文献速递重症综合管理与镇静镇痛复苏、休克与脓毒症

基于机器学习构建重症监护病房脓毒症患者院内死亡预测模型

[Construction of a predictive model for in-hospital mortality of sepsis patients in intensive care unit based on machine learning].
PubMed 文献 · Zhonghua Wei Zhong Bing Ji Jiu Yi Xue · 2023-01-01
一句话结论回顾性分析741例ICU脓毒症患者,随机森林模型预测院内死亡的AUC为0.871,优于SOFA(0.749)和APACHEⅡ(0.760)评分。

该研究回顾性分析2015年4月至2021年4月济宁医学院附属医院ICU收治的741例脓毒症患者临床数据,按是否发生院内死亡分为死亡组(351例)和存活组(390例),院内病死率47.4%。将70%病例随机分为训练集、30%为验证集,基于逻辑回归、K近邻、支持向量机、决策树、随机森林、极端梯度提升和人工神经网络7种机器学习模型构建院内死亡预测模型,并用ROC曲线、校准曲线和决策曲线分析评估。结果显示随机森林模型AUC最高(0.871),精确率0.750、准确率0.785、召回率0.773、F1评分0.761,校准和临床净获益均最佳,优于SOFA评分(AUC 0.749)和APACHEⅡ评分(AUC 0.760)。作者认为机器学习模型可作为预测脓毒症患者院内死亡的可靠工具,随机森林模型预测性能最佳,有助于临床医生识别高危患者并实施早期干预以降低病死率。

要点

  • 741例ICU脓毒症患者中351例院内死亡,院内病死率47.4%。
  • 随机森林模型预测院内死亡的AUC为0.871,高于SOFA评分0.749和APACHEⅡ评分0.760。
  • 随机森林模型精确率0.750、准确率0.785、召回率0.773、F1评分0.761,均为所有模型中最优。
  • 校准曲线显示随机森林模型在7种机器学习模型中表现最佳;决策曲线显示其净获益和阈值概率优于其他模型。
  • 死亡组与存活组在性别、年龄、APACHEⅡ评分、SOFA评分、GCS、心率、氧合指数、机械通气比例及时间、NE使用比例及最大剂量、Lac、APTT、ALB、SCr、BUN、BUA、pH、BE、K+等方面差异有统计学意义。

适用人群与场景

  • 2015年4月至2021年4月济宁医学院附属医院ICU收治的脓毒症患者,共741例,其中存活出院390例、院内死亡351例。

研究设计与关键结果

  • 研究类型:回顾性分析;样本量741例;人群:ICU脓毒症患者;按是否院内死亡分为死亡组(351例)和存活组(390例),院内病死率47.4%。
  • 方法:70%病例随机分为训练集构建模型,30%为验证集;基于逻辑回归、K近邻、支持向量机、决策树、随机森林、极端梯度提升和人工神经网络7种机器学习模型构建院内死亡预测模型,用ROC曲线、校准曲线和决策曲线分析评估,并与SOFA和APACHEⅡ评分比较。
  • ROC曲线分析:随机森林、极端梯度提升、逻辑回归、人工神经网络、决策树、支持向量机、K近邻模型及SOFA评分、APACHEⅡ评分预测院内死亡的AUC分别为0.871、0.846、0.751、0.747、0.677、0.657、0.555、0.749、0.760。
  • 随机森林模型精确率0.750、准确率0.785、召回率0.773、F1评分0.761,在所有模型中最高,判别能力最佳。
  • 校准曲线显示随机森林模型在7种机器学习模型中表现最佳;决策曲线显示随机森林模型较其他模型具有更大净获益和阈值概率,临床效用最佳。
  • 死亡组与存活组在性别、年龄、APACHEⅡ评分、SOFA评分、GCS、心率、氧合指数(PaO2/FiO2)、机械通气比例、机械通气时间、NE使用比例、最大NE、Lac、APTT、ALB、SCr、BUN、BUA、pH值、BE、K+方面差异有统计学意义。

证据类型、等级与局限

  • 证据类型:回顾性分析(单中心)。
  • 局限:原文未报告局限性。

对床旁实践的改变

  • 作者认为机器学习模型可作为预测脓毒症患者院内死亡的可靠工具,随机森林模型预测性能最佳,有助于临床医生识别高危患者并实施早期干预以降低病死率。

提炼说明

  • 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开

OBJECTIVE: To analyze the risk factors of in-hospital death in patients with sepsis in the intensive care unit (ICU) based on machine learning, and to construct a predictive model, and to explore the predictive value of the predictive model.

METHODS: The clinical data of patients with sepsis who were hospitalized in the ICU of the Affiliated Hospital of Jining Medical University from April 2015 to April 2021 were retrospectively analyzed,including demographic information, vital signs, complications, laboratory examination indicators, diagnosis, treatment, etc. Patients were divided into death group and survival group according to whether in-hospital death occurred. The cases in the dataset (70%) were randomly selected as the training set for building the model, and the remaining 30% of the cases were used as the validation set. Based on seven machine learning models including logistic regression (LR), K-nearest neighbor (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost) and artificial neural network (ANN), a prediction model for in-hospital mortality of sepsis patients was constructed. The receiver operator characteristic curve (ROC curve), calibration curve and decision curve analysis (DCA) were used to evaluate the predictive performance of the seven models from the aspects of identification, calibration and clinical application, respectively. In addition, the predictive model based on machine learning was compared with the sequential organ failure assessment (SOFA) and acute physiology and chronic health evaluation II (APACHE II) models.

RESULTS: A total of 741 patients with sepsis were included, of which 390 were discharged after improvement, 351 died in hospital, and the in-hospital mortality was 47.4%. There were significant differences in gender, age, APACHE II score, SOFA score, Glasgow coma score (GCS), heart rate, oxygen index (PaO2/FiO2), mechanical ventilation ratio, mechanical ventilation time, proportion of norepinephrine (NE) used, maximum NE, lactic acid (Lac), activated partial thromboplastin time (APTT), albumin (ALB), serum creatinine (SCr), blood urea nitrogen (BUN), blood uric acid (BUA), pH value, base excess (BE), and K+ between the death group and the survival group. ROC curve analysis showed that the area under the curve (AUC) of RF, XGBoost, LR, ANN, DT, SVM, KNN models, SOFA score, and APACHE II score for predicting in-hospital mortality of sepsis patients were 0.871, 0.846, 0.751, 0.747, 0.677, 0.657, 0.555, 0.749 and 0.760, respectively. Among all the models, the RF model had the highest precision (0.750), accuracy (0.785), recall (0.773), and F1 score (0.761), and best discrimination. The calibration curve showed that the RF model performed best among the seven machine learning models. DCA curve showed that the RF model exhibited greater net benefit as well as threshold probability compared to other models, indicating that the RF model was the best model with good clinical utility.

CONCLUSIONS: The machine learning model can be used as a reliable tool for predicting in-hospital mortality in sepsis patients. RF models has the best predictive performance, which is helpful for clinicians to identify high-risk patients and implement early intervention to reduce mortality.

来源:Zhonghua Wei Zhong Bing Ji Jiu Yi Xue,查看原文。版权归原作者所有。

参考来源

  1. Zhu M, Hu C, He Y, et al. [Construction of a predictive model for in-hospital mortality of sepsis patients in intensive care unit based on machine learning]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2023. PMID: 37545445. doi:10.3760/cma.j.cn121430-20221219-01104

供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。