Abstract Rationale Chronic obstructive pulmonary disease (COPD) is a heterogeneous disease that affects the pulmonary and extra-pulmonary systems, presenting with varied clinical characteristics, structural changes, and longitudinal outcomes. To better characterize this variability, the objective was to perform cluster analysis to identify distinct COPD endotypes using pulmonary and extra-pulmonary CT radiomics. Methods Subjects that obtained chest CT imaging and spirometry at baseline from the Genetic Epidemiology of COPD (COPDGene) study were included. A total of 39 CT densitometry-based imaging radiomics were extracted from 26 regions segmented using TotalSegmentator, including pulmonary, cardiac, nervous, digestive, and musculoskeletal regions. Any features with spearman correlations ≥0.80 or variance ≤0.05 were excluded. A principal component analysis (PCA) was performed on the remaining features as an unsupervised feature selection to identify top contributing features to components explaining 1% variance. The Silhouette index and Davies-Bouldin index were used to identify the optimal k clusters. Consensus clustering was performed by running K-means 1000 times and obtaining the stable cluster assignments. An ANOVA with Tukey’s post hoc or Chi-squared analysis was performed to compare clinical and CT imaging features between the identified clusters. Associations between cluster assignment and 5-year change in %Emphysema, 6-minute walk distance, SGRQ total, and FEV1 were evaluated using ordinary least squares (OLS) adjusted for demographic, clinical, and baseline measurements. Results A total of 9574 subjects were included in the cluster analysis and 4641 subjects with complete longitudinal data in the OLS sub-analysis. PCA identified 20 CT features, including features from the cardio-vascular (8), lungs (3), esophagus (2), stomach (2), pancreas (1), trachea (2), and adrenal glands (2), that were used to generate four unique clusters. The cluster characteristics include baseline and longitudinal clinical and CT imaging features (Table 1). Despite having mild emphysema and symptoms at baseline, Cluster 1 showed the largest increase in %Emphysema (β = 0.409, p = 0.002). Cluster 2, characterized by the most baseline %Emphysema, experienced the largest decline in exercise capacity (β=-31.6, p = 0.01). Although Cluster 3 had fewer symptoms and less baseline %Emphysema, these individuals experienced worsening symptoms (β = 1.17, p = 0.02) and the largest decline in FEV1 (β = 5.30, p = 0.01). Cluster 4, composed predominantly of symptomatic females with mild emphysema, remained stable. Conclusion CT-based cluster analysis incorporating pulmonary and extra-pulmonary radiomics identified four distinct clusters with differing clinical and imaging characteristics, while having limited variation between COPD severity. These cluster assignments have the potential to better risk stratify individuals that may benefit for targeted treatment. 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.