可解释机器学习模型预测ICU创伤性脑损伤患者静脉血栓栓塞:多中心研究
该研究基于MIMIC-IV数据库1564例创伤性脑损伤(TBI)患者训练并测试机器学习模型,外部验证采用解放军904医院ICU的304例TBI患者。逻辑回归模型表现最佳,测试集AUC为0.723,外部验证集AUC为0.759,Shapley分析显示ICU停留时间和肺部感染是主要预测因素。
要点
- 逻辑回归模型对TBI后早期VTE具有较强预测能力,外部验证AUC为0.759。
- ICU停留时间和肺部感染是影响VTE预测的首要因素。
- 基于逻辑回归的列线图在校准曲线和临床决策曲线上表现良好。
核心结论
- 逻辑回归模型对TBI后早期VTE表现出较强的预测能力,可能有助于减少VTE相关并发症并改善患者结局。
研究设计与样本
- 利用MIMIC-IV(2008-2021)中1564例TBI患者数据,按80%训练、20%测试划分。
- 外部验证队列为解放军904医院ICU的304例TBI患者(2021-2024)。
- 采用Boruta和LASSO回归进行特征选择,训练了支持向量机(SVM)和逻辑回归(LR)等四种机器学习算法。
- 使用AUC等指标评估性能,并用Shapley Additive explanations方法解释特征重要性。
主要发现
- VTE的独立危险因素包括住院时间、收缩压、下肢骨折和肺部感染。
- 最终选取6个变量:ICU停留时间、住院时间、肺部感染、心率、部分凝血活酶时间和下肢骨折。
- 4种机器学习模型中,LR模型表现优异,测试集AUC为0.723,外部验证集AUC为0.759。
- 基于LR模型的列线图在校准曲线和临床决策曲线上表现良好。
- Shapley分析强调ICU停留时间和肺部感染是影响预测的主要决定因素。
临床怎么用
- 可考虑使用基于LR模型的列线图对ICU的TBI患者进行VTE风险预测。
- 重点关注ICU停留时间较长和合并肺部感染的患者,以辅助VTE预防决策。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To develop and validate a machine learning model predicting venous thromboembolism (VTE) risk in intensive care unit (ICU) patients with traumatic brain injury (TBI). METHODS: Utilizing data from 1564 TBI patients in MIMIC-IV (2008-2021), models were trained (80%) and tested (20%). We used 304 TBI patients from the 904th Hospital of the Joint Support Force of the Chinese People's Liberation Army's ICU (2021-2024) as an external validation cohort. Feature selection employed Boruta and Least Absolute Shrinkage and Selection Operatoř regression. Four machine learning algorithms, including support vector machine (SVM) and logistic regression (LR), were trained. Performance was evaluated using area under the curve and other metrics. The Shapley Additive explanations method interpreted feature importance. RESULTS: Independent risk factors for VTE were identified as length of hospital stay, systolic blood pressure, lower limb fracture, and lung infection. Six variables were selected: ICU stay duration, length of hospital stay, lung infection, heart rate, partial thromboplastin time, and lower limb fracture. Among the 4 ML models, the LR model demonstrated excellent performance, with area under the curve values of 0.723 for the test set, and 0.759 for the external validation set. The nomogram based on the LR model showed good performance in the calibration curve and the clinical decision curve. Furthermore, Shapley Additive explanations analysis highlights that the length of ICU stay and lung infection are the primary determinants influencing the prediction. CONCLUSIONS: The logistic regression model demonstrates strong predictive capability for early VTE following TBI, which may potentially contribute to reducing complications associated with VTE and improving patient outcomes.
引用
- Qi H, Li L, Fang J, et al. Development and Validation of an Interpretable Machine Learning Model for Predicting venous Thromboembolism in ICU patients With Traumatic Brain Injury: A Multicenter Study. World Neurosurg. 2025. PMID: 40885477. doi:10.1016/j.wneu.2025.124399
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。