基于决策树和logistic回归构建恙虫病合并脓毒症的预测模型

Constructing a predictive model for tsutsugamushi disease complicated with sepsis based on decision tree and logistic regression

目的通过决策树模型和logistic回归模型探索恙虫病合并脓毒症的最佳预测模型。方法选取2012年6月—2023年12月昆明市第三人民医院收治的恙虫病患者235例作为研究对象,其中138例恙虫病合并脓毒症患者作为实验组,97例未合并脓毒症的恙虫病患者作为对照组,建立决策树模型和logistic回归模型,利用受试者工作(receiver operating characteristic,ROC)曲线进行2种预测模型的区分度比较,判断其预测能力。结果235例患者中,汉族占比91.1%,农民占比66.8%,居住地在郊县的占比57.4%,临床表现以发热、头痛、畏寒为主,焦痂发生率为100.0%,2组间一般资料比较差异有统计学意义(P<0.05)的因素包括年龄、白介素(interleukin,IL)-1β(IL-1β)、IL-2、IL-17、干扰素(interferon,IFN)-γ(IFN-γ)、肿瘤坏死因子(tumor necrosis factor,TNF)-α(TNF-α)。将以上因素分别带入logistic回归模型和决策树模型,logistic回归的多因素分析结果显示:年龄(OR=1.039,95%CI:1.017~1.061)和IFN-γ(OR=1.009,95%CI:0.999~1.018)是恙虫病合并脓毒症的独立影响因素。ROC曲线下面积(area under curve,AUC)为0.669(95%CI:0.599~0.738,P<0.001),模型的准确度为63.8%,灵敏度为71.1%,特异度为57.2%。决策树模型结果显示,根节点为IL-1β,子节点为年龄和IL-2,AUC为0.714(95%CI:0.646~0.782,P<0.001),模型的准确度为68.9%,灵敏度为58.8%,特异度为76.1%。2种模型之间比较,AUC差值为0.045,标准误为0.042,Z=1.074,P=0.283,95%CI:-0.037~0.128。在治疗和预后中,使用多西环素的治疗频率最高(78.3%),联合用药中多西环素联合莫西沙星的频率最高(87.7%),联合用药与用药时间呈强正相关(τ=0.119,P=0.021),住院时间与用药时间呈强正相关(τ=0.725,P<0.001),临床诊断率为94.5%,死亡率为1.7%。结论2种模型均具有一定的预测价值,其中决策树模型预测能力更佳,值得临床推广。

Objective To explore the optimal prediction model for tsutsugamushi disease complicated with sepsis through decision tree models and logistic regression models.Methods A total of 235 cases of tsutsugamushi disease patients admitted to the Third People's Hospital of Kunming from June 2012 to December 2023 were selected as the research subjects,including 138 cases of tsutsugamushi disease patients with sepsis as the experimental group and 97 cases of tsutsugamushi disease patients without sepsis as the control group.Decision tree models and logistic regression models were established,and the discriminability of two prediction models was compared using receiver operating curve(ROC)to assess their predictive abilities.Results Among 235 patients,Han ethnicity accounted for 91.1%,farmers accounted for 66.8%,and those residing in suburban areas accounted for 57.4%.The clinical manifestations were mainly fever,headache,and chills,with an incidence rate of scab of 100.0%.Factors that were statistically significant between the two groups included age,IL-1β,IL-2,IL-17,IFN-γ,and TNF-α(P<0.05).The above factors were brought into the logistic regression model and decision tree model,respectively.The multivariate analysis results of logistic regression showed that age(OR=1.039,95%CI:1.017-1.061)and IFNγ(OR=1.009,95%CI:0.999-1.018)were independent influencing factors for tsutsugamushi disease complicated with sepsis.The area under the ROC curve(AUC)was 0.669(95CI:0.599-0.738,P<0.001),with an accuracy of 63.8%,sensitivity of 71.1%,and specificity of 57.2%.The decision tree model results showed that the root node was IL-1β,and the sub-nodes were age and IL-2.The area under the ROC curve(AUC)was 0.714(95CI:0.646-0.782,P<0.001),with an accuracy of 68.9%,sensitivity of 58.8%,and specificity of 76.1%.Comparing the two models,the AUC difference was 0.045,the standard error was 0.042,Z=1.074,P=0.283,95%CI:-0.037-0.128.In terms of treatment and prognosis,doxycycline was used most frequently(78.3%),and the combination of doxycycline and moxifloxacin was the most common(87.7%)in combined treatments.There was a strong positive correlation between the duration of medication and the frequency of combined treatments(τ=0.119,P=0.021),as well as between the duration of medication and hospitalization time(τ=0.725,P<0.001).The clinical diagnostic rate was 94.5%,and the mortality rate was 1.7%.Conclusions Both models have certain predictive values,among which the decision tree model has better predictive ability and is worthy of clinical promotion.