2型炎症相关生物标志物在成人哮喘急性发作风险预测模型中的应用：一项系统评价。
全文（开放获取） · 中文译文
摘要 目的：系统评估2型（T2）炎症相关生物标志物如何被纳入用于预测成人哮喘患者未来急性发作的多变量模型中，并评估模型性能、验证、校准、偏倚风险及临床适用性。 方法：检索PubMed、Embase、Web of Science、Cochrane图书馆及Scopus数据库，检索时间从建库至2026年5月，随后通过PubMed补充更新至2026年7月31日。符合条件的报告须包含至少一个用于开发、验证或更新未来急性发作结局纵向多变量模型的成人队列，纳入至少一种T2相关生物标志物，并报告模型性能或验证结果。混合年龄队列不计入独立的成人证据；仅在成人队列中评估的性能纳入主要汇总分析。仅关联性研究、当前发作分类模型、发作后处置模型，以及预先指定基线风险背景之外的指数后复发模型均被排除。数据提取遵循CHARMS，偏倚风险及适用性采用PROBAST进行评估。 结果：共纳入8篇发表于2021年至2026年的报告。符合条件的成人队列规模为540至1，331，934名参与者；其中一篇报告还纳入了混合年龄的随机试验对照组（n = 345）。建模方法包括logistic回归和弹性网络回归、梯度提升和深度学习算法、贝叶斯网络，以及混合效应计数模型。嗜酸性粒细胞相关变量出现在全部8篇报告中，但并非在每项最终模型中均被保留；FeNO被纳入两项跨国重症哮喘注册研究模型中，血清IgE在一个跨来源分析中作为网络节点出现，并在一个独立复制的代谢组学模型中被保留。代谢组学模型在发现队列和复制队列中报告的AUC分别为0.901和0.893，而源自随机对照试验的贝叶斯网络在成人ISAR队列中验证时AUC为0.68。6篇报告被评定为总体偏倚风险较高，主要原因是分析领域存在局限性。 结论：T2相关生物标志物可能为成人哮喘急性发作风险模型增加补充信息，但现有证据尚不能确定BEC、FeNO或IgE为独立充分的预测因子，也不支持在常规实践中采用现有模型。未来研究应量化生物标志物在已确立的临床预测因素之外的增量价值，并优先考虑足够的样本量、外部验证、校准、临床实用性、对缺失生物标志物数据的透明处理，以及可复现的模型呈现方式。 系统评价注册：https://www.crd.york.ac.uk/prospero/，PROSPERO CRD420261415469。 引言 哮喘是一种异质性慢性气道疾病，以可变的呼吸道症状和气流受限为特征（）。急性加重是临床上重要的事件，可能需要全身性糖皮质激素、急诊治疗、住院或重症监护，并显著增加发病率和医疗资源使用（,）。准确的风险评估可支持针对性随访、依从性干预、优化抗炎治疗，以及及时转诊至专科医生。 T2炎症是哮喘中一种突出的炎症内型，涉及IL−4、IL-5和IL-13信号传导、嗜酸性粒细胞炎症以及IgE相关反应（,）。血嗜酸性粒细胞计数（BEC）和呼出气一氧化氮分数（FeNO）是反映该生物学过程最容易获得的临床指标之一，而总IgE和骨膜蛋白在一般急性加重风险评估中的作用则较不一致（）。生物标志物的解读依赖于具体情境，因为其数值可能随糖皮质激素暴露、感染、过敏合并症、吸烟、治疗依从性和测量时间而变化。因此，在异质性哮喘人群中，单次测量或固定阈值可能不足以用于个体化风险评估。 预后预测模型不同于生物标志物关联性研究。它结合多个预测因子来估计个体未来发生某种结局的概率或速率，并应报告区分度、校准度、验证或临床实用性（）。校正后的比值比或风险比可能显示预后关联性，但未必能产生可移植的个体风险工具。既往系统评价综合了基层医疗及更广泛成人人群中哮喘急性发作预测模型的研究（，），但并未专门评估T2相关生物标志物的测量、建模、报告及增量价值。因此，本综述聚焦于纳入至少一种T2相关生物标志物的成人多因素急性发作风险模型，并检查其表观性能及方法学可信度。 材料与方法 报告指南与注册 本系统评价依照PRISMA 2020（）进行，且方案已在PROSPERO注册（CRD420261415469）。在完成更新版筛选之前，对纳入标准框架进行了操作性澄清，以区分个体化预后模型与仅作关联性分析的研究，并明确预设的基线风险设定；该澄清已作为方案修订予以记录。 纳入标准 符合纳入标准的报告需满足以下条件：（1）包含年龄在18岁及以上、已确诊哮喘的成人队列，或可提取的纯成人开发或验证结果；（2）采用纵向设计，预测因子在目标结局发生之前测量；（3）开发、验证或更新了一个针对未来哮喘急性发作、严重急性发作、哮喘相关急诊就诊或住院、或接受全身性糖皮质激素治疗的发作的多变量模型，且预测起点与目标事件不同，并非在同一未缓解的急性发作期间开始；（4）纳入至少一种T2相关生物标志物作为预测因子、网络节点或预设分层变量，包括血嗜酸性粒细胞计数（BEC）/绝对嗜酸性粒细胞计数（AEC）、呼出气一氧化氮（FeNO）、免疫球蛋白E（IgE）、骨膜蛋白，或明确定义的T2生物标志物组合；（5）报告了区分度、校准度、分类性能、内部或外部验证，或临床净获益。原型和事后分析工具仅在成人个体水平数据产生了具有性能评估的个体化多变量风险估计时才可纳入。含有合格成人队列的报告可包括混合年龄辅助队列。仅成人队列中评估的性能纳入主要综合结果；仅在混合年龄队列中评估的结果作为背景信息，不计入独立的成人证据。 如果报告仅报告生物标志物与结局的关联，而未提供个体化的多变量预测输出；仅评估单一生物标志物阈值；检测或分类当前急性加重；预测严重程度、即时处置、住院、治疗失败或因同一未缓解的急性加重而短期返诊；以指标性急性加重作为预测起点，并针对预设基线风险情境之外的后续复发复合终点；预测新发哮喘；缺少符合条件的T2生物标志物；未提供可提取的成人发育或验证证据；或属于横断面研究、会议摘要、综述、评论、病例报告或基础研究，则排除该报告。无个体化验证工具的汇总亚组风险量表及事后关联分析亦被排除。此操作性边界总结如下。 合格预后模型与所排除的关联或分类研究之间的操作性边界。 HR，风险比；OR，比值比；RR，风险比；T2，2型。混合年龄队列不计入独立的成人证据；仅在成人队列中评估的性能贡献于主要综合分析。 信息来源与检索策略 检索了PubMed、Embase、Web of Science、Cochrane Library和Scopus数据库，检索时间从各数据库建库至2026年7月31日。检索概念涵盖哮喘、未来急性发作相关结局、T2相关生物标志物以及预测模型术语。针对每个数据库调整了受控词汇和自由词，并对纳入研究的参考文献列表及相关系统评价进行了筛检。检索报告遵循PRISMA-S原则（）。更新的PubMed检索策略见。 PubMed数据库检索策略。 AUC，受试者工作特征曲线下面积；FeNO，呼出气一氧化氮分数；MeSH，医学主题词表。补充的PubMed检索于2026年7月31日完成。 研究筛选与数据提取。 两位评审员独立筛选标题、摘要和全文，分歧通过讨论或与第三位评审员协商解决。数据提取遵循CHARMS（）并涵盖研究设置、人群、样本量、事件频率、结局定义、预测时间范围、候选及最终预测因子、T2-生物标志物的定义与时间点、缺失数据处理、变量选择、算法、模型呈现、验证、区分度、校准度、分类指标、决策曲线分析及研究局限性。当报告多个模型时，优先提取与预先指定的未来加重结局最相关的模型以及最独立的验证结果。对于包含仅成人及混合年龄队列的报告，仅在成人队列中评估的结果纳入主要汇总分析；仅在混合年龄队列中评估的结果则作为背景可迁移性证据予以描述。 偏倚风险和适用性 两名评价者使用PROBAST独立评估偏倚风险和适用性。偏倚风险在参与者、预测因子、结局和分析域中进行评估；适用性在参与者、预测因子和结局域中进行评估。当所有相关信号问题均回答为“是”或“可能是”时，该域评为低风险；当至少一个关键条目回答为“否”或“可能否”时，评为高风险；当报告不足以作出判断时，评为不明确。分析域考虑了有效样本量、预测因子处理、缺失数据、类别不平衡、过拟合、数据泄漏、校准和验证的独立性。 综合 由于人群、结局、预测时间范围、生物标志物定义、算法、事件患病率和性能指标存在显著异质性，采用结构化叙述性综合。未对区分度估计值进行合并。性能与校准、验证、事件患病率、报告透明度和偏倚风险共同解读，而非仅根据AUC对研究进行排名（）。 结果 研究选择 研究选择过程见。去重、标题和摘要筛选以及全文评估后，八篇符合条件的预测模型报告被纳入定性综合（–）。 研究选择的PRISMA 2020流程图。 纳入研究的特征 纳入的八篇报告发表于2021年至2026年，涵盖了门诊电子健康记录队列、基层医疗人群、基于生物标志物分层的临床队列、多国家重度哮喘登记研究、随机试验/真实世界比较研究，以及采用靶向代谢组学的成人生物样本库队列。符合条件的成人队列规模从540人到1,331,934人不等，预测时限从6个月到5年，最常见为12个月。嗜酸性粒细胞相关信息出现在全部八篇报告中，但并非在所有最终模型中均被保留。FeNO被纳入Yadav 2025和Chen W 2026的登记研究模型。Yadav 2026比较了来自DREAM/MENSA安慰剂组的345名年龄超过12岁的参与者与ISAR中6,814名未接受生物制剂治疗的成人；在成人ISAR中对源自随机试验的模型进行验证为主要综合结果，而在混合年龄试验队列中的反向验证仅作为背景参考。Yadav 2025、Yadav 2026和Chen W 2026部分使用了重叠的ISAR人群，但探讨了不同的问题和验证框架。Chen Y 2026在540名成人中开发了5年预测模型，并在823名成人中进行了重复验证。总结了纳入研究的特征。 纳入研究的特征。 A&E，意外与急诊；AEC，绝对嗜酸性粒细胞计数；BEC，血嗜酸性粒细胞计数；BN，贝叶斯网络；ED，急诊科；EHR/EMR，电子健康/医疗记录；ICS，吸入性糖皮质激素；ISAR，国际严重哮喘登记处；OCS，口服糖皮质激素。在Yadav 2026中，成人ISAR中的验证对主要综合结果有所贡献；在混合年龄RCT队列中的反向验证则属于背景性内容。 模型开发与验证 建模方法包括逻辑回归、LightGBM及相关机器学习方法、XGBoost和深度学习架构、基于Lasso或弹性网络的变量选择、贝叶斯网络，以及混合效应零膨胀负二项模型。缺失数据处理方式各异。Tibble等人将BEC表示为高、低或缺失；Yadav 2025使用随机森林插补；Yadav 2026使用MissForest插补、基于临床信息的离散化处理和自适应合成抽样；Chen Y 2026则使用完整病例的临床输入，并将比值构建限制在缺失率低于30%的代谢物上。多项EHR研究未充分量化选择性检测生物标志物的影响。验证方法涵盖同源数据划分或交叉验证，以及时间、国家分层、内部-外部、独立队列和跨源验证等（）方法。 模型开发与验证特征。 ADASYN，自适应合成抽样；BEC，血嗜酸性粒细胞计数；BN，贝叶斯网络；EHR，电子健康记录；LightGBM，轻量级梯度提升机；LSTM，长短期记忆；ML，机器学习；RESA，重度哮喘急性发作风险；SHAP，沙普利可加解释；XGBoost，极端梯度提升。 模型性能与主要预测因子 区分度差异很大。Zein 等人报告，LightGBM 对于非重度急性加重、急诊就诊和住院的 AUC 分别约为 0.71、0.88 和 0.85。Turcatel 等人报告，经内部测试的 XGBoost AUC 为 0.964，事件发生率为 1.2%。Tibble 等人报告的 AUC 约为 0.75，校准斜率为 1.02，截距为 0.004。Pham 等人报告，高血嗜酸性粒细胞（BEC）组的 AUC 为 0.84–0.85，低 BEC 组的 AUC 为 0.93–0.94，未进行校准或外部验证。在 Yadav 2026 的研究中，源自随机对照试验（RCT）的贝叶斯网络在成人 ISAR 队列中的 AUC 为 0.68；反向验证在混合年龄 RCT 队列中产生的 AUC 为 0.50，并被视作背景证据。源自成人 ISAR 的网络显示出接近理想的校准，对于至少一次和至少两次重度急性加重，截距分别为 0.02 和 −0.002，斜率分别为 0.95 和 1.02。Chen Y 2026 报告，对于结合鞘脂与类固醇比值、血清 IgE、种族和吸入性皮质类固醇使用的模型，发现队列和独立复制队列的 AUC 分别为 0.901 和 0.893。Yadav 2025 报告，交叉验证的 AUC 为 0.65，国家分层验证的 AUC 为 0.62，并提供了校准估计值。Chen W 2026 报告，对于至少一次和至少两次重度急性加重，合并的时间依赖性 AUC 分别为 0.63 和 0.68，并进行了校准和决策曲线分析。 在各模型中，常见的临床预测因子包括既往急性发作、哮喘控制情况、缓解药物或全身性皮质类固醇的使用、肺功能、吸烟、合并症以及治疗强度。T2相关生物标志物通常是补充而非替代这些变量。Chen Y 2026 是唯一提供消融式T2生物标志物比较的报道：去除血清IgE使发现队列的AUC从0.901降至0.793，验证队列的AUC从0.893降至0.724。由于这不是针对标准临床基础模型的预设比较，且使用的是完整病例，因此不应将其解释为确定的增量价值。没有研究直接分离出BEC或FeNO在预设临床模型之外的增量价值()。 纳入模型的性能与解读。 AUC，受试者工作特征曲线下面积；BEC，血嗜酸性粒细胞计数；BN，贝叶斯网络；DCA，决策曲线分析；ED，急诊科；IgE，免疫球蛋白E；ISAR，国际重症哮喘登记处；OCS，口服皮质类固醇；PPV，阳性预测值；RCT，随机对照试验。由于人群、结局、预测时间窗和验证设计不同，各估计值不可直接比较。 T2炎症相关生物标志物的应用 嗜酸性粒细胞相关变量是最常见的T2生物标志物输入，在Zein等人研究中作为候选EHR特征，在Turcatel等人研究中编码为SHAP解释特征，在Tibble等人研究中作为分类预测变量，在Pham等人研究中作为分层变量和预测变量，在Yadav 2025和Yadav 2026研究中作为贝叶斯网络节点，在Chen W 2026研究中作为直接预测变量，在Chen Y 2026研究中作为比较模型变量。FeNO被纳入Yadav 2025和Chen W 2026的重症哮喘登记模型。血清总IgE在多项报告中作为候选输入或网络节点被纳入，并在表现最佳的Chen Y 2026模型中被保留。在Yadav 2026中，学习到的网络包含从IgE到BEC以及从BEC到未来严重发作的弧；这些概率依赖性并不建立因果关系。FeNO被排除，因为在一项试验中不可用。Periostin未被纳入符合条件的模型。 Turcatel的报告需要谨慎解读，因为其方法和SHAP分析描述了嗜酸性粒细胞实验室编码，而讨论部分指出常规使用的T2生物标志物（包括BEC和FeNO）的可用性不完整。Chen Y 2026也需要谨慎对待：在消融比较中观察到添加血清IgE后判别能力提高，但IgE在发现队列中缺失67.2%，在验证队列中缺失56.9%，最终分析使用了完整病例。这些问题被视为关于透明度、选择性和适用性的担忧，而非临床生物标志物效用的确定性证据。 T2生物标志物在合格预测模型中的作用。 AEC，绝对嗜酸性粒细胞计数；BEC，血嗜酸性粒细胞计数；FeNO，呼出气一氧化氮分数；IgE，免疫球蛋白E；T2，2型。纳入此表并不表示具有独立的增量价值或临床就绪性。 偏倚风险与适用性 偏倚风险集中在分析领域。常见问题包括同源内部验证、校准报告不完整、生物标志物稀疏或选择性测量、缺失数据或类别不平衡处理不明确、完整病例分析、灵活的特征选择，以及生物标志物增量价值评估有限。Zein、Turcatel、Tibble、Pham、Yadav 2026和Chen Y 2026被评定为总体偏倚风险较高。对于Yadav 2026，问题包括混合年龄随机对照试验（RCT）对照、RCT队列中事件信息较少、大量插补和离散化处理，以及合成过采样。对于Chen Y 2026，问题包括完整病例分析、以口服皮质类固醇（OCS）处方作为急性发作的替代指标、缺乏校准，以及最终模型样本量报告不完整。Yadav 2025的验证和校准报告相对完整，而Chen W 2026显示出最完整的验证概况，包括内部-外部交叉验证、校准和临床净获益。适用性对于目标成人人群总体可接受，尽管重度哮喘登记处和代谢组学模型可能无法直接推广到非选择性初级保健中。 PROBAST偏倚风险与适用性评估。 RoB，偏倚风险。总体RoB反映PROBAST的四个领域。对于Yadav 2026，成人ISAR中的验证为主要综合分析提供了依据；混合年龄RCT队列为背景可迁移性及报告层面的参与者判定提供了信息。 讨论 本系统评价识别出8项具有可提取成人证据、纳入T2相关生物标志物的预后模型报告。主要发现并非BEC、FeNO或IgE构成了经过验证的独立解决方案，而是这些生物标志物日益被整合进更广泛的模型中，这些模型还同时使用了既往急性加重史、哮喘控制情况、药物使用、肺功能、合并症，以及
全文（开放获取） · 英文原文
【Abstract】 OBJECTIVE: To systematically evaluate how type 2 (T2) inflammation-related biomarkers are incorporated into multivariable models that predict future acute exacerbations in adults with asthma, and to appraise model performance, validation, calibration, risk of bias, and clinical applicability. METHODS: PubMed, Embase, Web of Science, the Cochrane Library, and Scopus were searched from inception to May 2026, followed by a supplementary PubMed update through 31 July 2026. Eligible reports contained at least one adult cohort used to develop, validate, or update a longitudinal multivariable model for a future exacerbation outcome, incorporated at least one T2-related biomarker, and reported model performance or validation. Mixed-age cohorts were not counted as independent adult evidence; only performance evaluated in an adult cohort contributed to the main synthesis. Association-only studies, current-attack classification models, post-exacerbation disposition models, and post-index relapse models outside the prespecified baseline-risk setting were excluded. Data extraction followed CHARMS, and risk of bias and applicability were assessed using PROBAST. RESULTS: Eight reports published from 2021 to 2026 were included. Eligible adult cohorts ranged from 540 to 1,331,934 participants; one report also included a mixed-age randomized-trial comparator (n = 345). Modeling approaches included logistic and elastic-net regression, gradient-boosting and deep-learning algorithms, Bayesian networks, and a mixed-effects count model. Eosinophil-related variables appeared in all eight reports but were not retained in every final model; FeNO was included in two multinational severe-asthma registry models, and serum IgE appeared as a network node in one cross-source analysis and was retained in an independently replicated metabolomics model. The metabolomics model reported AUCs of 0.901 and 0.893 in discovery and replication cohorts, whereas an RCT-derived Bayesian network achieved an AUC of 0.68 when validated in the adult ISAR cohort. Six reports were rated at high overall risk of bias, mainly because of analysis-domain limitations. CONCLUSION: T2-related biomarkers may add complementary information to adult asthma exacerbation-risk models, but the available evidence does not yet establish BEC, FeNO, or IgE as independently sufficient predictors or support routine implementation of an existing model. Future studies should quantify incremental biomarker value beyond established clinical predictors and prioritize adequate sample size, external validation, calibration, clinical utility, transparent handling of missing biomarker data, and reproducible model presentation. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/prospero/, PROSPERO CRD420261415469. 【Introduction】 Asthma is a heterogeneous chronic airway disease characterized by variable respiratory symptoms and airflow limitation (). Acute exacerbations are clinically important events that may require systemic corticosteroids, emergency care, hospitalization, or intensive care, and contribute substantially to morbidity and health-care use (, ). Accurate risk estimation could support targeted follow-up, adherence interventions, optimization of anti-inflammatory therapy, and timely specialist referral. T2 inflammation is a prominent inflammatory endotype in asthma involving IL−4, IL-5, and IL-13 signaling, eosinophilic inflammation, and IgE-associated responses (, ). Blood eosinophil count (BEC) and fractional exhaled nitric oxide (FeNO) are among the most accessible clinical indicators of this biology, whereas total IgE and periostin have less consistent roles in general exacerbation-risk assessment (). Biomarker interpretation is context dependent because values may vary with corticosteroid exposure, infection, allergic comorbidity, smoking, treatment adherence, and measurement timing. A single measurement or fixed threshold may therefore be insufficient for individualized risk estimation across heterogeneous asthma populations. A prognostic prediction model differs from a biomarker-association study. It combines multiple predictors to estimate an individual's future probability or rate of an outcome and should report discrimination, calibration, validation, or clinical utility (). An adjusted odds ratio or hazard ratio may show prognostic association without yielding a transportable individual-risk tool. Previous systematic reviews synthesized asthma-attack prediction models in primary care and broader adult populations (, ), but did not specifically evaluate the measurement, modeling, reporting, and incremental value of T2-related biomarkers. This review therefore focuses on adult multivariable exacerbation-risk models that incorporate at least one T2-related biomarker and examines apparent performance alongside methodological credibility. 【Materials and methods】 【Reporting guideline and registration】 This systematic review was conducted in accordance with PRISMA 2020 (), and the protocol was registered in PROSPERO (CRD420261415469). Before completion of the updated screening, the eligibility framework was operationally clarified to distinguish individualized prognostic models from association-only studies and to define the prespecified baseline-risk setting; this clarification was documented as a protocol amendment. 【Eligibility criteria】 Reports were eligible when they: (1) contained an adult cohort aged 18 years or older with established asthma, or extractable adult-only development or validation results; (2) used a longitudinal design in which predictors were measured before the target outcome; (3) developed, validated, or updated a multivariable model for a future asthma exacerbation, severe exacerbation, asthma-related emergency visit or hospitalization, or systemic-corticosteroid-treated attack, with prediction initiated from a baseline distinct from the target event and not during the same unresolved exacerbation; (4) incorporated at least one T2-related biomarker, including BEC/AEC, FeNO, IgE, periostin, or an explicitly defined T2-biomarker construct, as a predictor, network node, or prespecified stratification variable; and (5) reported discrimination, calibration, classification performance, internal or external validation, or clinical net benefit. Prototype and post hoc tools were eligible only when adult participant-level data produced individualized multivariable risk estimates with performance evaluation. Reports with an eligible adult cohort could include a mixed-age auxiliary cohort. Only performance evaluated in an adult cohort contributed to the main synthesis; results evaluated solely in a mixed-age cohort were contextual and were not counted as independent adult evidence. Reports were excluded when they reported biomarker-outcome associations without an individualized multivariable prediction output; evaluated only a single biomarker threshold; detected or classified a current exacerbation; predicted severity, immediate disposition, hospitalization, treatment failure, or short-term return for the same unresolved exacerbation; used an index exacerbation as the prediction origin and targeted a subsequent relapse composite outside the prespecified baseline-risk setting; predicted incident asthma; lacked an eligible T2 biomarker; provided no extractable adult development or validation evidence; or were cross-sectional studies, conference abstracts, reviews, commentaries, case reports, or basic research. Aggregate subgroup risk scales and post hoc association analyses without an individualized validated tool were also excluded. summarizes this operational boundary. Operational boundary between eligible prognostic models and excluded association or classification studies. HR, hazard ratio; OR, odds ratio; RR, risk ratio; T2, type 2. Mixed-age cohorts were not counted as independent adult evidence; only performance evaluated in an adult cohort contributed to the main synthesis. 【Information sources and search strategy】 PubMed, Embase, Web of Science, the Cochrane Library, and Scopus were searched from database inception through 31 July 2026. Search concepts covered asthma, future exacerbation-related outcomes, T2-related biomarkers, and prediction-model terminology. Controlled vocabulary and free-text terms were adapted for each database, and reference lists of included studies and relevant systematic reviews were screened. Search reporting followed PRISMA-S principles (). The updated PubMed strategy is presented in . Search strategy for the pubMed database. AUC, area under the receiver operating characteristic curve; FeNO, fractional exhaled nitric oxide; MeSH, Medical Subject Headings. The supplementary PubMed search was completed on 31 July 2026. 【Study selection and data extraction】 Two reviewers independently screened titles, abstracts, and full texts, with disagreements resolved by discussion or consultation with a third reviewer. Data extraction followed CHARMS () and covered study setting, population, sample size, event frequency, outcome definition, prediction horizon, candidate and final predictors, T2-biomarker definition and timing, missing-data handling, variable selection, algorithm, model presentation, validation, discrimination, calibration, classification measures, decision-curve analysis, and study limitations. When multiple models were reported, the model most relevant to the prespecified future-exacerbation outcome and the most independent validation result were preferentially extracted. For reports containing adult-only and mixed-age cohorts, only results evaluated in an adult cohort contributed to the main synthesis; results evaluated solely in a mixed-age cohort were described as contextual transportability evidence. 【Risk of bias and applicability】 Two reviewers independently assessed risk of bias and applicability using PROBAST (). Risk of bias was evaluated across the participants, predictors, outcome, and analysis domains; applicability was assessed across the participants, predictors, and outcome domains. A domain was rated as low risk when all relevant signaling questions were answered as yes or probably yes, high risk when at least one critical item was answered as no or probably no, and unclear when reporting did not permit a judgment. The analysis domain considered effective sample size, predictor handling, missing data, class imbalance, overfitting, data leakage, calibration, and independence of validation. 【Synthesis】 A structured narrative synthesis was used because populations, outcomes, prediction horizons, biomarker definitions, algorithms, event prevalence, and performance metrics were substantially heterogeneous. Discrimination estimates were not pooled. Performance was interpreted jointly with calibration, validation, event prevalence, reporting transparency, and risk of bias rather than by ranking studies according to AUC alone (). 【Results】 【Study selection】 The study-selection process is shown in . After deduplication, title and abstract screening, and full-text assessment, eight eligible prediction-model reports were included in the qualitative synthesis (–). PRISMA 2020 flow diagram of study selection. 【Characteristics of included studies】 The eight included reports, published from 2021 to 2026, represented outpatient EHR cohorts, primary-care populations, biomarker-stratified clinical cohorts, multinational severe-asthma registries, a randomized-trial/real-world comparison, and adult biobank cohorts with targeted metabolomics. Eligible adult cohorts ranged from 540 to 1,331,934 participants, and prediction horizons ranged from 6 months to 5 years, most commonly 12 months. Eosinophil-related information appeared in all eight reports but was not retained in every final model. FeNO was included in the Yadav 2025 and Chen W 2026 registry models. Yadav 2026 compared 345 participants aged over 12 years from DREAM/MENSA placebo arms with 6,814 biologic-naive adults in ISAR; validation of the RCT-derived model in adult ISAR contributed to the main synthesis, whereas reverse validation in the mixed-age trial cohort was contextual. Yadav 2025, Yadav 2026, and Chen W 2026 partly used overlapping ISAR populations but addressed distinct questions and validation frameworks. Chen Y 2026 developed a 5-year model in 540 adults and replicated it in 823 adults. summarizes study characteristics. Characteristics of included studies. A&E, accident and emergency; AEC, absolute eosinophil count; BEC, blood eosinophil count; BN, Bayesian network; ED, emergency department; EHR/EMR, electronic health/medical record; ICS, inhaled corticosteroid; ISAR, International Severe Asthma Registry; OCS, oral corticosteroid. In Yadav 2026, validation in adult ISAR contributed to the main synthesis; reverse validation in the mixed-age RCT cohort was contextual. 【Model development and validation】 Modeling approaches included logistic regression, LightGBM and related machine-learning methods, XGBoost and deep-learning architectures, Lasso- or elastic-net-based selection, Bayesian networks, and a mixed-effects zero-inflated negative binomial model. Missing-data handling varied. Tibble et al. represented BEC as high, low, or missing; Yadav 2025 used random-forest imputation; Yadav 2026 used MissForest imputation, clinically informed discretization, and adaptive synthetic sampling; and Chen Y 2026 used complete-case clinical inputs and restricted ratio construction to metabolites with less than 30% missingness. Several EHR studies did not fully quantify the implications of selectively measured biomarkers. Validation ranged from same-source splits or cross-validation to temporal, country-stratified, internal-external, independent-cohort, and cross-source approaches (). Model development and validation characteristics. ADASYN, adaptive synthetic sampling; BEC, blood eosinophil count; BN, Bayesian network; EHR, electronic health record; LightGBM, light gradient-boosting machine; LSTM, long short-term memory; ML, machine learning; RESA, Risk of Exacerbation in Severe Asthma; SHAP, Shapley additive explanations; XGBoost, extreme gradient boosting. 【Model performance and principal predictors】 Discrimination varied substantially. Zein et al. reported LightGBM AUCs of approximately 0.71 for nonsevere exacerbations, 0.88 for emergency department visits, and 0.85 for hospitalization. Turcatel et al. reported an internally tested XGBoost AUC of 0.964 with an event prevalence of 1.2%. Tibble et al. reported an AUC of approximately 0.75, with a calibration slope of 1.02 and intercept of 0.004. Pham et al. reported AUCs of 0.84–0.85 in the high-BEC group and 0.93–0.94 in the low-BEC group, without calibration or external validation. In Yadav 2026, the RCT-derived Bayesian network achieved an AUC of 0.68 in the adult ISAR cohort; the reverse validation produced an AUC of 0.50 in the mixed-age RCT cohort and was treated as contextual evidence. The adult ISAR-derived network showed near-ideal calibration, with intercepts of 0.02 and −0.002 and slopes of 0.95 and 1.02 for at least one and at least two severe exacerbations. Chen Y 2026 reported AUCs of 0.901 and 0.893 in discovery and independent replication for a model combining sphingolipid-to-steroid ratios, serum IgE, race, and inhaled corticosteroid use. Yadav 2025 reported AUCs of 0.65 in cross-validation and 0.62 in country-stratified validation, with calibration estimates. Chen W 2026 reported pooled time-dependent AUCs of 0.63 for at least one and 0.68 for at least two severe exacerbations, with calibration and decision-curve analysis. Across models, commonly represented clinical predictors included previous exacerbations, asthma control, reliever or systemic corticosteroid use, lung function, smoking, comorbidities, and treatment intensity. T2-related biomarkers generally supplemented rather than replaced these variables. Chen Y 2026 was the only report to provide an ablation-style T2-biomarker comparison: removing serum IgE reduced AUC from 0.901 to 0.793 in discovery and from 0.893 to 0.724 in replication. Because this was not a prespecified comparison with a standard clinical base model and used complete cases, it should not be interpreted as definitive incremental value. No study directly isolated the incremental value of BEC or FeNO beyond a prespecified clinical model (). Performance and interpretation of included models. AUC, area under the receiver operating characteristic curve; BEC, blood eosinophil count; BN, Bayesian network; DCA, decision-curve analysis; ED, emergency department; IgE, immunoglobulin E; ISAR, International Severe Asthma Registry; OCS, oral corticosteroid; PPV, positive predictive value; RCT, randomized controlled trial. Estimates are not directly comparable because populations, outcomes, horizons, and validation designs differed. 【Application of T2 inflammation-related biomarkers】 Eosinophil-related variables were the most common T2-biomarker input, serving as candidate EHR features in Zein et al., encoded SHAP-interpreted features in Turcatel et al., a categorized predictor in Tibble et al., a stratification variable and predictor in Pham et al., Bayesian-network nodes in Yadav 2025 and Yadav 2026, direct predictors in Chen W 2026, and comparator-model variables in Chen Y 2026. FeNO was included in the Yadav 2025 and Chen W 2026 severe-asthma registry models. Total serum IgE was included as a candidate input or network node in several reports and retained in the best-performing Chen Y 2026 model. In Yadav 2026, the learned network contained an arc from IgE to BEC and from BEC to future severe exacerbations; these probabilistic dependencies do not establish causality. FeNO was excluded because it was unavailable in one trial. Periostin was not incorporated in an eligible model. The Turcatel report requires cautious interpretation because its methods and SHAP analyses describe eosinophil laboratory encoding, whereas the discussion notes incomplete availability of routinely used T2 biomarkers, including BEC and FeNO. Chen Y 2026 also warrants caution: higher discrimination after adding serum IgE was observed in an ablation comparison, but IgE was missing in 67.2% of the discovery cohort and 56.9% of the replication cohort, and the final analysis used complete cases. These issues were treated as concerns about transparency, selection, and applicability rather than definitive evidence of clinical biomarker utility (). Role of T2 biomarkers in eligible prediction models. AEC, absolute eosinophil count; BEC, blood eosinophil count; FeNO, fractional exhaled nitric oxide; IgE, immunoglobulin E; T2, type 2. Inclusion in this table does not imply independent incremental value or clinical readiness. 【Risk of bias and applicability】 Risk of bias was concentrated in the analysis domain. Common concerns included same-source internal validation, incomplete calibration reporting, sparse or selectively measured biomarkers, unclear missing-data or class-imbalance handling, complete-case analysis, flexible feature selection, and limited assessment of incremental biomarker value. Zein, Turcatel, Tibble, Pham, Yadav 2026, and Chen Y 2026 were rated at high overall risk of bias. For Yadav 2026, concerns included the mixed-age RCT comparator, low event information in the RCT cohort, extensive imputation and discretization, and synthetic oversampling. For Chen Y 2026, concerns included complete-case analysis, an OCS-prescription proxy for exacerbations, absent calibration, and incomplete reporting of final model sample sizes. Yadav 2025 had comparatively complete validation and calibration reporting, while Chen W 2026 showed the most complete validation profile, including internal-external cross-validation, calibration, and clinical net benefit. Applicability was generally acceptable for the intended adult populations, although severe-asthma registry and metabolomics models may not transport directly to unselected primary care (). PROBAST risk-of-bias and applicability assessment. RoB, risk of bias. Overall RoB reflects the four PROBAST domains. For Yadav 2026, validation in adult ISAR contributed to the main synthesis; the mixed-age RCT cohort informed contextual transportability and the report-level participant judgment. 【Discussion】 This systematic review identified eight prognostic-model reports with extractable adult evidence that incorporated T2-related biomarkers. The main finding is not that BEC, FeNO, or IgE constitutes a validated stand-alone solution, but that these biomarkers are increasingly embedded in broader models that also use previous exacerbations, asthma control, medication use, lung function, comorbidities, and,