2026/01/24 by Chandra Prakash Yadav, Laura Huey Mien Lim, David Price +37 · 2 voices
Computer Science · Medicine · Environmental Science · #Machine Learning in Healthcare #Asthma and respiratory diseases #Health, Environment, Cognitive Aging
paper · doi:10.1016/j.chest.2026.01.009
openalex publication_date 2026/01/24 · openalex created_date 2026/01/25 · openalex updated_date 2026/07/22
BACKGROUND: The way in which risk predictors combine and contribute to severe asthma exacerbations may differ between clinical trials and real-world settings. RESEARCH QUESTION: How do the interactive pathways of risk predictors leading to severe asthma exacerbations compare under clinical trials vs real-world settings? STUDY DESIGN AND METHODS: The analysis involved 345 patients with severe asthma from the placebo arms of 2 international randomized controlled trials (RCTs), compared with 6,814 biologic-naïve patients from the International Severe Asthma Registry (ISAR). Sixteen key risk predictors, including demographics, biomarkers, lung function, health care use, exacerbation history, long-term oral corticosteroid use, asthma control, and nasal polyps, were covered. The outcome was the occurrence of severe asthma exacerbations over the 365 days after study enrollment. Bayesian networks (BNs), obtained from machine learning combined with expert knowledge, elucidated significant interplay processes of risk predictors that led to severe asthma exacerbations. External validation was performed in each cohort. RESULTS: The RCTs revealed 44 significant arcs (ie, probabilistic interdependency) between 16 risk factors, whereas the ISAR showed 170. Despite this difference, the main downstream prediction pathways were consistent across both settings, with 2 key pathways: total serum IgE level influenced blood eosinophils to predict future severe exacerbations, and severe exacerbation history directly predicted future severe exacerbations. In external validation, RCT-BN generalized well to ISAR patients (area under the receiver operating characteristic curve, 0.68), whereas ISAR-BN underperformed in RCT patients (area under the receiver operating characteristic curve, 0.50), and ISAR-BN demonstrated better calibration. INTERPRETATION: Our results show that the core pathways predicting severe asthma exacerbations were similar in both RCTs and real-world settings, with comparable predictive performance.