0858 Development and Validation of a Predictive Model for ICU Needs in Children with Severe OSA after Adenotonsillectomy: A Pilot Study

Abstract Introduction To develop and evaluate a predictive model to risk-stratify pediatric patients with severe obstructive sleep apnea (OSA) undergoing adenotonsillectomy, identifying those requiring post-operative ICU stay based on pre-surgical characteristics and polysomnogram (PSG) parameters. Methods A retrospective review of 352 pediatric patients (< 18 years) who underwent adenotonsillectomy with an apnea-hypopnea index (AHI) > 40 on overnight PSG between 2021 and 2024 at a single children’s hospital was conducted. Four machine learning models (Logistic Regression, Decision Tree, Random Forest, and XGBoost) were developed to predict post-operative ICU admission. Data was split into training (80%) and testing (20%) sets. A pipeline was created for preprocessing, including scaling, log transformation of continuous variables, and subsequent oversampling via SMOTE to address class imbalance. Hyperparameters were optimized via randomized search with 5-fold cross-validation targeting recall. Model performance was assessed using area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (PR-AUC), Recall, F1-score, Accuracy, and Precision. Feature importance and odds ratios were examined for interpretability. Results In a 72-patient test-cohort (8 ICU and 64 non-ICU), logistic regression identified higher OAHI (OR=167.65, 95%CI 5.88–4775.52, p=0.003) and a positive respiratory viral panel (OR=22.35, 95%CI 4.47–111.64, p< 0.001) as risk factors for ICU stay. Protective factors included higher nadir oxygen saturation (SpO2 nadir; OR=0.08, 95%CI 0.01–0.55, p=0.01) and older age (OR=0.11, 95%CI 0.02–0.64, p=0.014). BMI-Z score was not a significant predictor (OR 3.19, 95%CI 0.20-51.75, p=0.41). Random Forest and Logistic Regression were top performers with AUC of 0.88 and 0.86, respectively. Random Forest was the best performing model (recall 75%; F1-score 0.48), followed by logistic regression (recall 75%; F1-score 0.41). Conclusion This pilot study validates the feasibility of using machine learning to risk-stratify pediatric OSA patients for post-operative ICU admission based on demographic and PSG predictors. The high recall and transparent explainability are key factors for fostering clinical trust. As a proof of concept, this work establishes a baseline, though predictive power is limited by the dataset size. Future models trained on larger multi-institutional datasets and validated in randomized trials may offer improved performance and generalizability, serving as tools to augment clinical decision making. Support (if any)