人工智能驱动的ICU实时高乳酸血症预测:多队列国际回顾性研究及外部验证
该回顾性多队列研究基于AmsterdamUMCdb和MIMIC-III开发并内部验证两个机器学习模型,用于在普通ICU人群中每小时实时预测6小时和12小时 horizon 内高乳酸血症(乳酸>2 mmol/L),并在eICU和HiRID中进行外部验证。开发数据集包含13,573例ICU住院和577,414个小时样本。6小时模型内部验证AUC/AUPRC为0.87/0.384(AmsterdamUMCdb)和0.814/0.401(MIMIC-III),外部验证为0.772/0.295(eICU)和0.823/0.246(HiRID);12小时模型在所有队列中AUC≥0.76。校准和决策曲线分析证实了稳健的泛化能力和净临床获益,无需重新校准。
要点
- 模型可每小时实时预测普通ICU成人患者6小时和12小时内高乳酸血症(乳酸>2 mmol/L)。
- 6小时模型外部验证AUC为0.772(eICU)和0.823(HiRID),AUPRC为0.295和0.246。
- 12小时模型在所有队列中AUC≥0.76,性能一致。
- 校准和决策曲线分析显示稳健泛化且无需重新校准即有净临床获益。
- 纳入标准为成人ICU住院≥24小时且乳酸测量间隔≤12小时。
适用人群与场景
- 成人ICU住院患者,住院时长≥24小时,且乳酸测量间隔≤12小时。
- 研究场景为普通ICU人群,非特定ICU亚组。
诊断标准与关键阈值
- 高乳酸血症定义为乳酸>2 mmol/L。
- 预测 horizon 为6小时和12小时。
研究设计与关键结果
- 研究类型为回顾性多队列研究,使用AmsterdamUMCdb和MIMIC-III进行模型开发与内部验证,eICU和HiRID进行外部验证。
- 开发数据集包含13,573例ICU住院和577,414个小时样本。
- 6小时模型内部验证AUC/AUPRC:AmsterdamUMCdb为0.87/0.384,MIMIC-III为0.814/0.401;外部验证:eICU为0.772/0.295,HiRID为0.823/0.246。
- 12小时模型在所有队列中AUC≥0.76。
- 校准和决策曲线分析证实稳健泛化及净临床获益,无需重新校准。
证据类型、等级与局限
- 证据类型:回顾性多队列研究,使用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,这些经过验证的持续运行模型可在整个ICU住院期间实现实时、近未来乳酸风险评估,支持在未经选择的危重人群中更早识别组织低灌注和代谢性休克。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「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: Blood lactate is a key biomarker of tissue hypoperfusion and metabolic distress, reflecting the balance between oxygen delivery and cellular metabolism. Early identification of rising lactate levels is critical for timely intervention in critically ill patients, yet existing predictive tools are static and limited to specific ICU subgroups.
METHODS: We developed and externally validated two machine-learning models that provide hourly, real-time predictions of hyperlactatemia (lactate >2 mmol/L) within 6- and 12-hour horizons in general ICU populations. This retrospective, multi-cohort study used AmsterdamUMCdb and MIMIC-III for model development and internal validation, and eICU and HiRID for external validation. Adult ICU stays lasting ≥24 hours with lactate measurements ≤12 hours apart were included. Models based on routinely collected vital signs and laboratory data were trained using XGBoost and assessed for discrimination, calibration, decision-curve utility, and subgroup fairness.
RESULTS: The development dataset included 13,573 ICU stays and 577,414 hourly samples. The 6-hour model achieved AUROC/AUPR of 0.87/0.384 (AmsterdamUMCdb) and 0.814/0.401 (MIMIC-III) internally, and 0.772/0.295 (eICU) and 0.823/0.246 (HiRID) externally. The 12-hour model showed consistent performance (AUROC ≥0.76 across all cohorts). Calibration and decision-curve analyses confirmed robust generalization and net clinical benefit without recalibration.
CONCLUSIONS: These validated, continuously operating models enable real-time, near-future lactate risk assessment throughout the ICU stay, supporting earlier recognition of tissue hypoperfusion and metabolic shock in unselected critically ill populations.
参考来源
- Zappalà S, Rovati L, Alfieri F, et al. AI-Driven Real-Time Hyperlactatemia Prediction in ICU: A Multi-Cohort International Retrospective Study with External Validation. Shock. 2026. PMID: 42008773. doi:10.1097/SHK.0000000000002858
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。