基于机器学习算法的小儿重症病毒性脑炎预后预测模型构建

Establishment of prognosis prediction model for pediatric severe viral encephalitis based on machine learning algorithm

目的 应用机器学习算法构建小儿重症病毒性脑炎(SVE)预后预测模型,与经典Logistic多因素回归预测模型进行对比分析,比较不同算法模型的预测效果。方法 选取2018年5月-2022年5月浙江大学医学院附属第二医院临平院区收治的210例SVE患儿,均接受临床治疗。治疗结束后对患儿进行为期6个月的随访,根据格拉斯哥预后评分(GOS)将患儿分为预后良好组、预后不良组,收集患儿临床资料,单因素分析可能影响SVE患儿预后的危险因素,筛选影响SVE患儿预后的危险因素构建经典Logistic回归预测模型,并运用R软件构建随机森林预测模型、支持向量机(SVM)预测模型,绘制3种预测模型受试者工作特征(ROC)曲线,对比各预测模型ROC灵敏度、特异度、准确度、曲线下面积等指标。结果 210例患儿均未失访,且无死亡病例,其中预后不良组50例、预后良好组160例。Logistic多因素回归分析结果显示,惊厥持续状态、局灶性神经功能缺损、脑电图中重度异常、合并应激性高糖、脑脊液蛋白高是影响SVE患儿不良预后的危险因素(P<0.05)。基于多因素回归分析筛选出的5个危险因素分别构建Logistic回归预测模型、随机森林预测模型、SVM预测模型,结果显示3种预测模型的敏感度、准确度、阳性预测值、阴性预测值均>80%,其中SVM预测模型拥有最高的敏感度、准确度、曲线下面积(AUC)、阳性预测值,分别为93.00%、86.10%、0.913、87.90%,均高于Logistic预测模型的88.90%、82.70%、0.891、84.60%,随机森林预测模型的90.20%、83.20%、0.862、85.70%。结论 机器学习算法中SVM预测模型可有效预测SVE患儿不良预后,影响SVE患儿预后的危险因素较多,临床可基于上述危险因素给予针对性预防措施以改善SVE患儿预后。

OBJECTIVE To apply machine learning algorithm to construct a prognostic prediction model for paediatric severe viral encephalitis(SVE),and to compare the predictive effect of different algorithm models with the classical Logistic multivariate regression prediction model.METHODS A total of 210 children with SVE admitted to the Linping Hospital of the Second Affiliated Hospital of Zhejiang University College of Medicine from May 2018 to May 2022 were selected,and all of them received clinical treatment.After treatment,the children were followed up for 6 months and divided into good prognosis group and poor prognosis group according to the Glasgow outcome score(GOS).Clinical data of the children were collected,univariate analysis was used to identify the risk factors that may affect the prognosis of children with SVE,the risk factors that affect the prognosis of children with SVE were screened to construct the classical Logistic regression prediction model,and the random forest prediction model and support vector machine(SVM) prediction model were constructed using R software.The receiver operating characteristic(ROC) curves of the three prediction models were drawn,and the sensitivity,specificity,accuracy,area under the curve and other indicators were compared for each prediction model.RESULTS None of the 210 children were lost to follow-up,and there were no deaths,including 50 cases in the poor prognosis group and 160 cases in the good prognosis group.Logistic regression analysis showed that the presence of persistent convulsions,focal neurological deficits,moderate and severe EEG abnormalities,combined with stress hyperglycemia and high CSF protein were independent risk factors for poor prognosis of SVE children(P<0.05).Based on the five risk factors selected by multivariate regression analysis showed that the presence of persistent convulsions,focal neurological deficits,moderate to severe EEG abnormalities,combined stress hyperglycaemia and high cerebralspinal fluid protein were risk factors for poor prognosis in children with SVE(P<0.05).Logistic regression prediction model,random forest prediction model and SVM prediction model were constructed based on the five risk factors screened by multivariate regression analysis,and the results showed that the sensitivity,accuracy,positive predictive value and negative predictive value of the three prediction models were all greater than 80%,among which the SVM prediction model had the highest sensitivity,accuracy,AUC and positive predictive value,which were 93.00%,86.10%,0.913 and 87.90% respectively,all of them were higher than 88.90%,82.70%,0.891,84.60% of Logistic model and 90.20%,83.20%,0.862,85.70% of random forest model.CONCLUSION The support vector machine(SVM) prediction model in machine learning algorithm could effectively predict the poor prognosis of children with SVE,and there were various risk factors affecting the prognosis of children with SVE.Targeted preventive measures should be taken based on the above risk factors to improve the prognosis of children with SVE.