基于机器学习的床旁评分IPA-2T2C2DW用于重症非中性粒细胞减少免疫抑制宿主侵袭性肺曲霉病的开发与验证
本前瞻性多中心队列研究纳入重症非中性粒细胞减少免疫抑制合并肺炎呼吸衰竭患者,开发并验证基于机器学习(ML)的侵袭性肺曲霉病(IPA)诊断模型,并衍生床旁评分IPA-2T2C2DW。推导队列508例(121例IPA),外部验证队列229例(53例IPA)。评分将患者分为低危(≤7分)、中危(8-9分)、高危(≥10分),观察IPA率分别为8.2%、28.6%、62.4%。
要点
- IPA-2T2C2DW评分可快速床旁分层IPA风险,指导监测强度与经验性抗真菌治疗。
- 推导队列中逻辑回归完整模型内部验证AUC 0.915(95%CI 0.850-0.980),Brier评分0.094。
- 外部验证中CatBoost的AUC为0.844(95%CI 0.760-0.928),简化版逻辑回归AUC 0.863。
核心结论
- IPA-2T2C2DW评分可实现快速床旁IPA风险分层,指导监测强度与经验性抗真菌治疗,适用于不同医疗环境。
- 在未经选择的队列中进行前瞻性验证之前,尚不能临床实施。
研究设计与样本
- 前瞻性多中心队列研究,推导队列来自中日友好医院ICU的508例免疫抑制肺炎合并呼吸衰竭患者(121例IPA)。
- 外部验证队列来自全国15家中心的229例患者(53例IPA)。
主要发现
- 鉴定出9个独立预测因子:CD4+ T细胞计数、特征性CT表现、干啰音、D-二聚体、体温>38°C、WBC、7天累积糖皮质激素剂量、糖尿病、淋巴细胞计数。
- 测试了三种分层模型与六种机器学习算法;内部验证中逻辑回归完整模型区分度最佳(AUC 0.915,95%CI 0.850-0.980),校准度Brier评分0.094。
- 外部验证中CatBoost的AUC为0.844(95%CI 0.760-0.928),逻辑回归简化版AUC 0.863,最小版AUC 0.791。
- IPA-2T2C2DW评分将患者分为低危(≤7分)、中危(8-9分)、高危(≥10分),观察IPA率分别为8.2%、28.6%、62.4%。
关键数值与推荐
- IPA-2T2C2DW评分低危≤7分,中危8-9分,高危≥10分。
- 低危组观察IPA率8.2%,中危组28.6%,高危组62.4%。
临床怎么用
- 可用于ICU入院24小时内快速识别IPA高危患者,指导监测强度与经验性抗真菌治疗决策。
- 不同医疗环境可选用完整版、简化版或最小版模型进行风险分层。
局限性
- 需在未经选择的队列中进行前瞻性验证后方可临床实施。
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
OBJECTIVE: To develop and validate machine learning-based diagnostic models for IPA using data available within 24 hours of ICU admission, construct three model versions for different healthcare settings, and derive a bedside score for rapid risk stratification. METHODS: This prospective multi-center cohort study enrolled immunocompromised patients with pneumonia complicated by respiratory failure admitted to the ICU at China-Japan Friendship Hospital (formed the derivation cohort) and 15 additional centers nationwide (comprised the external validation cohort). RESULTS: The derivation cohort included 508 patients (121 IPA) and the external validation cohort 229 patients (53 IPA). Nine independent predictors were identified: CD4+ T-cell count (T), characteristic CT findings (C), dry rales (D), D-dimer, Temperature above 38°C (T), WBC (W), 7-day cumulative corticosteroid dose (C), diabetes mellitus (D), and lymphocyte count. We tested three hierarchical models with six machine learning algorithms. In internal validation, logistic regression showed the best discrimination (AUC 0.915, 95% CI 0.850-0.980) and calibration (Brier score 0.094) for the completed model. In external validation, CatBoost achieved an AUC of 0.844 (0.760-0.928), and logistic regression remained robust in simplified (AUC 0.863) and minimal versions (AUC 0.791). The IPA-2T2C2DW score stratified patients into low- (≤7 points), intermediate- (8-9 points), and high-risk (≥10 points) groups, with observed IPA rates of 8.2%, 28.6%, and 62.4%, respectively. CONCLUSIONS: The IPA-2T2C2DW score enables rapid bedside IPA risk stratification to guide monitoring intensity and empirical antifungal therapy across diverse healthcare settings. Prospective validation in unselected cohorts is warranted before clinical implementation.
引用
- Huang L, Rong R, Zhou X, et al. Development and Validation of a Machine Learning-Based Bedside Score (IPA-2T2C2DW) for Invasive Pulmonary Aspergillosis in Critically Ill Non-Neutropenic Immunocompromised Hosts. J Infect Dis. 2026. PMID: 42583797. doi:10.1093/infdis/jiag404
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。