Use of Principal Component and Cluster Analysis to Describe Phenotypes in COPD

Abstract

COPD is a multidimensional and heterogeneous disease. In the past years, classification methods in COPD context have been developed based on clinical observation with a limited number of variables to define phenotypes. In such studies, the selection of variables included in the analysis may influence the results and have missed some phenotypes. The main purpose of this study is to identify clinical phenotypes among adults suffering from COPD. Clustering was applied to understand, manage and better predict future risks and optimize treatment selection based on the new groupings of patients. Furthermore, missing data and dimension-reduction, which are present in any large observational dataset, were handled. In this application, 178 patients were described by 85 multiple and huge sets of variables that structured into seven groups. At the first step, the missing values were imputed by multiple factor analysis (MFA). After imputation, MFA was applied for reducing the complexity of high-dimensional data. After this step, hierarchical clustering was performed using Ward’s criterion. In this step, the optimal number of clusters was selected based on several methods. In the final step, K-means as performed to improve the clustering. Three different phenotypes were defined in COPD. Phenotype 1 included severe males with exacerbation-prone, bacterial colonization/neutrophilic and systemic inflammation. phenotype 2 comprised women moderate COPD with emphysema and phenotype 3 included males with moderate COPD and mild atopic traits. These clinically meaningful clusters of patients with common characteristics can be used to predict outcomes of patients with COPD, to aid in the development of personalized therapy.