Cohort vs case–control design for transformer-based prediction of asthma exacerbations in mild asthma

Acute asthma exacerbation (AAE) is among the most serious outcomes of asthma, and accurate prediction of its risk remains a key challenge. Existing electronic health record-based prediction models have typically adopted either cohort or case-control sampling designs, yet their comparative performance has not been systematically evaluated. To address this gap, we developed transformer-based deep learning models to predict AAE among adults with mild asthma, directly comparing cohort and case-control designs using identical predictors and architecture across two large integrated healthcare systems, Kaiser Permanente Southern California (KPSC) and Kaiser Permanente Northwest (KPNW). Models were trained on retrospective data from KPSC and externally validated in KPNW. Mean area under the receiver operating characteristic curve (AUC) was 0.85 for case-control models and 0.70 (KPSC)/0.71 (KPNW) for cohort models. Both designs generalized well across systems, indicating robust feature learning and population transferability. Although calibration appeared well aligned within each analytical framework, absolute predicted probabilities diverged between designs, reflecting how event-enriched sampling inflates apparent risk and affects interpretability. These findings demonstrate that study design strongly influences model behavior and should be aligned with the intended use when developing predictive models for clinical deployment.