Predicting phenotypes of asthma and eczema with machine learning

Area of Science:

• Computational biology
• Clinical informatics
• Precision medicine

Background:

• Asthma, wheeze, and eczema are increasingly recognized as heterogeneous diseases.
• Investigating machine learning (ML) to differentiate clinical subgroups within these conditions is crucial.
• A large, diverse dataset of attributes was used to explore disease manifestations.

Purpose of the Study:

• To assess the predictive capability of various ML methods in classifying asthma, wheeze, and eczema subgroups.
• To determine the extent to which heterogeneous information can be integrated to reveal specific clinical phenotypes.
• To identify novel predictors for conditions like eczema.

Main Methods:

• A cross-sectional study of 554 adults, including general population and asthma cohorts.
• Application of linear and non-linear ML models (e.g., logistic regression, random forests).
• Analysis of a comprehensive attribute set: demographic, clinical, laboratory, genetic (SNPs), and environmental factors.

Main Results:

• Non-linear ML models demonstrated superior sensitivity and specificity for asthma (AUC 84%) and wheeze (AUC 76%), with moderate performance for eczema (AUC 64%).
• Combined allergen sensitization and lung function data were more effective in characterizing asthma than individual factors.
• Genetic markers showed limited predictive power alone; bio-impedance emerged as a potential predictor for eczema.

Conclusions:

• Complex modeling is essential for understanding disease mechanisms and enabling personalized healthcare.
• Future advancements in disease classification and treatment will benefit from incorporating more diverse factors and longitudinal data.