创伤性脑损伤患者重症监护病房及院内死亡风险预测模型的构建与验证

Construction and validation of the predictive model for intensive care unit and in‑hospital mortalityrisk in patients with traumatic brain injury

目的构建创伤性脑损伤(TBI)患者重症监护病房(ICU)及院内死亡风险的预测模型并进行效能验证。方法采用回顾性队列研究分析截至2018年5月发布于eICU合作研究数据库v2.0(eICU⁃CRD v2.0)的3907例TBI患者的临床资料,其中男2397例,女1510例;年龄18~92岁[63.0(43.0,79.0)岁]。根据患者是否在ICU期间发生死亡分为ICU生存组(3575例)和ICU死亡组(332例),以及是否发生院内死亡分为院内生存组(3413例)和院内死亡组(494例)。提取患者一般资料、入院诊断、实验室检查、救治措施及临床预后变量。对生存组与死亡组进行单因素分析和多因素Logistic回归分析,筛选影响TBI患者ICU及院内死亡的独立危险因素,进而构建Logistic回归预测模型并以列线图呈现。将提取的数据资料按7∶3随机分为训练集(2735例)及验证集(1172例),对预测模型进行内部检验,同时提取MIMIC⁃Ⅲv1.4数据库中TBI患者数据对预测模型进行外部验证。采用受试者工作特征(ROC)曲线及曲线下面积(AUC)评估模型的区分度。采用Hosmer⁃Lemeshow(H⁃L)拟合优度检验及校准曲线评价模型的校准度。结果将单因素分析结果有统计学意义的变量纳入ICU死亡风险及院内死亡风险多因素Logistic回归分析中,结果表明,急性生理学与慢性健康状况评估Ⅳ(APACHEⅣ)评分(OR=1.04,95%CI 1.03,1.04,P<0.01)、格拉斯哥昏迷评分(GCS)(OR=0.66,95%CI 0.59,0.73,P<0.01)、合并脑疝(OR=6.91,95%CI 3.13,15.26,P<0.01)、国际标准化比值(INR)(OR=1.33,95%CI 1.09,1.62,P<0.01)、高渗盐水应用(OR=0.45,95%CI 0.21,0.94,P<0.05)、血管活性药物使用(OR=2.19,95%CI 1.36,3.52,P<0.01)是TBI患者ICU死亡的独立危险因素;年龄(每10岁为1个等级)(OR=1.28,95%CI 1.17,1.40,P<0.01)、APACHEⅣ评分(OR=1.03,95%CI 1.02,1.04,P<0.01)、GCS(OR=0.75,95%CI 0.71,0.80,P<0.01)、合并脑疝(OR=6.44,95%CI 2.99,13.86,P<0.01)、血肌酐水平(OR=1.07,95%CI 1.01,1.15,P<0.05)、INR(OR=1.49,95%CI 1.20,1.85,P<0.01)、高渗盐水应用(OR=0.41,95%CI 0.21,0.80,P<0.01)、血管活性药物使用(OR=2.27,95%CI 1.46,3.53,P<0.01)是TBI患者院内死亡的独立危险因素。依据上述ICU死亡的独立危险因素构建预测模型方程:Logit P(ICU)=7.12+0.03דAPACHEⅣ评分”-0.42דGCS”+1.93ד合并脑疝”+0.28דINR”-0.81ד高渗盐水应用”+0.79ד血管活性药物使用”;依据上述院内死亡的独立危险因素构建模型方程:Logit P(院内)=2.75+0.25ד年龄(每10岁为1个等级)”+0.03דAPACHEⅣ评分”-0.28דGCS”+1.86ד合并脑疝”+0.07ד血肌酐水平”+0.40דINR”-0.90ד高渗盐水应用”+0.82ד血管活性药物使用”。ICU死亡风险预测模型中,训练集的AUC为0.95(95%CI 0.94,0.97);验证集的AUC为0.91(95%CI 0.87,0.95)。训练集H⁃L拟合优度检验结果为P=0.495,校准曲线平均绝对误差为0.003;验证集H⁃L拟合优度检验结果为P=0.650,校准曲线平均绝对误差为0.012。院内死亡风险预测模型中,训练集的AUC为0.91(95%CI 0.89,0.93);验证集的AUC为0.91(95%CI 0.88,0.94)。训练集H⁃L拟合优度检验结果为P=0.670,校准曲线平均绝对误差为0.006;验证集H⁃L拟合优度检验结果为P=0.080,校准曲线平均绝对误差为0.021。在ICU死亡风险预测外部验证中,预测模型AUC为0.88(95%CI 0.86,0.90);H⁃L拟合优度检验结果为P=0.205,校准曲线绝对误差为0.031。在院内死亡风险预测外部验证集中,预测模型AUC为0.88(95%CI 0.85,0.91);H⁃L拟合优度检验结果为P=0.239,校准曲线绝对误差为0.036。模型的内部验证及外部验证显示,ICU及院内死亡风险预测模型均具有良好的区分度及校准度。结论由APACHEⅣ评分、GCS、合并脑疝、高渗盐水应用、血管活性药物使用、INR构建的ICU死亡风险预测模型及由年龄、APACHEⅣ评分、GCS、合并脑疝、血肌酐水平、高渗盐水应用、血管活性药物使用、INR构建的院内死亡风险预测模型均能够较好地预测TBI患者死亡风险。

Objective To construct a predictive model for intensive care unit(ICU)and in⁃hospital mortality risk in patients with traumatic brain injury(TBI)and validate its performance.Methods A retrospective cohort study was conducted to analyze the clinical data of 3907 patients with TBI published until May 2018 in the eICU Collaborative Research Database v2.0(eICU⁃CRD v2.0),including 2397 males and 1510 females,aged 18⁃92 years[63.0(43.0,79.0)years].According to whether the patients died in ICU or at hospital stay,they were divided into ICU survival group(n=3575)and ICU mortality group(n=332),and hospital survival group(n=3413)and hospital mortality group(n=494).The general data,admission diagnosis,laboratory tests,therapeutic interventions,and clinical outcomes were extracted as variables of interest.Univariate analysis and multivariate Logistic regression analysis were conducted on both the survival groups and the mortality groups to identify the independent risk factors that affect ICU and in⁃hospital mortality in TBI patients,based on which a Logistic regression prediction model was constructed and represented by Nomograms.The extracted dataset was randomly divided into training set(n=2735)and validation set(n=1172)with a ratio of 7∶3,and was applied for internal validation of the of the predictive model.Meanwhile,the data of TBI patients in the MIMIC-III v1.4 database were extracted for external validation of the predictive model.The area under the curve(AUC)of the receiver operating characteristic(ROC)curve was used for discriminability evaluation of the model,and the Hosmer-Lemeshow(H-L)goodness of fit test and calibration curve were used for calibration evaluation of the model.Results The statistically significant variables identified in the univariate analysis were included in the multivariate logistic regression analysis of ICU mortality and in⁃hospital mortality risk.The results revealed that acute physiology and chronic health evaluation IV(APACHE IV)score(OR=1.04,95%CI 1.03,1.04,P<0.01),Glasgow coma scale(GCS)(OR=0.66,95%CI 0.59,0.73,P<0.01),cerebral hernia formation(OR=6.91,95%CI 3.13,15.26,P<0.01),international normalized ratio(INR)(OR=1.33,95%CI 1.09,1.62,P<0.01),use of hypertonic saline(OR=0.45,95%CI 0.210.94,P<0.05),and use of vasoactive agents(OR=2.19,95%CI 1.36,3.52,P<0.01)were independent risk factors for ICU mortality in TBI patients.The age(with 10 years as a grade)(OR=1.28,95%CI 1.17,1.40,P<0.01),APACHE IV score(OR=1.03,95%CI 1.02,1.04,P<0.01),GCS(OR=0.75,95%CI 0.71,0.80,P<0.01),cerebral hernia formation(OR=6.44,95%CI 2.99,13.86,P<0.01),serum creatinine level(OR=1.07,95%CI 1.01,1.15,P<0.05),INR(OR=1.49,95%CI 1.20,1.85,P<0.01),use of hypertonic saline(OR=0.41,95%CI 0.21,0.80,P<0.01),and use of vasoactive agents(OR=2.27,95%CI 1.46,3.53,P<0.01)were independent risk factors of in⁃hospital mortality of TBI patients.Based on the forementioned independent risk factors for ICU mortality,the model equation was constructed:Logit P(ICU)=7.12+0.03×"APACHE IV score"-0.42×"GCS"+1.93×"cerebral hernia formation"+0.28×"INR"-0.81×"use of hypertonic saline"+0.79×"use of vasoactive agents".Based on the forementioned independent risk factors for in⁃hospital mortality,the model equation was constructed:Logit P(in⁃hospital)=2.75+0.25×"age"(with 10 years as a grade)+0.03×"APACHE IV score"-0.28×"GCS"+1.86×"cerebral hernia formation"+0.07×"serum creatinine level"+0.40×"INR"-0.90×"use of hypertonic saline"+0.82×"use of vasoactive agents".In the prediction model for ICU mortality,the AUC of the training set and validation set was 0.95(95%CI 0.94,0.97)and 0.91(95%CI 0.87,0.95).The result of H⁃L goodness of fit test of the training set was P=0.495 with the average absolute error in the calibration curve of 0.003,while the result of H⁃L goodness of fit test of the validation set was P=0.650 with the average absolute error in the calibration curve of 0.012.In the prediction model for in⁃hospital mortality,the AUC of the training set and validation set was 0.91(95%CI 0.89,0.93)and 0.91(95%CI 0.88,0.94).The result of H⁃L goodness of fit test of the training set was P=0.670 with the average absolute error in the calibration curve of 0.006,while the result of H⁃L goodness of fit test of the validation set was P=0.080 with the average absolute error in the calibration curve of 0.021.In the external validation set of ICU mortality risk,the AUC of the prediction model was 0.88(95%CI 0.86,0.90),while the result of H⁃L goodness of fit test was P=0.205 with the average absolute error in the calibration curve of 0.031.In the external validation set of in⁃hospital mortality risk,the AUC of the prediction model was 0.88(95%CI 0.85,0.91),while the result of H⁃L goodness of fit test was P=0.239 with the average absolute error in the calibration curve of 0.036.The internal and external validation of the model indicated that both the prediction models for ICU and in⁃hospital mortality had good discriminability and calibration.Conclusion The ICU mortality prediction model constructed by APACHE IV score,GCS,cerebral hernia formation,use of hypertonic saline,vasoactive agents use of and INR,and the in⁃hospital mortality prediction model constructed by age grading,APACHE IV score,GCS,cerebral hernia formation,serum creatinine level,hypertonic saline use of,use of vasoactive agents and INR can predict the mortality risk of TBI patients well.