B78-17 CT Digital Twin Lung-based Regional Functional Defects Explain Airflow Obstruction in COPD

Rationale

Chronic obstructive pulmonary disease (COPD) presents heterogeneous regional distribution of functional impairments, leading to airflow obstruction. However, gravitationally dependent non-uniform distribution has limited accurate quantitative evaluation of functional deficiencies at local lung regions. We aimed to develop quantitative computed tomography (qCT) digital twin lung-based novel functional imaging biomarkers and identify the importance of key functional abnormalities for predicting airflow obstruction in COPD.

Methods

Inspiratory and expiratory lung CTs, demographics, and pulmonary function test results were collected from 26 healthy individuals (58.0±5.8y;M:F=20:6) and 79 patients with COPD (71.2±7.4y;M:F=72:7) at Jeonbuk National University Hospital (Jeonju, Korea). Commercial (VIDA Vision 2.2, Coralville, IA) and in-house qCT software computed local lung structural-functional features including ventilation (relative regional air volume change, RRAVC), volumetric deformation (J), motion (dorsoventral [Δy] and apicobasal [Δz] displacements), and airway diameters (D). Local lung abnormality of each feature was computed subtracting LightGBM machine learning model-predicted healthy reference values from qCT-computed values. Then, percent lung volumes of the regions with ventilation and deformation defects (RRAVC-, J-, Δy-, and Δz-) and airway narrowing were used to predict airflow obstruction by post-bronchodilator FEV1/FVC, by multiple machine learning prediction approaches (decision tree, random forest, gradient boosting, XGBoost, LightGBM, and linear regression). Shapley (SHAP) feature importance analysis identified key features.

Results

Spearman correlations and age, sex, and BMI-adjusted linear regression showed strong/moderate association of showed strong association post-bronchodilator FEV1/FVC with proposed ventilation and deformation defect percentages (r=-0.66, -0.68, -0.62, and -0.56 for RRAVC-, J-, Δy-, and Δz- , respectively; all p < 0.001). The random forest model (RMSE=0.062; R2=0.71) showed the best performance to predict post-bronchodilator FEV1/FVC followed by XGBoost (RMSE=0.065; R2=0.69). Top two model-based Shapley (SHAP) feature importance analysis identified ventilation defect percentage (RRAVC-) as the most important feature to predict post-bronchodilator FEV1/FVC, followed by age, volumetric deformation defect percentage (J-), and motion defects (Δy-, and Δz-). Among others, RRAVC- also showed the strongest correlation with reduced diffusing capacity of the lungs for carbon monoxide (DLCO) and Global Initiative Obstructive Lung Disease (GOLD) stage (r=-0.63, 0.64; both p < 0.001).

Conclusion

Functional defects from the qCT and machine learning-based digital twin lung had strong association with and decent predictability for clinically measured airflow obstruction for COPD. The proposed imaging-based digital twin lung approach may allow CT imaging to be a useful tool for interpretation of regional lung functional defects and screening of airflow obstruction for individuals at risk for COPD.

This abstract is funded by: Korea National Research Foundation (NRF) grants RS-2025-00521675 and 2021R1C1C1009818; Seoul National University grant 800-20220619 through Mid-Career Bridging Program; National Institute of Environmental Research (NIER), funded by the Ministry of Environment (ME) of the Republic of Korea (NIER-2023-04-02-176)