Logistic回归和机器学习模型预测胸腔镜肺部分切除术患者单肺通气期间低SpO_(2)的比较  

Comparison of logistic regression and machine learning models predicting low SpO_(2)during one-lung ventilation in patients undergoing thoracoscopic partial pulmonary resection

作　　者：许斯洋 王君 渠磊秋 桂波 阮姗 XU Siyang;WANG Jun;QU Leiqiu;GUI Bo;RUAN Shan(Department of Anesthesiology and Pain,Geriatric Hospital of Nanjing Medical University,Nanjing 210024,China)

机构地区：[1]南京医科大学附属老年医院麻醉疼痛科,210024

出　　处：《临床麻醉学杂志》2024年第10期1022-1028,共7页Journal of Clinical Anesthesiology

摘　　要：目的比较Logistic回归和机器学习模型对胸腔镜肺部分切除术(TPPR)患者单肺通气(OLV)期间发生低SpO_(2)的预测效能,并探讨低SpO_(2)的危险因素。方法选择2022年8月1日至2023年4月30日行单侧TPPR患者127例,男61例,女66例,年龄18~80岁,ASAⅠ—Ⅲ级。根据术中OLV期间是否出现SpO_(2)<90%将患者分为两组:低SpO_(2)组(n=21)和正常SpO_(2)组(n=106)。收集患者围术期相关数据,采用Logistic回归构建预测模型,与采用随机森林(RF)、极限梯度提升(XGBoost)、决策树(DT)、逻辑回归(LogR)、支持向量机(SVM)5种机器学习模型构建的预测模型进行比较,绘制受试者工作特征(ROC)曲线,计算曲线下面积(AUC)评价预测模型的效能。采用沙普利加和解释法(SHAP)解释输出的最佳模型,确定TPPR患者OLV期间低SpO_(2)的危险因素。结果多因素Logistic回归分析显示,年龄增大(OR=1.087,95%CI 1.006~1.175,P=0.036)、BMI升高(OR=1.299,95%CI 1.050~1.608,P=0.016)、术前血糖浓度升高(OR=2.028,95%CI 1.378~2.983,P<0.001)、RV/TLC%_(Pred)降低(OR=0.936,95%CI 0.892~0.983,P=0.008)是OLV期间低SpO_(2)独立危险因素,预测模型为Logit(p)=-10.098+0.08×年龄+0.231×BMI+0.633×血糖-0.059×RV/TLC%_(Pred),该模型AUC为0.873(95%CI 0.803~0.943,P<0.001)。经过网格搜索与五折交叉验证结合优化机器学习模型参数,模型训练效果良好。ROC曲线分析结果显示,RF的AUC为0.921(95%CI 0.840~0.979),XGBoost的AUC为0.940(95%CI 0.812~0.981),DT的AUC为0.919(95%CI 0.828~0.982),LogR的AUC为0.892(95%CI 0.831~0.980),SVM的AUC为0.922(95%CI 0.832~0.982),XGBoost预测的AUC最高,且高于传统的Logistic回归预测模型。经SHAP方法处理后,XGBoost输出模型中最重要的危险因素是年龄增大、BMI和术前血糖浓度升高。结论年龄增大、BMI和术前血糖浓度升高是TPPR患者OLV期间低SpO_(2)的危险因素,机器学习模型XGBoost预测OLV期间低SpO_(2)发生的效能优于传统的Logistic回归,能分析变量与结局�Objective To compare the predictive effects of logistic regression and machine learning models on occurrence of low peripheral oxygen saturation(SpO_(2))during one-lung ventilation(OLV)in patients undergoing thoracoscopic partial pulmonary resection(TPPR),and to explore risk factors of low SpO_(2).Methods A total of 127 patients undergoing unilateral TPPR from August 1,2022 to April 30,2023 were enrolled,61 males and 66 females,aged 18-80 years,ASA physical statusⅠ-Ⅲ.Based on whether intraoperative SpO_(2)during OLV was less than 90%,the patients were divided into two groups:low SpO_(2)group(n=21)and normal SpO_(2)group(n=106).Perioperative data were collected and a predictive model was constructed using logistic regression.This model was compared with predictive models constructed using five machine learning models,including random forest(RF),extreme gradient boosting(XGBoost),decision tree(DT),logistic regression(LogR),and support vector machine(SVM).The receiver operating characteristic(ROC)curve was plotted,and the performance of the predictive models were evaluated by the area under the curve(AUC).The best output model was interpreted using Shapley additive explanations(SHAP)to identify the risk factors of low SpO_(2)during OLV in patients undergoing TPPR.Results Multivariate logistic regression analysis showed that increased age(OR=1.087,95%CI 1.006-1.175,P=0.036),increased BMI(OR=1.299,95%CI 1.050-1.608,P=0.016),increased preoperative blood glucose(OR=2.028,95%CI 1.378-2.983,P<0.001),and decreased RV/TLC%_(Pred)(OR=0.936,95%CI 0.892-0.983,P=0.008)were independent risk factors of low SpO_(2)during OLV.The predictive model was Logit(p)=-10.098+0.08×age+0.231×BMI+0.633×blood glucose-0.059×RV/TLC%_(Pred),with an AUC of 0.873(95%CI 0.803-0.943,P<0.001).After optimizing parameters of machine learning models using grid search combined with five-fold cross-validation,the model training results were satisfactory.ROC curve analysis showed that the AUC for RF was 0.921(95%CI 0.840-0.979),XGBoost was 0.940(95%

关 键 词：机器学习 胸腔镜肺部分切除术 单肺通气 年龄 体重指数 血糖 

分 类 号：R614[医药卫生—麻醉学]