Abstract
INTRODUCTION: Computational sciences have significantly contributed to characterizing airway disease phenotypes, complementing medical expertise. However, comparing studies that derive phenotypes is challenging due to varying decisions made during phenotyping. We conducted a systematic review to describe studies that utilized unsupervised computational approaches for phenotyping obstructive airway diseases in children and adults. METHODS: We searched for relevant papers published between 2010 and 2020 in PubMed, EMBASE, Scopus, Web of Science, and Google Scholar. Additional sources included conference proceedings, reference lists, and expert recommendations. Two reviewers independently screened studies for eligibility, extracted data, and assessed study quality. Disagreements were resolved by a third reviewer. An in-house quality appraisal tool was used. Evidence was synthesized, focusing on populations, variables, and computational approaches used for deriving phenotypes. RESULTS: Of 120 studies included in the review, 60 focused on asthma, 19 on severe asthma, 28 on COPD, 4 on asthma-COPD overlap (ACO), and 9 on rhinitis. Among asthma studies, 31 focused on adults and 9 on children, with phenotypes related to atopy, age at onset, and disease severity. Severe asthma phenotypes were characterized by symptomatology, atopy, and age at onset. COPD phenotypes involved lung function, emphysematous changes, smoking, comorbidities, and daily life impairment. ACO and rhinitis phenotypes were mostly defined by symptoms, lung function, and sensitization, respectively. Most studies used hierarchical clustering, with some employing latent class modeling, mixture models, and factor analysis. The comprehensiveness of variable reporting was the best quality indicator, while reproducibility measures were often lacking. CONCLUSION: Variations in phenotyping methods, study settings, participant profiles, and variables contribute to significant differences in characterizing asthma, severe asthma, COPD, ACO, and rhinitis phenotypes across studies. Lack of reproducibility measures limits the evaluation of computational phenotyping in airway diseases, underscoring the need for consistent approaches to defining outcomes and selecting variables to ensure reliable phenotyping.