Introduction: Despite widespread use of noninvasive positive pressure ventilation (NIPPV) in bronchiolitis, predicting which children require intubation remains difficult. This study uses five second interval physiologic data from the Etiometry Platform to identify early predictors of NIPPV failure, offering a novel approach beyond intermittent vitals or static clinical variables. Methods: Patients under 2 years old admitted to the PICU at Nemours Children’s Health, DE, between Jan 2023- June 2024 with viral bronchiolitis treated with NIPPV were included and categorized by “failure” (intubated) or “success” (weaned without escalation). Demographic, clinical, and continuous high frequency physiologic data (HR, RR, SpO2, BP) were collected from 74 pediatric ICU patients with RVP-positive bronchiolitis and included 63 patients in the analysis after excluding those with insufficient vital sign data or length of stay less than 16 hours. Vital signs were segmented into 4-hour observation windows with a 12-hour prediction horizon using a 1-hour sliding window. Summary statistics (mean, maximum, minimum, variability, and trend) were extracted and combined with demographic variables and illness severity scores. We compared logistic regression, random forest, XGBoost, and multilayer perceptron models (MLP) using nested cross-validation. Results: Of the Seventy-four patients who met criteria; 19 (25.7%) were intubated. Failure was associated with longer PICU stay (median 8.4 vs. 2.3 days, p< 0.001), prematurity (p=0.009), and history of pulmonary disease (p=0.017). Logistic regression achieved the best predictive performance (AUROC 0.78) with balanced sensitivity and specificity. SHAP (Shapley Additive exPlanations) analysis identified higher respiratory rate, higher mean noninvasive blood pressure, increased PRISM 3 probability of death, younger age, lower weight, and greater SpO2 variability as the most important predictors of intubation. This machine learning approach may enable earlier and more accurate risk stratification for intubation in critically ill children with bronchiolitis. Conclusions: Continuous monitoring data via Etiometry with application of a SHAP based machine learning algorithm may help identify patients at risk of NIPPV failure earlier in their PICU course and guide escalation decisions.
McCormick et al. (Sun,) studied this question.