Construction Of Prediction Model Of Necrotizing Enterocolitis In Preterm Infants Based On Machine Learning Algorithms

Objective: Necrotizing enterocolitis（NEC）is the most common and destructive intestinal inflammatory disease in preterm infants.In this study,we retrospectively analyzed the risk factors of NEC in preterm infants prior and at the onset of the disease,applied machine learning methods to the diagnosis of NEC,and compared the predictive performance of traditional logistic regression models and machine learning models.Methods:（1）In this study,564 preterm infants（264 NEC group and 300 non-NEC Group）with suspected NEC symptoms were selected from January 2015 to October 2021 in the Department of Neonatology,the First Hospital of Jilin University.Perinatal,clinical,and laboratory parameters were collected up to the onset of the disease.We analyzed and compared the variables between NEC group and non-NEC group,and the risk factors associated with the diagnosis of NEC were screened out via Logistic regression analysis.Then we established a risk regression equation to evaluate the accuracy of the prediction.（2）Recursive Feature Elimination（RFE）,Max-Relevance and Min-Redundancy（MRMR）and Elastic Net（EN）were used to select features from all the variables,and the results of the three methods were integrated to select the final variables.Support vector machine（SVM）,multilayer perceptron（MLP）and extreme gradient boosting（XGBoost）were used to establish the diagnostic prediction models of NEC and their performances were evaluated.The area under the receiver operating characteristic curve（AUC）,sensitivity,specificity,negative predictive value and positive predictive value of different models were compared.Results:（1）Multivariate Logistic regression analysis showed that independent risk factors associated with the diagnosis of NEC included: patent ductus arteriosus（PDA）[Odds Ratio（OR）: 1.802,95% Confidence Interval（CI）: 1.004-3.234,P = 0.048],gastric residual（OR:2.544,95%CI:1.417-4.567,P=0.002）,pneumatosis intestinalis（OR:4.026,95%CI:2.122-7.639）,P<0.001,portal venous gas（OR:7.115,95%CI:3.606-14.041,P<0.001）,dilated bowel（OR:13.562,95%CI:6.378-28.838,P<0.001）,bowel wall thickening（OR:3.948,95%CI:2.017-7.725,P<0.001）,enteral nutrition start（OR:0.227,95% CI:0.123-0.418,P<0.001）,daily milk increment（OR:6.248,95%CI:3.282-11.895,P<0.001）,breast milk feeding（OR:0.215,95%CI:0.112-0.415,P<0.001）,Probiotics（OR:0.310,95%CI:0.166-0.578,P<0.001）monocyte percentage（MO%）at birth（OR:0.878,95%CI:0.816-0.944,P=0.001）,neutrophil percentage（NEUT%）at onset（OR:1.035,95%CI:1.018-1.051,P<0.001）、mean corpuscular volume（MCV）at onset（OR:0.957,95%CI:0.918-0.997,P=0.037）,white blood cell count（WBC）change（OR:0.619,95%CI:0.425-0.902,P=0.013）.（2）Five features with the highest importance rankings screened by the MRMR algorithm were: dilated bowel（0.124370）,lymphocyte（LY）count change（0.118769）,portal venous gas（0.099929）,MO% at birth（0.099673）,hematocrit（HCT）at onset（0.097583）;five features with the highest importance rankings screened by the RFE algorithm were: dilated bowel（0.101730）,portal venous gas（0.092147）,pneumatosis intestinalis（0.072376）,NEUT% at onset（0.072009）,type of milk（0.062553）;five features with the highest importance rankings screened by the EN algorithm were: dilated bowel（0.107048）、LY count change（0.090874）,portal venous gas（0.086775）,NEUT% at onset（0.084808）,MO% at birth（0.079484）.（3）The variables determined by three feature selection methods were integrated,and two strategies of intersection（F-I）and union（F-U）of three variable subsets were used to establish the NEC prediction model.Among the diagnostic prediction models based on F-I feature selection strategy,SVM performed best（AUC: 0.919,95% CI: 0.866-0.972;Accuracy 0.854;Sensitivity 0.847;Specificity 0.880;Positive predictive value 0.850;Negative predictive value0.873）.Among the diagnostic predictive models based on F-U strategy,MLP performed best（AUC: 0.933,95% CI: 0.883-0.983;Accuracy 0.867;Specificity 0.873;Positive predictive value 0.867）.Conclusion:（1）"dilated bowel,pneumatosis intestinalis,portal venous gas,bowel wall thickening,type of milk,daily milk increment,enteral nutrition start,gestational diabetes mellitus,placenta abnormalities,MO% at birth,HCT at onset,early onset sepsis,late onset sepsis,PDA,emesis,bowel sound attenuation,gastric residual,bloody stools,WBC at birth,hemoglobin at birth,red blood cell count（RBC）at birth,mean corpuscular hemoglobin（MCH）、MO count at birth,NEUT% at onset,RBC at onset,MCV at onset,HCT change,LY count change,RBC change,red blood cell distribution width change,MCH change" are effective indicators for predicting NEC diagnosis.（2）Abdominal imaging examination,feeding strategy（type of milk,daily milk increment,enteral nutrition start,probiotics,etc.）and blood routine parameters are valuable in the differential diagnosis of NEC in preterm infants.（3）Both the traditional logistic model and machine learning models performed well and stably in predicting NEC diagnosis,and the comprehensive diagnosis performance of SVM model combined with F-I feature selection strategy is the best.The classification prediction model based on machine learning algorithm has certain clinical practical value for the differential diagnosis of NEC.