Late Breaking Abstract - Prediction of allergic disease trajectories up to adolescence

<bold>Background:</bold> The development of allergic diseases often starts in early childhood. Several potential disease trajectories and multiple early-life environmental risk and protective factors have been described. <bold>Aims and Objectives:</bold> We aimed to improve the prediction of allergic disease trajectories up to adolescence through machine learning models. <bold>Methods:</bold> Preceding research using data from the two German birth cohorts GINIplus and LISA identified seven allergic disease trajectories until the age of 15, for instance “early resolving dermatitis” or “multimorbid”. Data of 4646 adolescents were available. We applied least absolute shrinkage selection operator feature selection to derive a predictor subset consisting of early-life symptoms, family history of allergies, early lifestyle factors and environmental exposures. We implemented machine learning models to perform multiclass classification with the non-allergic cluster as reference group. <bold>Results:</bold> The models revealed a moderate classification success (multiclass area under the curve (AUC) = 0.68 and balanced accuracy = 0.58). Boosting reveals the highest macro-averaged sensitivity and specificity of 0.28 and 0.89, respectively. For instance, the AUC in the “multimorbid” cluster reaches 0.79 with a sensitivity of 0.18 and specificity of 0.98. The most important predictors encompass early-life rash, distance to the nearest busy road, respiratory symptoms and parental history of allergies. <bold>Conclusions:</bold> Early detection of subgroups with a high risk of disease progression is crucial for prevention. Boosting produces a prediction success comparable to established asthma prediction scores (e.g. Asthma Predictive Index) while accounting for multiple allergic disease trajectories.