Machine learning approach to identify non-linear contributors to asthma development

<bold>Background:</bold> Early prediction of childhood asthma has important implications for successful prevention and treatment. <bold>Methods:</bold> We included 411 children from the COPSAC cohort. Children with asthma (n=146) were subcategorized in 1) transient (remission before age 17, n=54) and 2) persistent asthma (no remission, n=92). We tested 3 different machine learning (ML) methods (LightGBM, Random forest, Neural networks). The children were classified into outcomes based on 70 features, mainly heritability, early life exposures and biomarkers. <bold>Results:</bold> LightGBM was superior for classifying persistent asthma vs controls/transient asthma (AUC=0.67, fig 1A). It identified 12 variables impacting the best predicted outcome (persistent vs control): total IgE, dust exposure, small particles in the child's bedroom, sIgE to aeroallergens, blood IL-1, genetic risk scores, neonatal FEF 0.5, gestational age, airway infections 0-3 years and paternal asthma (fig 1B). Post-hoc analyses showed that persistent asthma was mainly driven by sensitization and genetic risk score. We identified distinct subtypes of persistent asthma, revealing that paternal asthma and elevated sIgE to aeroallergens were associated with two subtypes, while high levels of small particles in the household provided protection for another. <bold>Conclusion:</bold> Our ML model revealed unknown features relevant for persistent asthma and hints to potential subtypes of asthma.