Abstract Rationale Although chronic obstructive pulmonary disease (COPD) is a respiratory disease, characterized by the presence of airflow limitations, extra-pulmonary manifestations have been shown to occur and contribute to disease burden, however they are often not included in prediction models. The aim of this study was to investigate prediction models that incorporate pulmonary and extra-pulmonary CT imaging features for predicting lung function decline. Methods Subjects from the Genetic Epidemiology of COPD (COPDGene) study obtained chest CT imaging at baseline and spirometry at baseline and 5-year follow-up visit. A total of 63 regions were segmented from the CT images using TotalSegmentator, including regions from the pulmonary, cardiac, nervous, digestive, and musculoskeletal system. From each region a total of 39 CT densitometry-based imaging features were extracted, including mean, standard deviation, skewness, kurtosis, volume, mass, high attenuation areas, low attenuation areas etc. Any features with a variance of zero were excluded from the analysis. Features were grouped in pulmonary and extra-pulmonary features sets. Machine learning models were constructed for the pulmonary features, extra-pulmonary features, and their combination, for predicting binary FEV1 decline (≥60mL/year) with a CatBoost classifier. Models were fine-tuned using GridSearch in the training dataset (80%) and evaluated in the testing dataset (20%) using the area under the receiver operating characteristics curve (AUC) and Shapely Additive Explanations (SHAP). Adjusted logistic regression models were constructed to investigate associations with FEV1 decline (covariates: age, sex, race, BMI, smoking status, baseline FEV1 %predicted, CT %Emphysema, and Pi10). Results 5495 subjects were included in the analysis and 808 (15%) experienced FEV1 decline ≥60mL/year. There was no significant difference between the pulmonary (AUC=0.638, Confidence Interval (CI): 0.602-0.674) and extra-pulmonary model (AUC=0.654, CI: 0.619-0.689, p = 0.54) or the combined model (AUC=0.643, CI: 0.607-0.679, p = 0.85). Although not significant, the extra-pulmonary model obtained the highest performance (AUC=0.654). The SHAP analysis (Figure 1) illustrates that 8/10 most important features in the combined model were extra-pulmonary features. Remarkably, 6/10 features were from the skeletal system, including 3 CT rib mass features. After adjusting for covariates, the rib features were significantly associated with FEV1 decline (p-value0.001). Conclusion CT imaging features extracted from extra-pulmonary organs, notably the ribs, were identified as important predictors of FEV1 decline. These extra-pulmonary CT rib characteristics may be providing information about complex structural changes to the ribs, which may be affecting the respiratory muscles and breathing mechanics, which can be targeted with pulmonary rehabilitation to improve lung function. This abstract is funded by: This work was supported by NHLBI grants 1R01HL149877, U01 HL089897, and U01 HL089856 and by NIH contract 75N92023D00011.
Makimoto et al. (Fri,) studied this question.