Unsupervised phenotyping of Severe Asthma Research Program participants using expanded lung data.

Area of Science:

• Pulmonary Medicine
• Computational Biology
• Immunology

Background:

• Previous asthma phenotyping relied on limited clinical, physiologic, or inflammatory markers.
• No prior studies have utilized machine learning with a comprehensive variable set for asthma subtyping.

Purpose of the Study:

• To identify asthma subphenotypes using unsupervised clustering on a wide range of data.
• To characterize identified subphenotypes using supervised learning approaches.

Main Methods:

• Applied unsupervised clustering to 112 variables (clinical, physiologic, inflammatory) from 378 subjects.
• Utilized variable selection and supervised learning to identify predictive and nonredundant features.

Main Results:

• Identified 10 variable clusters and 6 distinct subject clusters, some overlapping with prior classifications.
• Characterized clusters including early-onset allergic asthma, late-onset asthma with nasal polyps, and persistent inflammation phenotypes.
• 51 nonredundant variables distinguished clusters, achieving 88% classification accuracy (112 variables achieved 93%).

Conclusions:

• Unsupervised machine learning provides novel insights into asthma heterogeneity.
• Identified phenotypes complement existing classifications and reveal new disease endotypes.
• This approach enhances understanding of complex asthma presentations.