(705) Multilabel Deep Learning-Based Classification of Critical Deterioration Among Hospitalized Children

Introduction: Critical deterioration events, such as the initiation of mechanical ventilation (MV) or vasoactive infusion (VI), are a significant cause of mortality and morbidity among hospitalized children. We recently developed a model that predicts the composite risk of critical deterioration events. However, our model does not indicate the likelihood of each type of deterioration, which is crucial for accurate diagnosis and treatment. Here, we derive and validate a new multilabel deep learning classifier that simultaneously predicts the risk of multiple deterioration events

Methods: We conducted a retrospective study of all pediatric admissions to the University of Chicago Comer Children’s Hospital from 2009 to 2019. Data were split longitudinally into derivation ( < 2018) and validation (2018-19) cohorts. Predictors included age, hospital location (ED, ward, or ICU), vital signs, laboratory results, and prior comorbidities. Primary outcomes were MV, VI, and death within 12 hours of a vital or lab recording. We trained a multi-label recurrent neural network architecture to return predictions for all three outcomes. Model performance was compared to the composite model and outcome-specific models using the area under the receiver operating characteristic curve (AUC) as the primary metric.

Results: Out of 55,619 admissions in the overall cohort, 3,551 (6.4%) patients experienced MV, 1,368 (2.5%) had VI, and 313 (0.6%) patients died in hospital. The multilabel classifier demonstrated equivalent or better performance compared to the composite model and outcome-specific classifiers across all three outcomes (MV: 0.87 [0.87-0.87] vs. 0.86 [0.86-0.86], vs. 0.87 [0.86-0.87; VI: 0.84 [0.83-0.84] vs. 0.79 [0.79-0.79], vs. 0.83 [0.83-0.83]; death: 0.97 [0.97-0.98] vs. 0.97 [0.97-0.98], vs. 0.98 [0.97-0.98]). Variable importance revealed hospital location, FiO2, heart rate, and systolic blood pressure to be important across all outcomes.

Conclusions: We demonstrate that a multilabel classifier that predicts multiple outcomes simultaneously performs similarly to models that predict each outcome or models that predict the composite of outcomes. Our study lays the foundation for developing a single model in a pediatric hospital for predicting various outcomes, obviating the need for implementing several models