Pensier Joris, Fosset Maxime, Paschold Béla-Simon, von Wedel Dario, Redaelli Simone, Braeuer Ben L P, Novack Victor, Balzer Felix, Jung Boris, Amato Marcelo B P, Jaber Samir, Talmor Daniel, Baedorf-Kassis Elias, Schaefer Maximilian S
Anesthesiology and Intensive Care, Anesthesia and Critical Care Department B, Saint Eloi Teaching Hospital, PhyMedExp, University of Montpellier, INSERM U1046, Montpellier, France.
Department of Anesthesia, Critical Care and Pain Medicine, Center for Anesthesia Research Excellence Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, USA.
Intensive Care Med. 2025 Aug 21. doi: 10.1007/s00134-025-08089-4.
Inflammatory phenotypes of acute respiratory distress syndrome (ARDS) can predict patient outcomes and potentially response to treatment. The aim was to assess whether inflammatory phenotypes can be characterized over time using clinical surrogate data and used to guide therapy with corticosteroids.
Individual patient data and biomarkers from six multicenter randomized controlled trials (development, n = 1207; validation, n = 2751) were analyzed to establish an open-source AI Clinical Classifier ( https://bostonmontpelliercare.shinyapps.io/AIClarity ) for inflammatory phenotypes of ARDS using routine clinical data. Then, patients from a retrospective cohort (investigation, n = 5578) underwent classification from baseline to day 30. A discrete-time Bayesian Markov model assessed temporal stability at 3-day intervals. A target trial emulation and longitudinal logistic regression assessed corticosteroid effect on 30-day mortality depending on phenotype.
The AI Clinical Classifier identified 2169 (39%) hyperinflammatory and 3409 (61%) hypoinflammatory patients. 1053 (49%) and 826 (24%) patients died within 30 days, respectively (p < 0.001). Over 30 days, 49%(1072/2169) of hyperinflammatory patients at baseline transitioned to hypoinflammatory, and 7%(229/3409) of hypoinflammatory patients at baseline transitioned to hyperinflammatory (p < 0.001). Phenotypes predicted response to corticosteroids, with lower mortality in hyperinflammatory patients (IPW-weighted hazard ratio [HR]: 0.81 [0.67-0.98], p = 0.033), and higher mortality in hypoinflammatory patients (IPW-weighted HR: 1.26 [1.06-1.50], p = 0.009). At day 3, a positive response to corticosteroids only persisted among patients who remained hyperinflammatory (adjusted odds ratio = 0.51, 95% CI 0.32-0.80, p = 0.004).
Characterization of inflammatory ARDS phenotypes using clinical surrogate data allows physicians to monitor patients throughout the course of the disease and guide clinical treatment. Corticosteroids may be beneficial in hyperinflammatory ARDS and harmful in hypoinflammatory ARDS.
急性呼吸窘迫综合征（ARDS）的炎症表型可预测患者预后及对治疗的潜在反应。本研究旨在评估是否可利用临床替代数据随时间对炎症表型进行特征描述，并用于指导皮质类固醇治疗。
分析来自六项多中心随机对照试验（开发队列，n = 1207；验证队列，n = 2751）的个体患者数据和生物标志物，以建立一个使用常规临床数据对ARDS炎症表型进行分类的开源人工智能临床分类器（https://bostonmontpelliercare.shinyapps.io/AIClarity ）。然后，对来自回顾性队列（研究队列，n = 5578）的患者从基线至第30天进行分类。一个离散时间贝叶斯马尔可夫模型以3天为间隔评估时间稳定性。一个目标试验模拟和纵向逻辑回归评估了皮质类固醇根据表型对30天死亡率的影响。
人工智能临床分类器识别出2169名（39%）高炎症患者和3409名（61%）低炎症患者。分别有1053名（49%）和826名（24%）患者在30天内死亡（p < 0.001）。在30天内，基线时49%（1072/2169）的高炎症患者转变为低炎症，基线时7%（229/3409）的低炎症患者转变为高炎症（p < 0.001）。表型可预测对皮质类固醇的反应，高炎症患者死亡率较低（逆概率加权风险比[HR]：0.81[0.67 - 0.98]，p = 0.033），低炎症患者死亡率较高（逆概率加权HR：1.26[1.06 - 1.50]，p = 0.009）。在第3天，仅在仍为高炎症的患者中，皮质类固醇的阳性反应持续存在（调整优势比 = 0.51，95%CI 0.32 - 0.80，p = 0.004）。
利用临床替代数据对ARDS炎症表型进行特征描述，可使医生在疾病全过程中监测患者并指导临床治疗。皮质类固醇可能对高炎症ARDS有益，而对低炎症ARDS有害。
