiologists by reading the lower digestive tract angiographic images and the quantitative features of imaging,Logistic Regression was applied.LR),sparse multivariate logistic regression(Sparse Multinomial Logistic Regression SMLR),support vector machine(Support Vector Machines),SVM and feature-weighted support vector machine(FWSVM)four statistical machine learning algorithms are used to construct the S-HSCR prediction model.Receiver Operating characteristic Curve(Receiver Operating characteristic Curve,ROC Curve),Area under the ROC curve(Area Under the receiver operating characteristic Curve,AUC)and accuracy(Accuracy,ACC),sensitivity(Sensitivity,SEN)and specificity(SPE)as an evaluation index to evaluate the value of machine learning for the early diagnosis of S-HSCR.Results:A total of 298 clinical data were collected in this study,including 145 cases in the S-HSCR group and 153 cases in the non-S-HSCR group.By reading the lower digestive tract angiographic images,radiologists can divide the diagnosis results into S-HSCR confirmation,S-HSCR exclusion and S-HSCR uncertainty.After statistical analysis,the accuracy rate,sensitivity=63.45%and specificity=81.70%were obtained from radiologists in Children’s Hospital of Soochow University and Xuzhou Children’s Hospital.The diagnostic results of radiologists in the two hospitals were statistically analyzed,and the evaluation results were as follows:accuracy=74.60%,sensitivity=56.72%,specificity=94.92%in the Children’s Hospital of Soochow University;Accuracy=71.51%,sensitivity=69.23%,specificity=73.40%in Xuzhou Children’s Hospital.Univariate analysis showed that 24-hour meconium excretion(P<0.001),vomiting(P=0.048),abdominal distension(P<0.001)and anal digital examination(P<0.001)were statistically significant between the two groups.These four clinical features were used to assist the construction of diagnostic model.At the same time,the morphological characteristics and the clinical diagnosis results obtained by the radiologist through reading the lower digestive tract contrast images were used as the basis of modeling.Four prediction models of LR,SMLR,SVM and FWSVM were constructed based on clinical characteristics,morphological characteristics and diagnostic results of radiologists.The results were as follows:(LR):area under ROC curve=0.9280,accuracy=87.08%,sensitivity=87.85%,specificity=86.24%;(SMLR):Area under ROC curve=0.9271,accuracy=87.25%,sensitivity=87.38%,specificity=87.12%;(SVM):Area under ROC curve=0.9130,accuracy=85.94%,sensitivity=86.36%,specificity=85.56%;(FWSVM):ROC curve area=0.9056,accuracy=85.57%,sensitivity=86.03%,specificity=85.19%.However,after modeling without the diagnosis of radiologists,all values decreased significantly.Each of the four machine learning algorithms predicted better results than the radiologist.In order to further confirm the reliability of the four machine learning algorithms,four machine learning algorithms were applied to the data of Children’s Hospital Affiliated to Soochow University and Xuzhou Children’s Hospital respectively for prediction analysis based on clinical characteristics,morphological characteristics and radiologist diagnosis results.The prediction results obtained by the Children’s Hospital of Soochow University through the four machine learning algorithms were:(LR):Area under ROC curve=0.9549,accuracy=92.06%,sensitivity=91.36%,specificity=92.83%;(SMLR):Area under ROC curve=0.9688,accuracy=92.31%,sensitivity=88.98%,specificity=96.09%;(SVM):Area under ROC curve=0.9800,accuracy=93.42%,sensitivity=92.43%,specificity=94.53%;(FWSVM):ROC curve area=0.9452,accuracy=92.54%,sensitivity=92.66%,specificity=92.88%.The prediction results obtained by Xuzhou Children’s Hospital through four machine learning algorithms are as follows:(LR):area under ROC curve=0.9025,accuracy=82.20%,sensitivity=82.66%,specificity=81.59%;(SMLR):Area under ROC curve=0.8972,accuracy=81.45%,sensitivity=84.47%,specificity=78.94%;(SVM):Area under ROC curve =0.8907,accuracy=83.20%,sensitivity=84.45%,specificity=80.55%;(FWSVM):ROC area=0.8934,accuracy=83.07%,sensitivity=80.88%,specificity=84.62%.The prediction results of four machine learning algorithms established by the two hospitals based on clinical characteristics,morphological characteristics and the diagnosis of radiologists were all higher than the diagnosis of radiologists.Conclusions:1、The prediction of neonatal S-HSCR diagnosis was established based on clinical features,morphological features and diagnosis by radiologists using lower digestive tract angiography combined with four machine learning algorithms.The results were all higher than that of radiologists,and the combination of lower digestive tract angiography and machine learning could improve the diagnosis rate of neonatal S-HSCR.2、The combination of lower digestive tract angiography and machine learning can realize the early diagnosis of S-HSCR in neonate,which lays the foundation for early intervention and treatment of infants. |
PDF Full Text Request