基于时变机器学习模型的动态早期预警系统用于预测重症监护中血流动力学不稳定的开发与验证:一项多队列研究
该研究开发了时变血流动力学早期预警评分(TvHEWS),一种AI辅助模型,用于预测ICU患者血流动力学不稳定。模型通过台北荣民总医院VGHTPE 2010队列(2010–2021年)回顾性数据训练和内部验证,并用VGHTPE 2022队列前瞻性数据和MIMIC IV队列外部数据进一步验证。TvHEWS包含每小时更新模型,提供持续风险评估。VGHTPE 2010队列AUROC 0.93、精确率0.94、召回率0.77;VGHTPE 2022队列AUROC 0.92、精确率0.74、召回率0.76;MIMIC IV队列AUROC 0.82、精确率0.72、召回率0.36。校准图显示预测风险与观察风险对齐良好,Brier评分分别为0.082、0.085、0.116。模型在干预前预测血流动力学不稳定的提前时间分别为VGHTPE 2010队列7小时、VGHTPE 2022队列8.6小时、MIMIC IV队列21小时,假警报率低。结论认为TvHEWS通过整合多个临床参数提供持续动态风险评估,增强预见和管理关键循环问题的能力,可能通过早期干预改善患者结局,但需在其他医院进一步前瞻性验证以确认其在不同环境中的稳健性。
要点
- TvHEWS为每小时更新的时变机器学习模型,整合多个临床参数,对ICU患者血流动力学不稳定提供持续动态风险评估。
- 内部验证VGHTPE 2010队列AUROC 0.93、精确率0.94、召回率0.77;前瞻性VGHTPE 2022队列AUROC 0.92、精确率0.74、召回率0.76。
- 外部验证MIMIC IV队列AUROC 0.82、精确率0.72、召回率0.36,校准Brier评分0.116。
- 模型在干预前预测不稳定的提前时间:VGHTPE 2010队列7小时、VGHTPE 2022队列8.6小时、MIMIC IV队列21小时,假警报率低。
- 作者结论认为TvHEWS可能通过早期干预改善患者结局,但需在其他医院进一步前瞻性验证以确认其在不同环境中的稳健性。
适用人群与场景
- 重症监护病房(ICU)患者,用于预测血流动力学不稳定。
研究设计与关键结果
- 研究类型为多队列研究,开发并验证TvHEWS模型;训练和内部验证使用台北荣民总医院VGHTPE 2010队列(2010–2021年)回顾性数据,进一步验证使用VGHTPE 2022队列前瞻性数据和MIMIC IV队列外部数据。
- VGHTPE 2010队列:AUROC 0.93,精确率0.94,召回率0.77,Brier评分0.082。
- VGHTPE 2022队列:AUROC 0.92,精确率0.74,召回率0.76,Brier评分0.085。
- MIMIC IV队列:AUROC 0.82,精确率0.72,召回率0.36,Brier评分0.116。
- 校准图显示预测风险与观察风险对齐良好。
- 干预前预测血流动力学不稳定的提前时间:VGHTPE 2010队列7小时,VGHTPE 2022队列8.6小时,MIMIC IV队列21小时,假警报率低。
证据类型、等级与局限
- 证据类型:多队列研究(回顾性训练与内部验证、前瞻性验证及外部验证)。
- 局限:作者指出需在其他医院进一步前瞻性验证以确认模型在不同环境中的稳健性。
对床旁实践的改变
- 作者结论认为TvHEWS通过整合多个临床参数提供持续动态风险评估,增强预见和管理关键循环问题的能力,可能通过早期干预改善患者结局。
- 作者指出需在其他医院进一步前瞻性验证以确认其在不同环境中的稳健性。
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: Hemodynamic instability, a life-threatening condition marked by circulatory failure, presents a significant challenge in intensive care unit (ICU) settings, often leading to poor patient outcomes. Traditional monitoring methods that rely on single parameters may delay diagnosis. Machine learning models offer a solution by integrating multiple clinical parameters to more dynamically and accurately predict instability. METHODS: We developed the Time-varying Hemodynamic Early Warning Score (TvHEWS), an AI-assisted model used to predict hemodynamic instability in intensive care unit (ICU) patients. The model was trained and internally validated via retrospective data from the VGHTPE 2010 cohort (2010–2021) at Taipei Veteran General Hospital. It was further validated with prospective data from the VGHTPE 2022 cohort and external data from the MIMIC IV cohort. TvHEWS includes hourly updating models, providing continuous risk assessments. RESULTS: TvHEWS showed strong predictive performance. In the VGHTPE 2010 cohort, the AUROC was 0.93, with a precision of 0.94 and a recall of 0.77. In the VGHTPE 2022 cohort, the AUROC was 0.92, with precision and recall balanced at 0.74 and 0.76, respectively. The MIMIC IV cohort had a slightly lower AUROC of 0.82, with a precision of 0.72 and a recall of 0.36. The calibration plots showed good alignment between the predicted and observed risks, with Brier scores of 0.082, 0.085, and 0.116 for the VGHTPE 2010, VGHTPE 2022, and MIMIC IV cohorts, respectively. TvHEWS predicted hemodynamic instability for up to 7 h before intervention in the VGHTPE 2010 cohort, 8.6 h in the VGHTPE 2022 cohort, and 21 h in the MIMIC IV cohort, with low false alarm rates. CONCLUSIONS: TvHEWS addresses the challenge of early detection of hemodynamic instability by integrating multiple clinical parameters and offering continuous, dynamic risk assessments. It enhances the ability to anticipate and manage critical circulatory issues, potentially improving patient outcomes through earlier interventions. Further prospective validation in other hospitals is needed to confirm its robustness across diverse settings.
参考来源
- Chiang DH, Jiang Z, Tian C, et al. Development and validation of a dynamic early warning system with time-varying machine learning models for predicting hemodynamic instability in critical care: a multicohort study. Crit Care. 2025. PMID: 40702538. doi:10.1186/s13054-025-05553-x
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。