基于机器学习的新生儿坏死性小肠结肠炎预测模型的建立

Construction of prediction model of neonatal necrotizing enterocolitis based on machine learning algorithms

目的探讨将机器学习方法应用于新生儿坏死性小肠结肠炎(necrotizing enterocolitis,NEC)诊断及严重程度分类的意义。方法回顾性分析2015年1月至2021年10月吉林大学第一医院新生儿科疑似NEC并行腹部影像学检查的新生儿临床资料,根据入选患儿(数据集1)是否符合改良Bell分期Ⅱ期以上分为NEC组和非NEC组,用于诊断预测分析;NEC组(数据集2)再根据是否为Bell分期Ⅲ期以上分为外科NEC组和内科NEC组,用于NEC严重程度分析。采用特征选择算法中的极端随机树、弹性网和递归特征消除法对全部变量进行筛选,通过逻辑回归、支持向量机(support vector machine,SVM)、随机森林和光梯度增强机等机器学习分类模型建立NEC诊断和严重程度预测模型。应用受试者工作特征曲线下面积(the area under the receiver operating characteristic curve,AUROC)、敏感度、特异度、阳性预测值和阴性预测值评估模型性能,选出最优模型。结果共纳入536例疑似NEC患儿。数据集1包括非NEC组302例、NEC组234例;数据集2包括内科NEC组164例、外科NEC组70例。极端随机树法筛出的变量在两个数据集中预测性能最佳,在NEC诊断分类模型中,SVM模型预测性能最佳,AUROC为0.932(95%CI 0.891~0.973),准确度0.844(95%CI 0.793~0.895),共确定11个预测变量,包括门静脉积气、发病时中性粒细胞百分比、肠腔扩张、发病时单核细胞计数等;在NEC严重程度预测模型中,SVM模型预测性能最佳,AUROC为0.835(95%CI 0.737~0.933),准确度0.787(95%CI 0.703~0.871),共确定25个预测变量,包括发病时日龄、C反应蛋白、中性粒细胞计数等。结论利用机器学习中的特征选择算法和SVM分类模型建立的新生儿NEC预测模型有助于NEC的诊断和疾病严重程度分类。

Objective To construct prediction models of necrotizing enterocolitis(NEC)using machine learning(ML)methods.Methods From January 2015 to October 2021,neonates with suspected NEC symptoms receiving abdominal ultrasound examinations in our hospital were retrospectively analyzed.The neonates were assigned into NEC group(modified Bell's staging≥Ⅱ)and non-NEC group for diagnostic prediction analysis(dataset 1).The NEC group was subgrouped into surgical NEC group(staging≥Ⅲ)and conservative NEC group for severity analysis(dataset 2).Feature selection algorithms including extremely randomized trees,elastic net and recursive feature elimination were used to screen all variables.The diagnostic and severity prediction models for NEC were established using logistic regression,support vector machine(SVM),random forest,light gradient boosting machine and other ML methods.The performances of different models were evaluated using area under the receiver operating characteristic curve(AUC),sensitivity,specificity,negative predictive value and positive predictive value.Results A total of 536 neonates were enrolled,including 234 in the NEC group and 302 in the non-NEC group(dataset 1).70 were in the surgical NEC group and 164 in the conservative NEC group(dataset 2).The variables selected by extremely randomized trees showed the best predictive performance in two datasets.For diagnostic prediction models,the SVM model had the best predictive performance,with AUC of 0.932(95%CI 0.891-0.973)and accuracy of 0.844(95%CI 0.793-0.895).A total of 11 predictive variables were determined,including portal venous gas,intestinal dilation,neutrophil percentage and absolute monocyte count at the onset of illness.For NEC severity prediction models,the SVM model showed the best predictive performance,with AUC of 0.835(95%CI 0.737-0.933)and accuracy of 0.787(95%CI 0.703-0.871).A total of 25 predictive variables were identified,including age of onset,C-reactive protein and absolute neutrophil count at clincial onset.Conclusions NEC prediction model established using feature selection algorithm and SVM classification model in ML is helpful for the diagnosis of NEC and grading of disease severity.