Early predictors of persistent childhood asthma and allergic rhinitis: A machine learning study using post-hoc explainability.

Abstract

Background: Asthma and allergic rhinitis (AR) follow heterogeneous childhood trajectories, and reliable early prediction of persistence or remission remains challenging. We investigated whether machine learning models applied to early-life data could predict long-term asthma and AR outcomes.

Methods: We analyzed data from 411 children in the Copenhagen Prospective Studies on Asthma in Childhood (COPSAC₂₀₀₀) cohort using 113 early-life clinical, genetic, environmental, immunological, and lung function features. Five machine learning models were trained to predict four outcomes: persistent asthma versus controls, persistent asthma versus asthma in remission, persistent AR versus controls, and teen-onset AR versus controls. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).

Results: Predictive performance was limited across all outcomes, and no model achieved an AUROC above 0.65. The highest performance was observed for persistent AR versus controls, where the tree-based models performed similarly, with AUROCs around 0.63. For persistent asthma versus controls, Random Forest and XGBoost both achieved an AUROC of 0.59. Performance was weak for distinguishing persistent asthma from asthma in remission and for predicting teen-onset AR, with the best models reaching AUROCs of 0.60 and 0.59, respectively. SHAP analyses of the best-performing models identified biologically plausible features, including aeroallergen sensitization and genetic predisposition for persistent asthma, and familial, immunological, and growth-related features for persistent AR, but these signals did not translate into strong discriminative performance.

Conclusion: In this deeply phenotyped high-risk birth cohort, early-life data analyzed using machine learning did not reliably predict long-term asthma or AR outcomes. These findings suggest that early-life data alone may be insufficient for individual prognostic stratification and that repeated measurements later in childhood may be needed to improve prediction.