Combining supervised and unsupervised models to characterize asthma phenotypes in children

Background: Asthma is not a single disease and a number of phenotypes are now recognized. Characterizing these phenotypes aids in better management of asthma for individuals in the pursuit of personalized medicine. Aim: To identify and characterize the asthma phenotypes in an asthma-rich population of children. Methods: The study population included 752 twins, aged 9-14, selected from a larger twin study. Parents answered questions on symptoms, risk factors and medication use. Clinical tests assessed lung function and immunological biomarkers. A latent class analysis (LCA) using current and historical symptom data was used to generate asthma phenotypes (unsupervised). Multinomial regression was then applied to characterize these phenotypes further based on medication use, risk factors, biomarkers and lung function (supervised). Results: The LCA identified five phenotypes; healthy (61%), early transient wheeze (15%), persistent wheeze (5%), mild asthma (9%) and severe asthma (10%). All wheeze and asthma phenotypes were found to have reduced lung function, however early transient wheeze did not show reversibility with bronchodilators. Persistent wheeze was identified as an at-risk group as they had wheeze and sensitization but no asthma diagnosis and only used controller medication intermittently. Those with severe asthma were more likely to have used preventer medication and β2 agonists in the last year than those with mild asthma. They were also more likely to have a family history of asthma and an allergic profile. Conclusion: Characterizing asthma using supervised and unsupervised methods we have been able to identify several phenotypes that may improve the clinical management of asthma.