Background: Emphysema extent in the upper lobes on computed tomography (CT) imaging is associated with lung function decline in COPD (Hoessein, F.A. et al. Eur Respir J 2012; 844-850). However, numerous CT disease-related features can be quantified regionally, but it is unknown what features contribute more to lung function decline in COPD.

Aim: To compare upper lobe emphysema extent only (measured by low attenuation areas below -950HU, LAA₉₅₀) and all CT features extracted from the upper lobes for predicting rapid FEV₁ decline in COPD with machine learning. We hypothesize that including all CT features will have an improved performance for predicting rapid FEV₁ decline compared to LAA₉₅₀ only.

Methods: COPD participants from the CanCOLD study underwent CT and spirometry at baseline; spirometry was repeated at a 1.5-3-year follow-up. A total of 110 measurements were used: 102 upper lobe CT features (95 texture-based radiomics and 7 conventional CT features), 5 demographics and 3 spirometry. Machine learning models were evaluated using all CT features or LAA₉₅₀ alone, with demographics and spirometry for predicting rapid FEV₁ decline?60mL/yr in COPD. Models were compared using the area under the receiver operating characteristic curve (AUC) and DeLong's test.

Results: A total of 486 COPD participants were evaluated. The model using all CT features (AUC=0.73) obtained a higher performance compared to LAA₉₅₀ only (AUC=0.59, p-value=0.049). In the model with all CT features, baseline FEV₁, low attenuation cluster and 3 texture-based radiomic features were selected.

Conclusion: CT features in the upper lobes reflecting emphysema clustering are more predictive than emphysema extent for rapid FEV₁ decline in COPD.