用简约机器学习模型量化肺炎与脓毒症的免疫失调:跨诊疗场景的多队列分析及对氢化可的松随机对照试验的再分析
研究基于三个独立队列中398例社区获得性肺炎患者的35种血浆生物标志物,构建免疫失调分期(DIP1-3)与连续评分(cDIP),并开发由降钙素原、可溶性TREM-1和IL-6组成的三标志物机器学习模型。该模型在五个外部队列(n=1191)中验证,对DIP分期准确率91.2%,cDIP均方根误差0.056;cDIP每增加10%与死亡率(OR 1.26)和继发感染(OR 1.50)升高相关。对CAPE COD试验的再分析显示,氢化可的松仅在模型判定为重度免疫失调者中带来生存获益。
要点
- 三标志物(降钙素原、可溶性TREM-1、IL-6)模型可量化免疫失调,DIP分期准确率91.2%。
- cDIP每增加10%,死亡风险OR 1.26、继发感染OR 1.50,独立于临床严重程度。
- 氢化可的松的生存获益仅见于模型判定的重度免疫失调者,按临床严重程度分层未见此效应修饰。
核心结论
- 作者提供了一个公开可用的三标志物框架来判定宿主反应失调程度,认为其对精准指导免疫调节治疗具有潜在价值。
研究设计与样本
- 本研究为多队列分析及对一项随机对照试验的再分析,主要目的是推导并验证不依赖临床表现或结局的免疫失调分类与连续评分。
- 研究在急诊科、普通病房和重症监护室等不同诊疗场景中,用三个独立队列的社区获得性肺炎患者数据测量35种反映关键宿主反应域的血浆生物标志物。
- 研究采用无监督轨迹推断分析识别免疫失调梯度,形成离散的免疫失调阶段(DIP)和连续评分(cDIP,0-1),并基于35种生物标志物开发两个简约机器学习模型预测DIP分期和cDIP评分。
- 研究在五个独立队列中验证模型捕捉免疫失调和预测临床结局的能力,并对评估氢化可的松治疗重症社区获得性肺炎的CAPE COD试验(NCT02517489)做事后分析。
- CAPE COD试验事后分析评估不同DIP分期和cDIP连续谱上的治疗效果,以及氢化可的松如何影响随时间变化的失调轨迹。
主要发现
- 研究基于35种血浆生物标志物将398例社区获得性肺炎患者沿免疫失调从轻到重 continuum 排列,得到三个失调阶段(DIP1-3)和一个连续评分(cDIP)。
- 临床严重程度被证明是免疫失调的不充分替代指标。
- 由降钙素原、可溶性TREM-1和IL-6组成的三标志物机器学习框架准确预测了基于35种生物标志物推导的失调程度,DIP分期准确率91.2%,cDIP均方根误差0.056。
- 免疫失调加重(以DIP和cDIP反映)与死亡率逐渐升高相关(cDIP每增加10%,OR 1.26 [95%CI 1.13-1.40],p<0.0001),也与继发感染相关(每增加10%,OR 1.50 [95%CI 1.22-1.93],p=0.0005),且独立于临床严重程度。
- 三标志物工具在五个外部队列中得到验证,这些队列涵盖不同感染、严重程度和诊疗场景(n=1191)。
关键数值与推荐
- 三标志物机器学习框架(降钙素原、可溶性TREM-1、IL-6)预测基于35种生物标志物的DIP分期准确率为91.2%,cDIP均方根误差为0.056。
- cDIP每增加10%,死亡率OR 1.26(95%CI 1.13-1.40,p<0.0001),继发感染OR 1.50(95%CI 1.22-1.93,p=0.0005)。
- CAPE COD试验中,DIP3者氢化可的松30天死亡率OR 0.25(95%CI 0.05-0.85,p=0.042),cDIP≥0.63者OR 0.21(95%CI 0.10-0.72,p=0.011)。
临床怎么用
- 作者认为该三标志物框架可用于判定宿主反应失调程度,对精准指导免疫调节治疗具有潜在价值。
- CAPE COD试验再分析结果按原文表述为氢化可的松仅在被模型判定为重度失调的参与者中带来生存获益。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「局限性」,故未列出该栏;全文请点「阅读原文」。
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
Lancet Respir Med . 2026 Apr;14(4):327-340. doi: 10.1016/S2213-2600(25)00429-1. Epub 2026 Mar 16. ABSTRACT BACKGROUND: Sepsis is a dysregulated host response to infection resulting in life-threatening organ failure. Although immune dysregulation is central to the sepsis definition, immunomodulation trials enrol participants based on clinical severity, not the extent of dysregulation, which could contribute to treatment heterogeneity. A pragmatic way to quantify immune dysregulation could improve prognostication, help to evaluate treatment responses, and identify individuals most likely to benefit from immunomodulation. We aimed to construct a parsimonious machine-learning tool that defines and quantifies immune dysregulation, thereby supporting biologically informed immunomodulation. METHODS: In this multicohort analysis and reanalysis of a randomised controlled trial, the primary objective was to derive and validate a categorical and continuous immune dysregulation score that is independent of clinical presentation or outcome. We measured 35 plasma biomarkers reflecting key host response domains in individuals with community-acquired pneumonia (CAP) across different care settings (emergency department, general ward, and intensive care unit) and disease severities using data from three independent cohorts. We applied unsupervised trajectory inference analysis to identify an immune dysregulation gradient captured as discrete immune dysregulation stages (Dysregulated Immune Profile [DIP]) and a continuous score (cDIP; 0-1). We developed two parsimonious machine-learning models to predict the DIP stages and cDIP scores based on 35 biomarkers, and validated their ability to capture immune dysregulation and predict clinical outcomes in five independent cohorts. On the basis of our hypothesis that only individuals with severe immune dysregulation benefit from immunomodulation, we carried out a post-hoc analysis of a randomised trial evaluating hydrocortisone in severe CAP (CAPE COD trial, NCT02517489), assessing treatment effects across DIP stages and the cDIP continuum, and how hydrocortisone influenced dysregulation trajectories over time. FINDINGS: We organised 398 participants with CAP along a continuum of immune dysregulation from mild to severe on the basis of 35 plasma biomarkers, yielding three dysregulation stages (DIP1-3) and a continuous score (cDIP). Clinical severity proved to be an inadequate proxy for immune dysregulation. A three-biomarker machine-learning framework (procalcitonin, soluble TREM-1, and IL-6) accurately predicted the degree of dysregulation derived from 35 biomarkers (DIP stage accuracy 91·2%; cDIP root mean square error 0·056). Although the framework was not designed for outcome prediction, increased immune dysregulation-reflected in DIP and cDIP-was associated with a gradual rise in mortality (cDIP odds ratio [OR] 1·26 [95% CI 1·13-1·40] per 10% increase, p<0·0001) and secondary infections (OR 1·50 [1·22-1·93] per 10% increase, p=0·0005), independent of clinical severity. The three-biomarker tool was validated in five external cohorts of varying infections, severities, and care settings (n=1191). Reanalysis of the CAPE COD trial showed that hydrocortisone conferred a survival benefit only in participants classified as severely dysregulated by our model (30-day mortality: DIP3 OR 0·25 [0·05-0·85], p=0·042; cDIP ≥0·63 OR 0·21 [0·10-0·72], p=0·011), accompanied by faster immune recovery (time × treatment interaction, p<0·0001). No such effect modification was observed when stratifying participants by clinical severity. Interpretation We have provided a publicly available three-biomarker framework to determine the extent of host response dysregulation with potential value for precision-guided immunomodulatory therapy. Funding EU Horizon 2020.
引用
- PubMed 顶刊 · 肺炎:Quantifying immune dysregulation in pneumonia and sepsis with a parsimonious machine-learning model: a multicohort analysis across care settings and reanalysis of a hydrocortisone randomised controlled trial
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。