LATE-BREAKING ABSTRACT: Predicting asthma at age 8; the application of machine learning

Abstract

Background: Asthma is among the most common chronic conditions in childhood. We aimed to develop and validate a simple and robust model to predict asthma at 8 years of age. Methods: The data come from 3 UK cohorts in the STELAR consortium. We studied 1,145 children from Ashford and Aberdeen and externally validated the predictive model using data on 348 children from Manchester. Information on characteristics of the children, family related factors and asthma-like symptoms were collected at recruitment and at 1 and 2 or 3 years of age. We defined asthma at age 8 by the presence of at least two of the following: (1) current wheeze; (2) asthma treatment; (3) a doctor9s diagnosis of asthma. We developed a predictive model using the Least Absolute Shrinkage and Selection Operator method (LASSO) and assessed predictive performance by discrimination and calibration measures. Results: Current asthma was present in 117 (22.5%), 119 (21%) and 80 (21.5%) children living in Ashford, Aberdeen and Manchester, respectively.The final model included 20 predictors such as gender, maternal history of hay-fever or eczema, parental history of asthma, older sibling(s), domestic crowding; a doctor9s diagnosis of eczema, current wheeze and antibiotic use, each before the age of 3. Predictive accuracy in the external validation was 0.83, the area under the receiver operating characteristic curve was 0.65 and positive and negative predictive values were 0.81 and 0.83. Conclusion: After validation, our LASSO model demonstrated good discrimination ability for asthma; its predictive performance compares favorably with that from previous tools, such as the API and PIAMA, due to increased specificity and higher positive predictive value.