1175: A Machine Learning Computable Phenotype for Iatrogenic Withdrawal in Critically Ill Children

Introduction: Iatrogenic Withdrawal Syndrome affects over 100,000 children annually, particularly following mechanical ventilation termination. Current nursing-based screening is time-consuming and episodic, limited by staffing availability. We developed and validated a computable phenotype using standard electronic health record variables to identify withdrawal through autonomic dysregulation patterns and medication histories. Methods: This single-center, retrospective and prospective multi-cohort study included PICU patients from 2012-2025. Training/testing cohorts comprised 2,100 patients (2012-2023), with prospective validation in 400 patients (2023-2025). Inclusion criteria: extubation from mechanical ventilation and Withdrawal Assessment Tool (WAT-1) evaluation. Data were organized into 4-hour epochs. Variables included vital signs metrics (average, maximum, minimum, measurement frequency) and opioid/sedative exposures (cumulative dose, prior 4-hour dosage, average 4-hour dose). WAT-1 scores ≥3 defined withdrawal. Machine learning utilized XGBoost algorithm. Results: Test cohort performance: AUROC 0.87, AUPRC 0.62, calibration curve R2 0.96 (slope 1.08, intercept -0.02). Prospective validation maintained performance despite lower withdrawal prevalence: AUROC 0.85, AUPRC 0.35, calibration R2 0.93 (slope 1.19, intercept -0.09). Pearson’s correlation between computable phenotype risk and WAT-1 scores: r = 0.95 (p < 0.05). Prediction accuracy exceeded 87% at high/low WAT-1 values and ranged 54.8-73.3% near diagnostic thresholds. Conclusions: We successfully developed and validated a novel machine learning computable phenotype using standard EHR data for pediatric ICU withdrawal risk identification. This objective, personalized approach demonstrated clinically suitable performance metrics with potential for real-time monitoring while reducing nursing administrative burden. The model’s reliance on autonomic dysregulation indicators and medication profiles enables continuous, automated withdrawal surveillance in critically ill children.