loading page

Machine learning-derived asthma phenotypes in a representative Swedish adult population
  • +12
  • Muwada Bashir Awad Bashir,
  • Daniil Lisik,
  • Saliha Selin Özuygur Ermis,
  • Rani Basna,
  • Reshed Abohalaka,
  • Selin Ercan,
  • Helena Backman,
  • Teet Pullerits,
  • Roxana Mincheva,
  • Göran Wennergren,
  • Madeleine Rådinger,
  • Jan Lötvall,
  • Linda Ekerljung,
  • Hannu Kankaanranta,
  • Bright Nwaru
Muwada Bashir Awad Bashir
University of Gothenburg

Corresponding Author:muwada.bashir@gu.se

Author Profile
Daniil Lisik
University of Gothenburg
Author Profile
Saliha Selin Özuygur Ermis
University of Gothenburg
Author Profile
Rani Basna
Lunds universitet Institutionen for kliniska vetenskaper Malmo
Author Profile
Reshed Abohalaka
University of Gothenburg
Author Profile
Selin Ercan
University of Gothenburg
Author Profile
Helena Backman
Umea universitet Institutionen for folkhalsa och klinisk medicin
Author Profile
Teet Pullerits
Goteborgs universitet Avdelningen for invartesmedicin och klinisk nutrition
Author Profile
Roxana Mincheva
University of Gothenburg
Author Profile
Göran Wennergren
Goteborgs universitet Pediatrik
Author Profile
Madeleine Rådinger
University of Gothenburg
Author Profile
Jan Lötvall
University of Gothenburg
Author Profile
Linda Ekerljung
University of Gothenburg
Author Profile
Hannu Kankaanranta
University of Gothenburg
Author Profile
Bright Nwaru
University of Gothenburg
Author Profile

Abstract

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.