AI Identifies Four Distinct Asthma Phenotypes in Adults

FOUR distinct asthma phenotypes have been identified using machine learning, highlighting how clinical characteristics, lifestyle behaviours, and nutritional factors may contribute to differences in asthma presentation among adults.

Researchers analysing nationwide health data from South Korea identified asthma subgroups characterised by differences in body mass index (BMI), smoking status, lung function, and allergic characteristics. The findings suggest that integrating multiple health-related factors could support more personalised approaches to asthma assessment and management.

Using Machine Learning to Understand Asthma

Asthma is a heterogeneous respiratory condition influenced by genetic, environmental, and lifestyle factors. Patients can experience substantially different symptoms, patterns of airway inflammation, and responses to treatment, highlighting the limitations of treating asthma as a single disease entity.

To investigate this variability, researchers analysed data from the Korea National Health and Nutrition Examination Survey (KNHANES VIII, 2019).

The study included data from 198,964 participants, of whom 3,529 had physician-diagnosed asthma.

Researchers applied cluster analysis alongside two machine learning algorithms, Gradient Boosting and Random Forest, to investigate asthma phenotypes using clinical, lifestyle, and nutritional characteristics.

Variables included BMI, forced expiratory volume in 1 second (FEV₁%), smoking status, stress levels, and dietary nutrient intake.

The researchers also calculated odds ratios to examine associations between demographic, environmental, and nutritional factors and asthma within the identified clusters.

Four Distinct Asthma Phenotypes Identified

The analysis identified four asthma phenotypes with different combinations of respiratory and lifestyle characteristics.

The first phenotype was characterised by higher BMI, smoking, and a greater prevalence of atopy, suggesting a subgroup in which excess body weight, tobacco exposure, and allergic sensitisation frequently coexist.

The second phenotype demonstrated a higher prevalence of smoking but relatively low atopy, indicating a different clinical profile in which tobacco exposure was more prominent than allergic sensitisation.

The third phenotype was characterised by higher FEV₁% and a greater prevalence of allergies, suggesting comparatively preserved lung function alongside allergic characteristics.

The fourth phenotype demonstrated lower FEV₁%, together with lower prevalences of smoking, atopy, and allergies.

These differences highlight the considerable variation in asthma characteristics and suggest that reduced lung function does not necessarily occur alongside smoking or allergic sensitisation.

AI Models Achieve High Prediction Accuracy

The machine learning models achieved accuracy exceeding 95% in predicting asthma status.

These findings demonstrate the ability of computational approaches to identify patterns across multiple clinical and lifestyle variables.

However, overall accuracy alone does not establish clinical usefulness, particularly when asthma represents a relatively small proportion of the study population. Additional measures, including sensitivity and specificity, would be important for determining how reliably the models identify individuals with asthma.

Beyond clinical characteristics, researchers identified associations between environmental and nutritional factors and asthma risk that differed across the four phenotypes.

Higher stress levels and urban residence were associated with increased odds of asthma in specific clusters.

Associations involving omega-3 fatty acids, thiamine, vitamin D, and dietary fibre also varied between subgroups, suggesting that nutritional characteristics may have different relationships with asthma across patient populations.

Importantly, these findings were observational and do not establish that dietary factors, stress, or urban residence directly cause asthma.

Implications for Precision Asthma Management

The findings suggest that asthma phenotyping could benefit from considering a broader range of characteristics beyond conventional clinical measurements.

Integrating lung function, smoking history, BMI, allergic status, environmental exposures, and nutritional information may help researchers better understand differences between patient groups.

Such approaches could eventually contribute to more targeted prevention strategies and personalised asthma management.

However, the study does not demonstrate that using these phenotypes to guide treatment improves patient outcomes. Further validation in independent populations would also be needed to establish whether the identified patterns apply beyond Korean adults.

The researchers concluded that machine learning provides a promising framework for investigating asthma heterogeneity and identifying clinically relevant subgroups.

Further research is needed to determine whether integrating these multifactorial phenotypes into clinical practice can improve risk assessment, treatment selection, and long-term respiratory outcomes.

Reference

Choi, Y.J., Jeong, S. Machine learning-based identification of multifactorial asthma phenotypes in Korean adults: insights from the Korea National Health and Nutrition Examination Survey. BMC Pulm Med (2026). https://doi.org/10.1186/s12890-026-04868-x