Prediction of allergic disease trajectories from birth up to adolescence

Area of Science:

• Pediatric Allergy and Immunology
• Computational Epidemiology
• Machine Learning in Healthcare

Background:

• Allergic diseases frequently co-occur in early childhood, yet prediction models often focus on single conditions.
• Existing approaches lack models for predicting allergic multimorbidity and disease trajectories over time.
• Predicting the course of allergic diseases from birth through adolescence requires novel methodologies.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting allergic disease trajectories from birth to adolescence.
• To identify key early-life factors associated with distinct allergic disease pathways.
• To assess the utility of polygenic risk scores (PRS) in enhancing predictive models for allergic multimorbidity.

Main Methods:

• Utilized data from 4646 adolescents in the German GINIplus and LISA birth cohorts.
• Employed an XGBoost machine learning model for multiclass classification of seven identified allergic disease trajectories.
• Included parental, perinatal, symptom-based, lifestyle, and environmental factors as predictors; PRS were analyzed in a subsample.

Main Results:

• Achieved moderate classification success with a multiclass AUC of 0.69.
• Identified early-life skin rash, respiratory symptoms, and air pollution as significant predictors.
• Polygenic risk scores showed importance but did not improve prediction performance on external data.

Conclusions:

• The developed prediction model shows performance comparable to established scores using only early-life factors.
• The study highlights the potential of machine learning for understanding allergic disease trajectories.
• While not yet suitable for individual clinical prediction, the findings inform future research directions.