Rationale: Unsupervised statistical learning techniques, such as exploratory factor analysis (EFA) and hierarchical clustering (HC), have been used to identify asthma phenotypes, with partly consistent results. Some of the inconsistency is caused by the variable selection and demographic and clinical differences among study populations.
Objectives: To investigate the effects of the choice of statistical method and different preparations of data on the clustering results; and to relate these to disease severity.
Methods: Several variants of EFA and HC were applied and compared using various sets of variables and different encodings and transformations within a dataset of 383 children with asthma. Variables included lung function, inflammatory and allergy markers, family history, environmental exposures, and medications. Clusters and original variables were related to asthma severity (logistic regression and Bayesian network analysis).
Measurements and main results: EFA identified five components (eigenvalues ≥ 1) explaining 35% of the overall variance. Variations of the HC (as linkage-distance functions) did not affect the cluster inference; however, using different variable encodings and transformations did. The derived clusters predicted asthma severity less than the original variables. Prognostic factors of severity were medication usage, current symptoms, lung function, paternal asthma, body mass index, and age of asthma onset. Bayesian networks indicated conditional dependence among variables.
Conclusions: The use of different unsupervised statistical learning methods and different variable sets and encodings can lead to multiple and inconsistent subgroupings of asthma, not necessarily correlated with severity. The search for asthma phenotypes needs more careful selection of markers, consistent across different study populations, and more cautious interpretation of results from unsupervised learning.