Machine learning-derived asthma phenotypes in a representative Swedish adult population

Background Asthma is a heterogenous airway disease characterized by multiple phenotypes. Unbiased identification of these phenotypes is paramount for optimizing asthma management. Objectives To identify and characterize asthma phenotypes based on a broad set of attributes using a novel machine learning approach in a representative sample of Swedish adults. Methods Deep learning clustering was used to derive asthma phenotypes in a sample of 1,895 subjects aged 16-75, drawn from the ongoing West Sweden Asthma Study. The algorithm integrated 47 variables encompassing demographics, risk factors, asthma triggers, pulmonary function, disease severity, allergy, and comorbidity profiles. The optimal clustering solution was selected by combining statistical metrics and clinical interpretation. Results A four-cluster solution was determined to reliably represent the data, resulting in distinct phenotypes described as: (1) troublesome, late-onset, non-atopic asthma with smoking ( n =458, 24.2%); 2) female-dominated early adult-onset asthma ( n =545, 28.7%); 3) adult-onset asthma with high inflammation ( n =358, 18.9%); and 4) early-onset, mild, atopic asthma ( n =534, 28.2%). The phenotypes also differed with respect to demographics, risk factors, asthma triggers, pulmonary function, symptom profiles, and markers of inflammation. Current asthma was more common in phenotypes with later age of asthma onset than phenotypes with early onset. Conclusion Four clinically meaningful asthma phenotypes, distinguishable by age of onset, severity, risk factors, and prognosis, were found in Swedish adults. This provides a setting for future research to profile the immunological basis of the phenotypes, and further our understanding of their pathophysiology, therapeutic possibilities, future clinical outcomes, and societal burden.