摘要
目的采用不同机器学习方法构建非心脏手术老年患者术后谵妄(POD)的预测模型其性能。方法收集2022年4月至2024年4月行非心脏手术的905例老年患者围术期资料,包括人口学特征、既往合并症、术前认知功能评估、实验室检查结果、麻醉记录单等共102个变量。采用单因素分析初步筛选危险因素,将有统计学差异(P<0.05)的变量纳入最小绝对收缩与选择算子(LASSO)筛选特征变量,应用随机森林(RF)、支持向量机(SVM)、自适应增强算法(Adaboost)和神经网络(NN)4种机器学习方法构建POD预测模型,采用受试者工作特征曲线下面积(AUROC)、精确度-召回率(PR)曲线的平均精度(AP)、Brier评分等对模型进行综合评估,引入Shapley加性解释(SHAP)对最优机器学习模型进行可解释化分析。结果有155例(17%)患者发生POD,经LASSO回归分析后,确定10个特征变量用于构建机器学习模型。4种机器学习模型中,RF的AUROC最高为0.90(95%CI 0.86~0.93),AP为0.8,Brier评分为0.086。SHAP模型解释性分析显示,对POD贡献度最高的是手术时间。结论在应用4种机器学习方法构建的非心脏手术老年患者POD预测模型中,RF的预测效能最佳。
Objective To develop and compare predictive models for postoperative delirium(POD)in elderly non-cardiac surgery patients using machine learning-based methods.Methods A total of 905 elderly patients undergoing non-cardiac surgery were collected for perioperative data from April 2022 to April 2024.Demographic characteristics,past comorbidities,preoperative cognitive function assessments,laboratory test results,and anesthesia records among the 102 variables were collected.Univariate analysis was initially employed to screen for risk factors,and variables with significant statistical differences(P<0.05)were selected for further analysis using the least absolute shrinkage and selection operator(LASSO)method.Four machine learning methods including random forest(RF),support vector machine(SVM),adaptive boosting(Adaboost),and neural network(NN)were utilized to construct predictive models for POD.The models were comprehensively evaluated using the area under the receiver operating characteristic curve(AUROC),average precision(AP)of the precision-recall curve and Brier score.The Shapley additive explanations(SHAP)method was employed to interpret the optimal machine learning model.Results A total of 155 patients(17%)developed POD.After LASSO regression analysis,10 feature variables were identified and used to construct the machine learning models.Among the four machine learning models,RF had the highest AUROC of 0.90(95%CI 0.86-0.93),an AP of 0.8,and the lowest Brier score of 0.086.SHAP model interpretability analysis revealed that the duration of surgery contributed the most to POD.Conclusion Among the four machine learning methods used to construct the POD predictive models in non-cardiac surgery patients,RF demonstrats the best predictive performance.
作者
石金云
陈荣
李文媛
纪木火
李青
SHI Jinyun;CHEN Rong;LI Wenyuan;JI Muhuo;LI Qing(Department of Anesthesiology,Integrative Medicine Affiliated Hospital of Nanjing University of Traditional Chinese Medicine,Nanjing 210028,China)
出处
《临床麻醉学杂志》
北大核心
2025年第3期240-245,共6页
Journal of Clinical Anesthesiology
基金
国家自然科学基金(82172131)。
