(383) Development and Validation of a Predictive Model for Non-Escalation in Bronchiolitis

Introduction: Over the last decade, the use of non-invasive ventilation (NIV) for bronchiolitis has increased, straining ICU capacity during viral surges and delaying ICU admissions. To support safe critical care outside the ICU and identify a data-driven method to triage patients, we hypothesized that EHR data could be used to predict children with bronchiolitis initiated on NIV in the emergency department who are unlikely to require escalation.

Methods: We developed and temporally validated a predictive model at a single center. We included children with bronchiolitis who were less than 3 years of age, admitted from 2011-2024, and started and maintained on low-level NIV (expiratory/inspiratory positive airway pressure ≤ 12 and fiO2 ≤ 0.40) for at least 3 hours. The outcome was not requiring escalation above low-level NIV at any point during hospitalization. The model was trained on admissions from 2011-2022 and validated on a temporally distinct cohort admitted from 1/1/23 -7/31/2024. We trained an extreme gradient-boosted model using EHR data, 5-vold cross validation, and a grid search for hyperparameter tuning. Missing data were imputed with the median value. Evaluation metrics included Area Under the Receiver Operating Characteristic (AUROC), sensitivity, specificity, positive predictive value (PPV), and calibration curve.

Results: The training cohort included 2,199 patients, and the validation cohort included 318 patients. In the training cohort, 71.5% were classified as non-escalators, compared to 63.2% in the validation cohort. The model used 88 predictor variables, including demographics, vitals, respiratory assessments, laboratory data, blood gases, and therapies. The model achieved an AUROC of 0.78 (95% CI 0.73-0.83). The probability threshold was adjusted to a specificity of 85% and resultant sensitivity of 45%, and PPV of 0.84 for predicting non-escalators. Visual inspection of calibration curve demonstrated good alignment.

Conclusions: Our model demonstrated modest accuracy and good calibration to predict non-escalation of support for children with bronchiolitis initiated on low NIV. We achieved adequate specificity and PPV to identify patients that may be best suited for monitoring outside of an ICU during times of high census. Future steps include optimization and real-world implementation.