Logo image
Back
B110-02 Combining Quantitative CT, CFD, and Clinical Data Using Machine Learning to Identify Severe Asthma Subtypes
Abstract   Peer reviewed

B110-02 Combining Quantitative CT, CFD, and Clinical Data Using Machine Learning to Identify Severe Asthma Subtypes

J Choi, A Abdolijomoor, D H Lee, J S Boomer, S Haworth, E A Hoffman, S B Fain, L C Denlinger, E Israel, N N Jarjour, …
American journal of respiratory and critical care medicine, Vol.212(Supplement_1), aamag162517
05/01/2026
DOI: 10.1093/ajrccm/aamag162.517

View Online

Abstract

Rationale Pathophysiologic characteristics of severe asthma subtypes are not fully understood. Computational fluid dynamics (CFD) analysis showed novel structural-functional characterization of asthma sub-types (PMID 30888242). We aimed to identify comprehensive phenotypic characterization of asthma, integrating quantitative computed tomography (qCT), CFD, and clinical data. Methods Inspiratory and expiratory CTs and clinical data were collected from 97 asthma patients in the Severe Asthma Research Program III. Commercial and in-house software were employed to conduct qCT and CFD airflow simulations (tidal volume: 6 ml/kg). Unsupervised machine learning approaches of principal component analysis (PCA) and k-means clustering were used to identify subgroups from multiple variables: 29 qCT, 10 CFD, and 27 clinical features. Kruskal-Wallis and post-hoc Dunn’s tests were used to compare clusters (statistical significance by p < 0.05). Then, 6 machine learning prediction models (Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, LightGB, and XGBoost) with majority rule and the Shapley additive explanations (SHAP) feature importance analysis identified the key discriminative features. Results Cluster analysis identified 5 distinct clusters: obese (N = 12), older (N = 14), early onset male (N = 19), airway thickened female (N = 27), and non-severe (N = 25) asthmatics. To highlight significant differences, the “obese” cluster showed the worst asthma control, the worst quality of life, widespread high attenuation area (HAA), and segmental airway narrowing along with excessive CFD-based airway resistance and transpulmonary pressure. The “older” cluster was characterized by high sputum eosinophil levels, high mucus plug score, widespread air trapping, segmental airway narrowing, excessive airway resistance and reduced transpulmonary pressure. The “early onset male” cluster was similar to the older age cluster but distinguished by lower sputum eosinophil, moderate air trapping, no segmental airway narrowing, leading to relatively lower airway resistance. The “airway thickened female” cluster showed airway wall thickening but had preserved lung function attributable to no significant luminal narrowing or air trapping resulting in relatively normal airway resistance and transpulmonary pressure. Asthma severity, gender, air trapping percentage (AirT%), normalized segmental airway wall thickness (WT*), and high attenuation area percentage (HAA%) served to characterize one of the five clusters. These also were the top five discriminative features, followed by age of onset, asthma control test (ACT) score, and normalized segmental airway hydraulic diameter, which predicted cluster memberships with an accuracy of 0.94. Conclusion qCT, CFD, and clinical data-integrated clustering differentiated lung structure-function impairment across novel clinical phenotypes in severe asthma. These findings, if replicated in larger cohorts, would provide novel insights into potential targetable therapeutic endpoints. This abstract is funded by: NIH Grants KL2-TR002346, TL1-TR002344, U10-HL109257, RO1 HL091762, U01HL146002, U10-HL109086, KL2-TR000450, KL2-TR002345, UL1TR0002366, UL1-TR000448, and TL1-TR000449.
 Asthma Machine Learning Cluster analysis

Details

Metrics

1 Record Views
Logo image