Structural and functional assessments of longitudinal COPD subpopulations via unsupervised machine learning

Chronic obstructive pulmonary disease (COPD) is highly heterogeneous in terms of syndromes and has tremendous variability in the rate of lung function decline among COPD patients. Tobacco smoking is the main cause of the high prevalence of COPD. With increasing availability of CT images, imaging-based measures are becoming an objective approach to assessing risk outcomes across multiple populations. In NIH-funded multicenter studies, such as SubPopulations and InteRmediate Outcome Measures in COPD Study (SPIROMICS), CT scans for large numbers of patients at multiple visits enabled interrogation of a broad sampling of the COPD population. We selected 899 subjects from SPIROMICS including 472 former smokers, 372 current smokers and 55 never smokers with a baseline visit and a one-year follow-up visit. A total of 150 quantitative computed tomography (qCT) imaging-based variables, comprising 75 variables at baseline and 75 corresponding progression rates over one year, were derived from the respective inspiration and expiration scans of the two visits. Based on the 150 variables, we perform longitudinal clustering and identified four clusters in former smokers and current smokers, respectively. The clusters were then associated with subject demography, clinical variables, and biomarkers. The longitudinal clusters were associated with unique progression patterns and clinical characteristics and smoking status resulted in different progression patterns. Compared with cross-sectional clusters which only utilized imaging-based variables at baseline, combination of baseline variables and their progression rates enables identification of longitudinal clusters, resulting in a refinement of cross-sectional clusters.