AI for predicting exacerbations in KIDs with asthma (AIRE-KIDS).

Area of Science:

• Pediatric Asthma Management
• Health Informatics
• Machine Learning in Healthcare

Background:

• Recurrent acute care visits for pediatric asthma are a significant, preventable burden.
• Electronic medical records (EMR) offer potential for identifying high-risk children.
• Targeted interventions can reduce emergency department (ED) visits and hospitalizations.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting repeat asthma-related ED visits or hospital admissions in children.
• To assess the performance of boosted tree methods and large language models (LLMs) using retrospective and prospective data.
• To identify key predictors of acute care utilization in pediatric asthma patients.

Main Methods:

• Retrospective data (pre-COVID-19) from a tertiary children's hospital (CHEO) were used for model training.
• Models included boosted tree methods (LGBM, XGBoost) and LLMs (DistilGPT2, Llama variants).
• Environmental pollutant exposure and neighborhood marginalization data were integrated; models were validated on post-COVID-19 data.

Main Results:

• The LGBM model demonstrated the best performance with an AUC of 0.712 and an F1 score of 0.51, outperforming current best practices (F1 0.334).
• Key predictors identified include prior asthma ED visits, triage acuity, medical complexity, food allergy, prior non-asthma respiratory ED visits, and age.
• The AIRE-KIDS models showed accuracy in predicting future acute care needs.

Conclusions:

• Machine learning models, particularly LGBM, can accurately predict recurrent acute care visits in children with asthma.
• AIRE-KIDS models can aid emergency department decision-making for timely referral to preventative care.
• This approach has the potential to improve equitable access to preventative asthma care for high-risk children.