Background Chronic Obstructive Pulmonary Disease (COPD) is a heterogeneous condition characterized by persistent, poorly reversible airflow obstruction. While some patients experience accelerated decline in forced expiratory volume in one second (FEV1), others remain stable. We hypothesized that unsupervised analysis of chest computed tomography (CT) scans using machine learning-derived radiomic features may identify endophenotypes associated with distinct clinical and biological characteristics and FEV1 decline trajectories. Methods We analyzed 101 radiomic features from 1759 chest CT scans of COPD patients in the ECLIPSE study. Unsupervised consensus clustering identified six mutually exclusive radiomic clusters, and we derived six corresponding average patient scores (APSC1–6). Random-coefficient models assessed associations between each APSC and baseline clinical characteristics, FEV1, and its three-year change, adjusting for relevant covariates. Associations with baseline gene expression in sputum and blood were also evaluated. Results Radiomic scores were associated with multiple baseline clinical features. Higher APSC2 (−5.3 mL·year −1 ; 95% CI −9.5, −1.0; p=0.01) and APSC6 (−5.5 mL·year −1 ; 95% CI −9.6, −1.3; p=0.01) predicted greater FEV1 decline, whereas higher APSC3 was associated with slower decline (+5.2 mL·year −1 ; 95% CI 0.8, 9.6; p=0.02). APSC6 was associated with increased sputum expression of genes enriched in respiratory infection pathways, relevant COPD loci such as TRIM38 and IFIT3 , and higher blood neutrophil counts. Conclusions Unsupervised CT radiomic analysis identifies distinct COPD endophenotypes associated with variability in FEV1 decline and biological markers, supporting potential stratified treatment.
Cáceres et al. (Thu,) studied this question.