文献速递重症综合管理与镇静镇痛创伤与烧伤

创伤性脊髓损伤危重患者7天死亡风险机器学习预测模型的开发与验证:多中心回顾性研究

Development and Validation of Machine Learning Models for Predicting 7-Day Mortality in Critically Ill Patients with Traumatic Spinal Cord Injury: A Multicenter Retrospective Study.
PubMed 文献 · Neurocrit Care · 2025-06-25
一句话结论基于MIMIC数据库261例与新疆医科大学第一附属医院45例ICU创伤性脊髓损伤患者,随机森林模型在预测7天死亡风险上表现最优,并已开发在线计算器。

本研究利用MIMIC 2.2数据库中261例创伤性脊髓损伤危重患者数据,采用Boruta与LASSO回归筛选关键特征,比较10种机器学习算法构建7天死亡风险预测模型,并以新疆医科大学第一附属医院ICU的45例患者进行外部验证。随机森林模型在原始数据集与SMOTE平衡合成数据集上,其ROC曲线、校准曲线、决策曲线及各项性能指标均持续优于其他模型。SHAP分析显示最低体温、最低收缩压和Charlson合并症指数为随机森林模型的关键预测因子。外部验证初步显示模型稳健性与临床适用性,并据此开发在线计算器供临床医生估算7天生存概率。

要点

  • 随机森林模型在原始与SMOTE平衡数据集上预测创伤性脊髓损伤危重患者7天死亡的ROC曲线、校准与决策曲线均持续优于其他9种算法。
  • SHAP分析提示最低体温、最低收缩压和Charlson合并症指数是随机森林模型的关键预测因子。
  • 模型基于MIMIC 2.2数据库261例与新疆医科大学第一附属医院ICU 45例创伤性脊髓损伤危重患者开发与外部验证。
  • 已开发在线计算器,供临床医生估算创伤性脊髓损伤危重患者的7天生存概率。
  • 作者认为随机森林模型在预测7天死亡风险方面表现良好,可能有助于支持临床决策。

适用人群与场景

  • 重症监护病房(ICU)中创伤性脊髓损伤的危重患者,数据来自MIMIC 2.2数据库(261例)与新疆医科大学第一附属医院ICU(45例)进行开发与外部验证。

研究设计与关键结果

  • 研究类型为多中心回顾性研究,开发队列来自MIMIC 2.2数据库261例创伤性脊髓损伤危重患者,外部验证队列为新疆医科大学第一附属医院ICU 45例同类患者。
  • 采用Boruta与LASSO回归算法筛选出10个关键特征,使用自适应提升、类别提升、梯度提升机、k近邻、轻量梯度提升机、逻辑回归、神经网络、随机森林、支持向量机与极端梯度提升共10种机器学习算法构建7天死亡风险预测模型。
  • 随机森林模型在原始数据集与SMOTE平衡合成数据集上,其ROC曲线、校准曲线、决策曲线分析及准确率、敏感度、特异度、精确率、F1分数等性能指标均持续优于其他算法。
  • SHAP分析显示最低体温、最低收缩压和Charlson合并症指数是随机森林模型中的关键预测因子。
  • 外部验证初步显示模型稳健性与临床适用性,并据此开发了在线计算器供临床医生估算7天生存概率。

证据类型、等级与局限

  • 证据类型:多中心回顾性研究(机器学习模型开发与外部验证)。
  • 局限:原文未报告局限性。

对床旁实践的改变

  • 作者认为随机森林模型在预测创伤性脊髓损伤危重患者7天死亡风险方面表现良好,可能有助于支持临床决策。
  • 已开发在线计算器,可估算创伤性脊髓损伤危重患者的7天生存概率,供临床医生使用。

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

BACKGROUND: Traumatic spinal cord injury (TSCI), a severe central nervous system injury, despite treatment advances, critically ill patients with TSCI face high short-term mortality. This study leverages machine learning to integrate standard intensive care unit (ICU) indicators, identifying 7-day high-mortality risk patients with TSCI to optimize treatment.

METHODS: Using critically ill patients with TSCI data from the Medical Information Mart for Intensive Care 2.2 database, this study employs the Boruta and LASSO regression algorithms to identify key features, developing a 7-day mortality risk prediction model in critically ill patients with TSCI using ten machine learning algorithms including Adaptive Boosting, Categorical Boosting, Gradient Boosting Machine, k-Nearest Neighbors, Light Gradient Boosting Machine, Logistic Regression, Neural Network, Random Forest (RF), Support Vector Machine, and Extreme Gradient Boosting. Model Performance is evaluated via receiver operating characteristic curves, calibration curves, decision curve analysis, accuracy, sensitivity, specificity, precision, and F1 score, whereas Shapley Additive Explanations ensure model interpretability. External validation with ICU data from the First Affiliated Hospital of Xinjiang Medical University further assesses the model's generalizability.

RESULTS: This study, collecting data from 261 and 45 critically ill patients with TSCI from the Medical Information Mart for Intensive Care database and the First Affiliated Hospital of Xinjiang Medical University's ICU, respectively, identified ten key features for model development, in which the RF model consistently outperformed others across raw and Synthetic Minority Over-sampling Technique-balanced synthetic datasets in receiver operating characteristic curves, calibration curves, decision curve analysis, and performance metrics. Shapley Additive Explanation analysis highlighted minimum body temperature, lowest systolic blood pressure, and Charlson Comorbidity Index as critical predictors in the RF model. External validation initially demonstrated the model's robustness and clinical applicability, leading to an online calculator that enables clinicians to estimate the 7-day survival probability of critically ill patients with TSCI.

CONCLUSIONS: The RF model exhibits favorable performance in predicting 7-day mortality risk among critically ill patients with TSCI, indicating its potential utility in supporting clinical decision-making.

来源:Neurocrit Care,查看原文。版权归原作者所有。

参考来源

  1. Wang Y, Luo X, Wang J, et al. Development and Validation of Machine Learning Models for Predicting 7-Day Mortality in Critically Ill Patients with Traumatic Spinal Cord Injury: A Multicenter Retrospective Study. Neurocrit Care. 2025. PMID: 40563047. doi:10.1007/s12028-025-02308-y

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。