Childhood asthma remains one of the most burdensome chronic conditions in pediatrics, yet clinicians still lack reliable, validated tools to identify which children are most likely to land in the emergency department or hospital within the next year. A machine-learning framework that can flag high-risk patients before a crisis strikes — drawing on the full richness of the clinical record — could meaningfully shift care from reactive to preventive.
Investigators mined over a decade of electronic health records from UF Health (2011–2023), applying two established asthma computable phenotypes — CAPriCORN and COMPAC — to define patient cohorts and outcome events. Exacerbations were operationalized through a validated composite of diagnosis codes from emergency, inpatient, and outpatient settings paired with systemic corticosteroid prescriptions. Multiple machine-learning architectures were benchmarked using stratified five-fold cross-validation and Bayesian hyperparameter tuning, with XGBoost emerging as the top performer across 6-, 12-, and 24-month prediction horizons. Crucially, SHAP (SHapley Additive exPlanations) analysis revealed that free-text symptom terms extracted from clinical notes and rescue-inhaler usage patterns were the dominant predictive signals — features that structured billing codes alone would miss.
This work sits at the intersection of natural language processing and clinical decision support, an area gaining rapid traction but still short on pediatric-specific models. The reliance on free-text note mining is both a strength and a limitation: unstructured text captures clinical nuance, but note quality varies widely across institutions, raising real questions about generalizability beyond a single academic health system. The study also remains single-center and retrospective, meaning prospective validation is essential before any clinical deployment. That said, the interpretable SHAP framework addresses a longstanding criticism of black-box models in medicine. This is solid incremental progress — not yet paradigm-shifting — but the text-integration approach offers a credible path toward practical, explainable pediatric risk stratification tools.