结直肠癌患者根治术后切口感染发生风险的预测模型研究  

Research on risk prediction model of postoperative wound infection in patients with colorectal cancer undergoing radical resection surgery

作　　者：梁舒婷 盛崴宣[1] 高丹阳 宋明雪 缪慧慧 李天佐[1] Liang Shuting;Sheng Weixuan;Gao Danyang;Song Mingxue;Miao Huihui;Li Tianzuo(Department of Anesthesiology,Beijing Shijitan Hospital,Capital Medical University,Beijing 100038,China)

机构地区：[1]首都医科大学附属北京世纪坛医院麻醉科,100038

出　　处：《北京医学》2025年第1期42-50,共9页Beijing Medical Journal

摘　　要：目的使用机器学习算法建立结直肠癌患者根治术后切口感染(postoperative wound infection,PWI)发生风险的预测模型。方法选取2021年1月至2022年12月首都医科大学附属北京世纪坛医院择期行全身麻醉下结直肠癌根治术患者317例,收集患者一般资料、骨骼肌参数相关临床资料,采用Boruta算法筛选PWI发生风险的特征变量。将患者分为训练集(n=227例)和测试集(n=90例),在训练集上使用重复交叉验证、超参数优化建立PWI发生风险的预测模型,包括逻辑回归(logistic regression,LR)、极度梯度提升树(extreme gradient boosting,XGBoost)和随机森林(random forest,RF)等8种预测模型,选择最佳模型绘制特征变量排序图和单变量偏依赖图。在测试集上计算预测模型的混淆矩阵参数,分别使用ROC曲线和精度召回率曲线(precision recall curve,PRC)评价预测模型的区分能力和校准能力。结果317例患者中男112例、女205例;年龄31~91岁,平均(64.8±10.8)岁。24例(7.57%)发生PWI。Boruta筛选纳入的特征变量按重要性排序依次是骨骼肌放射密度(skeletal muscle radiodensity,SMD)、手术时间、预后营养指数(prognostic nu‐tritional index,PNI)、术前静脉血栓栓塞症(venous thromboembolism,VTE)评分、HU均值(HU average calculation,HUAC)、年龄、第三腰椎骨骼肌指数(L3-skeletal muscle index,L3-SMI)和年龄校正Charlson合并症指数(age-adjusted Charlson comorbidity index,aCCI)。SMD<42 HU、手术时间>6 h、PNI<45、VTE>4分、HUAC<30 HU、年龄>70岁和aCCI>8的患者易发生PWI,HUAC>42.5 HU和L3-SMI>55 cm^(2)/m^(2)的患者发生PWI的风险降低。RF预测模型的性能最好,准确度为0.9667,马修斯相关系数(Matthews correlation coefficient,MCC)为0.7765,ROC曲线的AUC为1.0000(95%CI:0.0998~1.0000,P<0.001),PRC的AUC为1.0000(95%CI:0.0997~1.0000,P<0.001)。结论基于机器学习和骨骼肌参数构建的结直肠患者根治术后PWI发生风险预测模型的性能较好。术前骨Objective To construct a risk prediction model of postoperative wound infection(PWI)in patients with colorectal cancer undergoing radical resection surgery by using machine learning algorithms.Methods A total of 317 patients with colorectal cancer undergoing radical resection surgery under general anesthesia in Beijing Shijitan Hospital,Capital Medical University from January 2021 to December 2022 were selected,and the general data and clinical data related to skeletal muscle parameters were collected.Boruta algorithm was used to screen the characteristic variables of risks of PWI.Patients were divided into training set(n=227 cases)and test set(n=90 cases).On the training set,the risk prediction model of PWI was established by repeated cross-validation and hyperparameter optimization,including eight prediction models,such as logistic regression(LR),extreme gradient boosting tree(XGBoost)and random forest(RF),and the best model was selected to draw the characteristic variable ranking diagram and the univariate partial dependence diagram.On the test set,the confusion matrix parameters of the prediction model were calculated,and the distinguishing ability and calibration ability of the prediction model were evaluated by ROC curve and precision recall curve(PRC),respectively.Results Among the 317 patients,there were 112 males,205 females,aged from 31 to 91 years,with an average age of(64.8±10.8)years.PWI occurred in 24 cases(7.57%).The filtered characteristic variables included in Boruta screening in order of importance were skeletal muscle radiodensity(SMD),operation time,prognostic nutritional index(PNI),preoperative venous thromboembolism(VTE)score,HU average calculation(HUAC),age,L3-skeletal muscle index(L3-SMI)and age-adjusted Charlson comorbidity index(aCCI).Patients with SMD<42 HU,operation time>6 h,PNI<45,VTE>4 points,HUAC<30 HU,age>70 years and aCCI>8 were more likeyly to occur PWI,while patients with HUAC>42.5 HU and L3-SMI>55 cm^(2)/m^(2) were less likely to occur PWI.The performance of RF prediction was

关 键 词：机器学习 术后切口感染 骨骼肌参数 混淆矩阵 偏依赖图 

分 类 号：R735.34[医药卫生—肿瘤]