Challenges in identifying asthma subgroups using unsupervised statistical learning techniques.

Area of Science:

• Pediatric Allergy and Immunology
• Computational Biology
• Biostatistics

Background:

• Unsupervised statistical learning, including exploratory factor analysis (EFA) and hierarchical clustering (HC), is utilized for identifying asthma phenotypes.
• Inconsistent results in asthma phenotype identification stem from variations in variable selection and study population characteristics.

Purpose of the Study:

• To evaluate how statistical method choice and data preparation impact clustering results for asthma phenotypes.
• To correlate identified asthma clusters with disease severity.

Main Methods:

• Applied multiple variants of EFA and HC to a dataset of 383 children with asthma.
• Utilized diverse variable sets, encodings, and transformations, including lung function, inflammatory markers, and environmental exposures.
• Related clusters and variables to asthma severity using logistic regression and Bayesian network analysis.

Main Results:

• EFA identified five components explaining 35% of variance; HC results varied with data encoding but not linkage functions.
• Derived asthma clusters were less predictive of severity than original variables.
• Key prognostic factors for asthma severity included medication use, symptoms, lung function, paternal asthma, BMI, and age of onset.

Conclusions:

• Different unsupervised learning methods and data processing choices can produce inconsistent asthma subgroupings.
• These subgroupings may not consistently correlate with disease severity.
• Future research requires rigorous marker selection and cautious interpretation of unsupervised learning outcomes for robust asthma phenotyping.