全身麻醉下肿瘤细胞减灭术联合腹腔热灌注化疗术患者术后肺部并发症的随机森林预测模型

A random forest prediction model for postoperative pulmonary complications in patients undergoing tumor cell cytoreduc⁃tive surgery combined with intraperitoneal hyperthermic chemoperfusion under general anesthesia

目的分析行肿瘤细胞减灭术(CRS)联合腹腔热灌注化疗术(HIPEC)的患者术后肺部并发症(PPC)的危险因素,并构建预测模型。方法收集行CRS+HIPEC的298例患者围手术期信息[性别、年龄、美国麻醉医师协会(ASA)分级、手术时间、术中总入量、术中总出量、出血量、尿量、胶体液输注量、晶体液输注量、自体血回输量、红细胞输注量、血浆输注量、围手术期进行目标导向液体治疗(GDFT)时参考的每搏变异度(SVV)值]。根据患者术后有无PPC,将患者分为PPC组(106例)和非PPC组(192例)。采用逐步回归分析筛选PPC的特征变量并建立随机森林预测模型,计算随机森林预测模型的袋外误差率,分别在训练集和测试集上计算混淆矩阵及参数(包括准确度、Kappa值、灵敏度、特异度、精准度、召回率、F1⁃Score);绘制受试者操作特征曲线(ROC曲线)[并计算曲线下面积(AUC)及95%置信区间(CI)]、校准曲线,绘制自变量排序图和各特征变量的偏依赖图。结果与非PPC组比较,PPC组的手术时间较长(P<0.05),术中总入量、术中总出量、出血量、胶体液输注量、尿量和红细胞输注量均较多(均P<0.05),围手术期进行GDFT时参考的SVV值较低,差异有统计学意义(P<0.05)。逐步回归分析显示手术时间、出血量、红细胞输注量和围手术期进行GDFT时参考的SVV值为PPC的特征变量(P<0.05)。随机森林预测模型的袋外误差率为1.400%。训练集准确度1.000,测试集准确度0.952,说明模型整体预测准确性高。Kappa值训练集1.000,测试集为0.894,说明模型整体预测能力的一致性高。训练集的灵敏度为1.000,特异度为1.000,测试集的灵敏度为0.871,特异度为1.000,说明模型的整体区分度较好。训练集的精准度为1.000,召回率为1.000,F1⁃Score为1.000,测试集的精准度为1.000,召回率为0.871,F1⁃Score为0.931,说明模型对于阳性结果的预测能力高。训练集ROC曲线的AUC为1.000(95%CI 1.000~1.000),测试集ROC曲线的AUC为0.997(95%CI 0.962~1.000),表明该预测模型具有较好的判别能力。从自变量排序图可以看出特征变量对PPC的贡献程度:特征变量对PPC影响大小排序为围手术期进行GDFT时参考的SVV值>手术时间>出血量>红细胞输注量。从偏依赖图可以看出每个特征变量对PPC的影响及PPC随特征变量的变化趋势:PPC基本随手术时间的增加而波动上升;当出血量<1000 ml时,PPC波动改变,上升不明显,当术中出血量>1000 ml时,PPC概率明显上升;红细胞输注量>1000 ml时,PPC上升明显;围手术期进行GDFT时参考的SVV值与PPC变化呈负相关。结论影响PPC的特征变量有手术时间、出血量、红细胞输注量和围手术期进行GDFT时参考的SVV值。构建的随机森林预测模型具有良好的区分度与准确度,能很好地运用于进行CRS+HIPEC患者PPC的预测。

Objective To analyze the risk factors for postoperative pulmonary complications(PPC)in patients undergoing cy⁃toreductive surgery(CRS)combined with hyperthermic intraperitoneal chemotherapy(HIPEC)and to construct a prediction model.Methods Collect perioperative information of 298 patients undergoing CRS+HIPEC[including gender,age,American Society of Anes⁃thesiologists(ASA)classification,operation duration,total intraoperative infusion volume,total intraoperative output volume,blood loss,urine volume,colloid infusion volume,crystalloid infusion volume,autologous blood transfusion volume,red blood cell transfusion vol⁃ume,plasma transfusion volume,and stroke volume variation(SVV)value referred to during goal⁃directed fluid therapy(GDFT)during the perioperative period].According to the presence or absence of PPC after surgery,patients were divided into PPC group(106 cases)and non⁃PPC group(192 cases).Stepwise regression analysis was used to screen the characteristic variables of PPC and establish a ran⁃dom forest prediction model.The out⁃of⁃bag error rate of the random forest prediction model was calculated,and the confusion matrix and parameters(including accuracy,Kappa value,sensitivity,specificity,precision,recall,F1⁃Score)were calculated on the training set and test set,respectively.Receiver operating characteristic(ROC)curves were plotted[and the area under the curve(AUC)and 95%confi⁃dence interval(CI)were calculated],calibration curves were drawn,and the ranking of independent variables and partial dependence plots of each characteristic variable were plotted.Results Compared with the non⁃PPC group,the PPC group had a longer operation du⁃ration(P<0.05),increased total intraoperative infusion volume,total intraoperative output volume,blood loss,colloid infusion volume,urine volume,and red blood cell transfusion volume(all P<0.05),and a decreased SVV value during the perioperative period,with statistically sig⁃nificant differences(P<0.05).The stepwise regression analysis showed that operation time,blood loss,red blood cell transfusion volume,and GDFT used the SVV value as the characteristic variable of PPC(P<0.05).The out⁃of⁃bag error rate of the random forest prediction model was 1.400%.The accuracy of the training set was 1.000,while the accuracy of the test set was 0.952,indicating high overall prediction accuracy of the model.The Kappa value of the training set was 1.000,and the Kappa value of the test set was 0.894,indicating high consistency in the overall prediction ability of the model.The sensitivity and specificity of the training set were both 1.000,while the sensitivity and specificity of the test set were 0.871 and 1.000,respectively,indicating good overall discrimination of the model.The degree of accuracy,recall rate,and F1⁃Score of the training set were all 1.000,while the degree of accuracy,recall rate,and F1⁃Score of the test set were 1.000,0.871,and 0.931,respectively,indicating a high predictive ability of the model for positive results.The AUC of the ROC curve for the training set was 1.000(95%CI 1.000,1.000),and the AUC for the test set was 0.997(95%CI 0.962,1.000),indicating good discrimination ability of the pre⁃diction model.From the plot of the sorted independent variables,we can observe the contribution degree of the feature variables to PPC:the order of the impact of feature variables on PPC when SVV value used as a reference is GDFT>operation time>blood loss>red blood cell trans⁃fusion volume.The partial dependence plot can show the impact of each feature variable on PPC and the trend of PPC with the change of the feature variables:PPC fluctuates and rises with the increase of operation time;when the blood loss is less than 1000 ml,the fluctuation of PPC changes and the rise is not obvious,but when the intraoperative blood loss is greater than 1000 ml,the probability of PPC rises signifi⁃cantly;when the red blood cell transfusion volume is greater than 1000 ml,PPC rises significantly;there is a negative correlation between the SVV value used as a reference for GDFT value and PPC changes.Conclusions The characteristic variables affecting PPC include operation time,blood loss,and transfusion volume of suspended red blood cells when SVV value was used as a reference for GDFT.The constructed random forest prediction model has good discrimination,and accuracy and can be effectively applied to predicting PPC in patients undergoing CRS+HIPEC.