Exploratory Identification of Multidimensional COPD Clusters Using Unsupervised Analysis.

Background and Objectives: COPD is a heterogeneous disease in which conventional clinical classifications may not fully capture the complexity of patient profiles. This exploratory study aimed to examine whether multidimensional COPD phenotypes could be identified using unsupervised cluster analysis integrating clinical, functional, radiological and laboratory features. Materials and Methods: We enrolled 161 patients with confirmed COPD evaluated between January 2020 and January 2024. Demographic, clinical, functional, radiological and laboratory findings were collected. Mixed-type data were analyzed using Gower distance and Partitioning Around Medoids (PAM) clustering. The optimal solution was selected by average silhouette width; stability was assessed by 1000 bootstrap resamples and sensitivity analyses. Results: The two-cluster solution had the highest silhouette width (0.184), although separation was modest. Cluster 1 (n = 78) was characterized by greater symptom and exacerbation burden, worse lung function, greater static hyperinflation, shorter 6-min walking distance and more frequent emphysema than cluster 2 (n = 83). Bootstrap resampling indicated internal stability, although concordance with the primary partition varied across sensitivity analyses. After correction for multiple post hoc comparisons, only LAMA/LABA/ICS use differed between clusters, whereas demographic characteristics, comorbidity burden and blood eosinophil levels were comparable. Conclusions: These exploratory findings suggest multidimensional assessment may complement conventional classifications, but external and longitudinal validation is needed before clinical implementation.