基于机器学习方法对非心脏手术后急性肾损伤预测价值研究  

Predictive value of machine learning methods for acute kidney injury after non⁃cardiac surgery

作　　者：吴卓熙 陈勤 陈凤 任玉坤 王卓 李洪 WU Zhuo-xi;CHEN Qin;CHEN Feng;REN Yu-kun;WANG Zhuo;LI Hong(Department of Anesthesiology,The Second Affiliated Hos-pital of Army Military Medical University,Chongqing 400037,China)

机构地区：[1]陆军军医大学第二附属医院麻醉科,重庆400037

出　　处：《临床军医杂志》2023年第7期678-682,687,共6页Clinical Journal of Medical Officers

基　　金：重庆市技术创新与应用发展专项重点项目(CSTB2022TIAD⁃KPX079)。

摘　　要：目的探讨基于机器学习方法对非心脏手术后急性肾损伤(AKI)的预测价值。方法选取陆军军医大学第二附属医院自2014年6月至2022年9月收治的24611例接受非心脏手术患者为研究对象。采用随机数字表法将患者分为训练组(n=17227)与验证组(n=7384)。采用包括逻辑回归、随机森林、支持向量机、简单决策树、极端梯度提升和集成模型在内的机器学习方法训练模型。采用受试者工作特征(ROC)曲线及ROC曲线下面积(AUC)评价模型的整体性能。采用沙普利可加性特征解释法(SHAP)评估每个特征对预测结果的贡献程度。结果非心脏手术后AKI发生率为11.3%(2780/24611)。训练组与验证组患者各项一般资料比较,差异均无统计学意义(P>0.05)。集成模型预测非心脏手术后AKI的AUC为0.749,大于逻辑回归、随机森林、支持向量机、简单决策树、极端梯度提升单独预测。随机森林模型中贡献度较高的20个特征中,手术相关变量有8个,分别为麻醉时长、普通胸外科手术、骨科手术、手术中总失液量、急诊手术、手术中晶体液输注总量、手术中血液制品使用量、手术中胶体液输注总量和择期手术。结论采用机器学习方法开发的非心脏手术后AKI的预测模型可进一步加强对AKI的预防,从而优化和改善患者预后。Objective To investigate the predictive value of machine learning methods for acute kidney injury(AKI)after non⁃cardiac surgery.Methods A total of 24611 patients who underwent non⁃cardiac surgery from June 2014 to September 2022 in the Second Af⁃filiated Hospital of Army Military Medical University were selected as the study objects.Patients were divided into training group(n=17227)and verification group(n=7384)using random number table method.The model was trained using machine learning methods including logistic regression,random forests,support vector machines,simple decision trees,extreme gradient lifting,and integrated models.Receiver operator characteristic(ROC)curve and area under ROC curve(AUC)were used to evaluate the overall performance of the model.Shapley additive explanation(SHAP)was used to evaluate the contribution of each feature to the predicted results.Results The incidence of AKI after non⁃cardiac surgery was 11.3%(2780/24611).There was no significant difference in the general information between the training group and the verification group(P>0.05).The AUC of non⁃cardiac surgery AKI predic⁃ted by the integrated model was 0.749,which was larger than that predicted by logistic regression,random forest,support vector ma⁃chine,simple decision tree,and extreme gradient lift alone.Among the 20 features with high contribution in the random forest model,8 were surgery⁃related variables,namely,duration of anesthesia,general thoracic surgery,orthopaedic surgery,total fluid loss during sur⁃gery,emergency surgery,total intraoperative crystalline fluid infusion,amount of intraoperative blood products used,total intraoperative colloidal fluid infusion,and elective surgery.Conclusion The predictive model of AKI after non⁃cardiac surgery developed by ma⁃chine learning can further strengthen the prevention of AKI and optimize and improve the prognosis of patients.

关 键 词：急性肾损伤 非心脏手术 预测模型 机器学习 随机森林 

分 类 号：R614[医药卫生—麻醉学]