Identification of Chronic Obstructive Pulmonary Disease Phenotype using PCA and Clustering Methodologies

Context
The approach to classification of chronic obstructive pulmonary disease patients is complicated by the heterogeneous and multidimensional nature of the disease. In the past years, application of classification methods in this context have been developed by limited and selected variables to define phenotypes with comparable results.
Objective
In the present study, we aim to identity distinct phenotypes of adults suffering from chronic obstructive pulmonary disease (COPD). Clustering was applied to understanding, management and better predict future risks and optimize treatment selection based on the new groupings of patients.
Method
In this application, patients were described by multiple and huge sets of variables that structured into groups. Two step approach was used for reduction of the huge variables and identification of the cluster. In the first step, the choice of variables to input into the clustering algorithm was one of the most important considerations. Most of previous studies employed manual variable extraction based on expert advice or dimension reduction techniques such as PCA. However, in this study we used multiple factor analysis (MFA), which is an extension of principal component analysis (PCA), for reducing the complexity of high-dimensional data. After this first step, a hierarchical and partition clustering method around medoids based on Jensen–Shannon distance was applied to the multiple factor analysis. This method is more robust to noise and outliers. All statistical analyses were performed using R software.
Results
The best cluster number was found to be equal to 5. Those clusters were identified as: mild COPD, Eosinophilic COPD, Atopic COPD, COPD + Emphysema and neutrophilic COPD + Emphysema + Systemic inflammation. Use of other criterion to compute distance didn’t present a significant improvement in the results obtained from clinical data.
Conclusions
Application of this new cluster approach allowed to identify five groups within stable COPD patients based on huge clinical variables while accounting for noise and outlier. Derived results can be used to predict outcomes of patients with COPD and to aid in development of personalized therapy.