人工智能/机器学习脓毒症预测系统:从开发到部署的检验科考量
综述指出临床检验科在机器学习脓毒症预测系统的开发、验证和治理中基本缺席,而实验室变量是多数已部署工具的核心输入。平台转换、试剂配方调整和参考区间更新等分析事件可导致输入分布偏移,进而降低模型性能,但实验室特异性变化的影响尚未充分表征。支持已部署工具的对照结局数据仍有限,且报告性能常无法在外部站点复现。
要点
- 检验科应参与脓毒症AI工具的开发、验证与治理,因其实验室变量是核心输入。
- 平台转换、试剂配方调整和参考区间更新可导致输入分布偏移并降低模型性能。
- 已部署工具的对照结局数据有限,报告性能常无法在外部站点复现。
核心结论
- 检验科在脓毒症机器学习预测系统的开发、验证和治理中基本缺席,构成重要缺口。
- 实验室变量是多数已部署脓毒症AI工具的核心输入,分析事件可导致输入分布偏移并降低模型性能。
- 支持已部署工具的对照结局数据仍有限,报告性能常无法在外部站点复现。
研究设计与样本
- 综述,面向在实践或实施评估中遇到脓毒症AI工具的临床检验人员。
- 描述了模型输入的数据来源及分析背景、分析前变异和医嘱模式对模型行为的影响。
- 剖析了五种有真实世界实施证据的已部署算法及更广泛的临床前研究格局。
主要发现
- 常规分析事件如平台转换、试剂配方调整和参考区间更新可改变实验室输入变量的分布。
- 分布偏移是模型性能下降的已知原因,但实验室特异性变化对临床在用脓毒症工具的影响尚未充分表征。
- 临床团队和检验科目前均不具备良好条件来检测这些影响。
- 已部署工具的对照结局数据仍有限,报告性能常无法在外部站点复现。
临床怎么用
- 检验科应参与脓毒症AI工具的实施团队,提供分析质量、数据溯源和变更管理方面的专长。
- 需关注平台转换、试剂配方调整和参考区间更新等分析事件对模型输入分布的影响。
- 随着领域向脓毒症亚型分型、持续监测和本地模型校准发展,检验科在分析质量和监测方面的专长将日益核心。
局限性
- 实验室特异性变化对临床在用脓毒症工具的影响尚未充分表征。
- 支持目前已部署工具的对照结局数据仍有限。
- 报告的性能频繁无法在外部站点复现。
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
Sepsis prediction algorithms based on machine learning are increasingly embedded in electronic health record systems, and clinical laboratories have been largely absent from their development, validation, and governance. This is a consequential gap. Laboratory variables are central inputs to most deployed sepsis artificial intelligence (AI) tools, and routine analytical events such as platform transitions, reagent reformulations, and reference interval updates can shift the distributions of these inputs. Such distribution shifts are a recognized cause of degraded model performance; however, the effects of laboratory-specific changes remain insufficiently characterized for sepsis tools in clinical use. Neither clinical teams nor laboratories are currently well positioned to detect these effects. This review is written for clinical laboratorians who encounter these tools in practice or are asked to evaluate them for implementation. We describe the data sources used as model inputs and how analytical context, preanalytical variability, and ordering patterns influence model behavior. Five deployed algorithms with real-world implementation evidence are profiled alongside the broader preclinical landscape, including early warning systems, mortality and complication prediction models, and sepsis subtyping research. Development and implementation pitfalls are reviewed with attention to laboratory-specific considerations, and the regulatory and governance landscape is addressed with emphasis on where laboratorian involvement is most needed. Controlled outcomes data supporting currently deployed tools remain limited, and reported performance frequently does not replicate at external sites. Laboratorians bring expertise in analytical quality, data provenance, and change management that is not replicated elsewhere on implementation teams. As the field moves toward sepsis subtyping, continuous monitoring, and local model calibration, laboratorian expertise in analytical quality and monitoring will become increasingly central to how these tools are built, deployed, and maintained.
引用
- Durant, Aldewereld, Peaper, et al. AI/ML-enabled sepsis prediction systems: laboratory considerations from development to deployment. J Clin Microbiol. 2026. PMID: 42937359. doi:10.1128/jcm.01461-25
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。