基于机器学习建立脓毒症心肾综合征患者早期死亡风险预测模型

Early mortality risk prediction models for patients with sepsis-induced cardiorenal syndrome based on machine learning

目的探讨机器学习算法构建脓毒症心肾综合征患者早期死亡风险预测模型的方法,为临床早期识别高危患者及精准治疗提供依据。方法入选同济大学附属同济医院2015年1月1日至2019年5月31日期间入院的脓毒症心肾综合征患者为研究对象,收集患者确诊时的临床表现、实验室检查及治疗情况等资料。研究终点事件定义为患者确诊后30 d内死亡。运用Python软件构建不同机器学习算法模型,采用受试者工作特征曲线下面积(AUC)评估各模型的预测效能。运用构建的最优模型筛选疾病相关风险因素,构建可视化决策树模型和半朴素贝叶斯(sNB)模型。结果340例患者入选本研究,其中114例(33.5%)患者确诊后30 d内死亡。支持向量机(SVM)、随机森林(RF)、梯度提升树(GBDT)、极端梯度提升(XGBoost)和轻量梯度提升(LGBM)5种模型的AUC值分别为0.652、0.868、0.870、0.754和0.852,其中GBDT模型预测患者发生终点事件的AUC值最优。依据GBDT模型特征重要度评分筛选出前10项患者预后的影响因素,包括序贯器官衰竭评估(SOFA)总评分、神经系统SOFA评分、血管活性药物应用史、高敏肌钙蛋白(cTNI)、年龄、肌红蛋白(MYO)、循环系统SOFA评分、慢性肾脏病史、心率和基线血肌酐值等参数,建立可视化决策树模型,模型共4层,15个节点,8个终端节点。依据SOFA总评分、MYO变化率、基线血肌酐值和年龄等4项影响因素建立决策树流程,模型预测患者发生终点事件的AUC值为0.690。sNB模型提示总SOFA总评分与神经系统SOFA评分、SOFA总评分与血管活性药物、cTNI与基线血肌酐值间的相互作用影响患者的短期预后。结论基于机器学习建立的脓毒症心肾综合征患者早期死亡风险预测模型结果提示,高SOFA评分仍然是预测脓毒症心肾综合征患者预后不良的首要危险因素。本研究建立的可视化决策树模型和sNB模型可在疾病早期针对高危患者进行临床判断,为脓毒症患者的精准治疗提供预测依据。

Objective To explore the method of constructing an early mortality risk prediction model for patients with sepsis-induced cardiorenal syndrome by machine learning algorithm,so as to provide a basis for early clinical identification of high-risk patients and accurate treatment.Methods Patients with sepsis-induced cardiorenal syndrome from January 1,2015 to May 31,2019 in Tongji Hospital,Tongji University were enrolled.Basic characteristics,laboratory indexes,hospitality treatment and other relevant baseline data were collected.Thirty-day mortality was defined as the primary end-point event after the enrolled patients were diagnosed.Python software was applied to establish different machine learning models,and the area under the receiver-operating characteristic curve(AUC)was used to evaluate the predictive value of models.Disease-related risk factors were selected according to the most optimal model.Importantly,visualized decision tree and semi-naive Bayesian(sNB)models were established to further explore the interrelationship between these risk factors.Results A total of 340 patients were included,of whom 114 patients(33.5%)died within 30 days after diagnosis.The AUC of support vector machine(SVM),random forest(RF),gradient boosting decision tree(GBDT),extreme gradient boosting(XGBoost),and light gradient boosting machine(LGBM)prediction models were 0.652,0.868,0.870,0.754,and 0.852,respectively.The AUC of GBDT model had the most efficiency to predict end-point events,and the prediction AUC value was better.According to the feature ranking of GBDT model,the relevant influencing factors were selected,including total sequential organ failure assessment(SOFA)score,neural SOFA score,vasoactive drug application,cardiac troponin I(cTNI),age,myoglobin,circulation system SOFA score,chronic kidney disease,heart rate and baseline serum creatinine.Visualized decision tree model had 4 layers,15 nodes and 8 terminal nodes as evidenced by total SOFA score,myoglobin,baseline serum creatinine and age.The total SOFA score,change rate of myoglobin,serum creatinine and age were included into the visualized decision model.The AUC value of the model for predicting end-point event was 0.690.sNB model revealed complex correlation between the risk factors,in which neural SOFA score was related to total SOFA score,vasoactive drug application was related to total SOFA score,and cTNI was related to baseline serum creatinine.Conclusions A risk prediction model for patients with sepsis-induced cardiorenal syndrome is established and the model showes that high SOFA score remains the primary risk factor for patients with sepsis-induced cardiorenal syndrome based machine learning.Visualized decision tree and sNB models help clinicians to further identify the dependence and logic relationship among these risk factors clearly and provide a novel method to predict mortality risk for patients with sepsis-induced cardiorenal syndrome.