2024 年第 7 期 第 0 卷

急性心肌梗死合并肾功能不全患者主要不良心脑血管事件风险预测模型构建

Constructing risk prediction models for major adverse cardiovascular and cerebrovascular events in patients with acute myocardial infarction and renal insufficiency

作者：韩福生田雪米玉红何华

英文作者：Han Fusheng Tian Xue Mi Yuhong He Hua

单位：首都医科大学附属北京安贞医院急诊重症监护病房，北京100029

英文单位：Emergency and Critical Care Center Ward Beijing Anzhen Hospital Capital Medical University Beijing 100029 China

关键词：急性心肌梗死；肾功能不全；机器学习；主要不良心脑血管事件；Shapley加性解释算法

英文关键词：Acutemyocardialinfarction;Renalinsufficiency;Machinelearning;Majoradversecardiovascularandcerebrovascularevents;Shapley′sadditiveexplanationalgorithm

• 摘要：

• 目的 应用不同机器学习算法构建急性心肌梗死（AMI）合并肾功能不全患者主要不良心脑血管事件（MACCE）的风险预测模型。方法 选取2014年1月至2019年8月于首都医科大学附属北京安贞医院住院治疗的740例AMI合并肾功能不全患者为研究对象，收集患者一般特征、生命体征、合并症和实验室检查结果等临床资料。采用简单随机抽样法按80%∶20%将研究对象分为训练集（592例）和测试集（148例），采用逻辑回归、随机森林、极限梯度提升（XGBoost）、支持向量机和深度神经网络5种机器学习算法分别构建MACCE的预测模型。采用受试者工作特征曲线下面积（AUC）评估模型的可靠性，选择最优模型。使用Shapley加性解释算法评估特征影响并进行特征选择构建最终模型。结果 740例AMI合并肾功能不全患者中有473例（63.9%）发生MACCE。XGBoost模型的AUC最大（AUC=0.862）。在根据特征重要性等级对特征进行减少后，建立了具有5个特征的可解释的最终XGBoost模型。最终的模型可以在内部验证中准确预测MACCE发生（AUC=0.955），其中影响XGBoost模型重要临床特征分别为血尿酸、白蛋白、糖化血清白蛋白，体重和血小板计数。结论 基于机器学习算法的5种模型中XGBoost模型预测AMI合并肾功能不全患者发生MACCE效果最佳。

• Objective To construct risk prediction models for major adverse cardiovascular and cerebrovascular events (MACCE) in patients with acute myocardial infarction (AMI) and renal insufficiency using different machine learning algorithms. Methods A total of 740 patients with AMI and renal insufficiency who were hospitalized in Beijing Anzhen Hospital, Capital Medical University from January 2014 to August 2019 were selected as the research objects. Clinical data such as general characteristics, vital signs, comorbidities and laboratory examination results of the patients were collected. The subjects were divided into training set (592 cases) and test set (148 cases) by the simple random sampling method at a ratio of 80%∶20%. Five machine learning algorithms, including Logistic regression, random forest, extreme gradient boosting (XGBoost), support vector machine and deep neural network, were used to construct the prediction model of MACCE. The area under the receiver operating characteristic curve (AUC) was used to evaluate the reliability of the model and select the optimal model. Shapley′s additive explanation algorithm was used to evaluate the feature influence and perform feature selection to construct the final model. Results  MACCE occurred in 473 (63.9%) of 740 patients with AMI and renal insufficiency. The XGBoost model had the largest AUC (AUC=0.862). After reducing the features according to their importance rank, an interpretable final XGBoost model with five features was built. The final model could accurately predict the occurrence of MACCE in the internal validation (AUC=0.955). The important clinical characteristics affecting the XGBoost model were serum uric acid, albumin, glycated albumin, body weight and platelet count. Conclusion  Among the five models based on machine learning algorithms, XGBoost model has the best effect in predicting MACCE in patients with AMI and renal insufficiency.