Bayesian additive regression trees for predicting childhood asthma in the CHILD cohort study

Area of Science:

• Pediatric respiratory health
• Computational epidemiology
• Machine learning in medicine

Background:

• Asthma diagnosis in children lacks a universal gold standard, relying on age-dependent clinical assessments.
• Machine learning offers potential for improved asthma diagnosis and classification.
• Limited research exists on Bayesian machine learning for childhood asthma prediction.

Purpose of the Study:

• To develop and evaluate a prediction model for childhood asthma using Bayesian Additive Regression Trees (BART).
• To compare BART's performance against other machine learning algorithms for asthma diagnosis.
• To identify key predictors and interaction effects for childhood asthma using BART.

Main Methods:

• Utilized data from 2794 participants in the CHILD Cohort Study, focusing on variables available by age 3 to predict asthma at age 5.
• Trained BART and six other machine learning algorithms (adaptive boosting, logistic regression, decision tree, neural network, random forest, support vector machine).
• Assessed model performance using sensitivity, specificity, and area under the ROC curve, with confidence intervals from bootstrapping.

Main Results:

• BART, logistic regression, and random forest demonstrated the highest predictive accuracy (area under the ROC curve).
• Key predictors identified by BART included recurrent wheeze, respiratory infection, and food sensitization at age 3.
• Significant interaction effects involved combinations of respiratory infection, wheezing, parental asthma, and child sensitization.

Conclusions:

• BART shows strong potential for predicting childhood asthma compared to other machine learning methods.
• The model identified crucial early-life risk factors and their interactions.
• External validation of BART is recommended to confirm its reliability and generalizability in diverse populations.