Structural determinants of pulmonary diffusing capacity identified by network analysis and machine learning on quantitative CT.
Authors
Affiliations (5)
Affiliations (5)
- Division of Pulmonary and Critical Care Medicine, Department of Internal Medicine, Ilsan Paik Hospital, Inje University College of Medicine, Juhwa-ro 170, Ilsanseo-gu, Goyang, 10380, Republic of Korea.
- Department of Neurology, Ilsan Paik Hospital, Inje University College of Medicine, Goyang, Korea.
- Department of Radiology, Ilsan Paik Hospital, Inje University College of Medicine, Goyang, Korea.
- Division of Pulmonary and Allergy, Department of Internal Medicine, Konkuk University Medical Center, Konkuk University School of Medicine, Seoul, Republic of Korea.
- Division of Pulmonary and Critical Care Medicine, Department of Internal Medicine, Ilsan Paik Hospital, Inje University College of Medicine, Juhwa-ro 170, Ilsanseo-gu, Goyang, 10380, Republic of Korea. [email protected].
Abstract
Diffusing capacity for carbon monoxide (DLCO) reflects pulmonary gas exchange efficiency, but its measurement and interpretation remain challenging due to physiological and technical variability. Quantitative computed tomography (CT) provides structural insights that may help address these limitations. This study investigated associations among demographic factors, spirometric values, DLCO, and CT-derived metrics in patients with various lung conditions. Additionally, we developed predictive models for DLCO. We analyzed mean lung density (MLD), percentile index 15 (PI15), percentages of low and high attenuation areas (LAA% and HAA%), emphysema size heterogeneity (D-slope), airway wall thickness (AWTPi10), and pulmonary vascular indices. Network analysis and random forest regression were employed to identify determinants and predictors of DLCO. DLCO correlated positively with weight and whole lung volume, and negatively with age, MLD variation, HAA%, and D-slope. DLCO/alveolar volume was positively associated with weight, body mass index, and lung density metrics (PI15), and negatively with LAA% and D-slope. CT-derived metrics exhibited distinct correlation patterns compared to spirometric measurements. The random forest model predicted DLCO with correlation coefficient of 0.82, an RMSE of 3.04 mL/min/mmHg, and an R<sup>2</sup> of 0.58, indicating considerable predictive performance. Integrating quantitative CT metrics improves the understanding of DLCO. Imaging-based models may enhance diagnostic precision and support personalized management strategies for pulmonary diseases.