Abstract Rationale Increased short-acting β2-agonist (SABA) use is a well-recognized risk factor for poor asthma outcomes. Patterns of SABA use and reduction in peak flow typically occur prior to the onset of exacerbations. Digital inhalers capture SABA use and inhalation parameters that enable predictive modeling to identify impending exacerbations. Methods We analyzed a 52-week single center observational cohort of adults with severe asthma using the ProAir® Digihaler (N = 86). The outcome was a SABA-defined clinical deterioration event (CDE) from previously specified Delphi rules. For each index day, we predicted whether a CDE would occur between day +5 and day +9 (5-day risk window with 5-day lead time). Predictors were recent SABA use and inhalation parameters over short-term and multi-week windows relative to each individual’s baseline. We trained an XGBoost classifier with a patient-level 80/20 split. The primary analysis was performed in those with uncontrolled asthma (ACT19) on the held-out test set. We reported ROC AUC (95% CI) and Brier score. A single threshold chosen on training was carried forward to test for sensitivity/specificity. Separately, in an exploratory analysis, we used the same approach to classify whether the model can predict OCS-exacerbations during the 52-week follow-up. This analysis used the same 80/20 split and was evaluated using ROC AUC. Results Our model achieved AUC 0.878 (95% CI 0.867-0.890) for predicting CDEs 5-9 days ahead. Calibration was strong (Brier = 0.092). Using a training-selected operating point near p≈0.213, test performance was sensitivity 0.84 and specificity 0.79. We utilized the same machine learning algorithm to predict OCS-requiring exacerbations over the 52-week follow-up period and the model was not predictive (AUC=0.525). Conclusions In uncontrolled asthma, digital inhaler parameters provide a high-discrimination, well-calibrated signal for early deterioration. These findings reinforce that patterns over time in SABA use and inhalation characteristics can identify early markers of decline offering an opportunity for intervention. Predictive models can be used to develop alerts that could drive interventions such as automated outreach, adherence coaching, and changes to treatment during early clinical deterioration. This abstract is funded by: The original study was funded by Teva Pharmaceuticals.
Aslam et al. (Fri,) studied this question.