Computational phenotyping of obstructive airway diseases and allergy

Abstract

Allergy and obstructive airway diseases, e.g., asthma, are common and may inflict substantial hardship. They are also interlinked and heterogeneous in their clinical presentations (phenotypes). Artificial intelligence (AI) is useful for phenotyping, by facilitating pattern discovery in complex data. The aim of this thesis was, using AI, to further our understanding of phenotypes of obstructive airway diseases (focusing on asthma), symptoms thereof, and allergies. In addition, potential risk factors and outcomes were evaluated to provide further clinical utility.

Paper I was a systematic review on AI-derived longitudinal phenotypes (trajectories) of asthma and allergy in children. A total of 71 studies were identified with differing methodologies and populations. However, overall, few trajectories were consistently identified. The methodological approaches were generally flawed, with insufficient reporting. Further, most studies used binary measures of disease, without inclusion of assessment of severity or other disease characteristics, which may have underestimated the variety of identified phenotypes. Meta-analyses were limited by heterogeneity, but confirmed the association with actionable risk factors, e.g., prenatal/childhood smoking exposure.

Paper II was a trajectory analysis on phenotypes of asthma, allergic rhinitis, and eczema in a population-based cohort of children. The input data consisted of parental report of disease and register data on dispensed medication. Nine trajectories were identified, differing in disease and medication patterns, even between trajectories that appeared similar from a binary disease/no disease perspective. These findings illustrate how richer data enable identification of more informative phenotypes.

Paper III was a cluster analysis of late-onset asthma phenotypes, utilizing interview/clinical data from representative adult samples. Clustering was done separately on those with asthma debuting at ≥12 years, ≥20 years, and ≥40 years. The identified sets of clusters were similar across onset age groups, but varied substantially among each other in symptoms, lung function, asthma control, and comorbidities. Importantly, they appeared relatively easy to distinguish in the clinical setting.

Paper IV was a cluster analysis based on self-reported respiratory symptoms, utilizing data from population-based adult cohorts. Five clusters were identified, ranging widely in symptom locality (e.g., lower respiratory symptoms), type, and frequency/persistence. Further, it was found that all clusters, except the allergic nasal symptoms cluster, even the low-symptomatic cluster, were associated with increased (cause-specific) mortality. As the clusters only partially overlapped with diagnosed respiratory disease, our findings underscore the clinical relevance of respiratory symptoms in their own right.

In conclusion, the present work demonstrates that AI, particularly on multimodal data, can guide us towards a more comprehensive subtyping of obstructive respiratory disease and allergy, ultimately paving way for personalized management and preventive measures.