The application of bayesian and frequentist regularization and variable selection methods for the prediction of asthma in later childhood

Abstract

Asthma is a global health problem and among the most common chronic conditions in childhood. Several models were proposed to predict asthma in children, but their reproducibility in external populations was limited and none was developed to predict asthma in adolescence. I conducted a systematic review of asthma predictive models validated in external populations; validation studies showed poorer predictive performances than development studies. I developed predictive models for asthma between 15 and 20 years, using data from the Study Team for Early Life Asthma Research (STELAR) consortium of five UK asthma cohorts. For one of these cohorts, the Ashford study, I developed an questionnaire to collect follow-up information when study subjects were age 20 years. I harmonised 41 variables across the STELAR cohorts, 39 of which were used as candidate predictors to develop predictive models, while the others were used to define asthma at 15–20 years. Asthma at that age was defined as positive responses to ‘current wheezing’ and ‘asthma medications in the last year’.Two of the five STELAR cohorts (development data) were combined to develop predictive models using stepwise regression and frequentist, Bayesian and empirical Bayes regularization models. The remaining cohorts (validation data) were used to assess predictive performance using discrimination and accuracy measures. Analyses were performed in two populations - all children and a subgroup with reported wheezing between two and five years (high-risk population). Sex, eczema, sensitization to house dust mite and doctor’s diagnosis of asthma in early childhood (4-7 years) were identified as asthma predictors at 15-20 years in both populations. Additional predictors in the general population included early wheezing symptoms and parental allergies, while in the high-risk population maternal allergies and pet in the house at one year were important for asthma prediction in adolescence. Sensitivity was higher in the general population, whereas positive predictive value was higher in the high-risk population. Although accuracy was good in both populations, the predictive ability of the models developed was limited.