(1444) Using Machine Learning to Predict Cardiac Arrest in the Pediatric Cardiac Intensive Care Unit

Introduction: Pediatric cardiac arrest (CA) occurs in >15,000 hospitalized patients each year in the United States and results in high morbidity and mortality. Although >90% of pediatric in-hospital CA events occur in the data-rich environment of an intensive care unit (ICU), these events remain difficult to predict. The objective of this study was to use machine learning algorithms to develop a CA prediction model from routinely collected pediatric cardiac ICU electronic health record (EHR) data.

Methods: Retrospective cohort study of all patients admitted to a single quaternary, academic pediatric cardiac ICU from 1/1/2010 to 12/31/2023. Patient demographics, vital sign and assessment flowsheets, laboratory values, and code blue flowsheets were extracted from the EHR. The primary outcome was CA. Admissions were split 50/50 to create two datasets with: 1) the observed incidence of CA in the full cohort (3.9%) and 2) a balanced incidence of CA (35%) through down-sampling of controls. Data from the first 24 hours of admission were used for model development, CA events in this time window were excluded. Laboratory features present in < 60% of admissions were removed, as were flowsheet features with text values. Models to predict CA were developed and evaluated using the following machine learning methods: 1) nearest neighbors; 2) random forests; 3) gradient boosting models; 4) tabular neural networks.

Results: There were 15,145 admissions, including 177 with CA events, that met inclusion criteria. A total of 3 demographic, 47 laboratory, and 24 vital sign and assessment features were used for model development. The best performing model was a tabular neural network, NeuralNetFastAI, the area under the receiver operating characteristic (AUROC) was 0.63 in the unbalanced dataset and 0.71 in the balanced dataset. After optimal threshold identification, the NeuralNetFastAI model had a positive predictive value (PPV) of 43% with a sensitivity of 85% in the balanced dataset, and PPV of 11% with low sensitivity (27%) in the unbalanced dataset.

Conclusions: The best performing model, a tabular neural network, performed well in the balanced dataset but not in the unbalanced dataset, which is more similar to real-world conditions. Further work is needed to improve model performance.