Objectives: Accurate prediction of time to death after withdrawal of life-sustaining therapies may improve counseling for families and help identify candidates for organ donation after cardiac death. The study objectives were to: 1) train a long short-term memory model to predict cardiac death within 1 hour after terminal extubation, 2) calculate the positive predictive value of the model and the number needed to alert among potential organ donors, and 3) examine associations between time to cardiac death and the patient’s characteristics and physiologic variables using Cox regression. Design: Retrospective cohort study. Setting: PICU and cardiothoracic ICU in a tertiary-care academic children’s hospital. Patients: Patients 0–21 years old who died after terminal extubation from 2011 to 2018 (n = 237). Interventions: None. Measurements and Main Results: The median time to death for the cohort was 0.3 hours after terminal extubation (interquartile range, 0.16–1.6 hr); 70% of patients died within 1 hour. The long short-term memory model had an area under the receiver operating characteristic curve of 0.85 and a positive predictive value of 0.81 at a sensitivity of 94% when predicting death within 1 hour of terminal extubation. About 39% of patients who died within 1 hour met organ procurement and transplantation network criteria for liver and kidney donors. The long short-term memory identified 93% of potential organ donors with a number needed to alert of 1.08, meaning that 13 of 14 prepared operating rooms would have yielded a viable organ. A Cox proportional hazard model identified independent predictors of shorter time to death including low Glasgow Coma Score, high Pao 2-to-Fio 2 ratio, low-pulse oximetry, and low serum bicarbonate. Conclusions: Our long short-term memory model accurately predicted whether a child will die within 1 hour of terminal extubation and may improve counseling for families. Our model can identify potential candidates for donation after cardiac death while minimizing unnecessarily prepared operating rooms.

Supplemental Digital Content is available in the text.

Electronic Medical Records (EMR) are a rich source of patient information, including measurements reflecting physiologic signs and administered therapies. Identifying which variables are useful in predicting clinical outcomes can be challenging. Advanced algorithms, such as deep neural networks, were designed to process high-dimensional inputs containing variables in their measured form, thus bypass separate feature selection or engineering steps. We investigated the effect of extraneous input variables on the predictive performance of Recurrent Neural Networks (RNN) by including in the input vector extraneous variables randomly drawn from theoretical and empirical distributions. RNN models using different input vectors (EMR variables only; EMR and extraneous variables; extraneous variables only) were trained to predict three clinical outcomes: in-ICU mortality, 72-hour ICU re-admission, and 30-day ICU-free days. The measured degradations of the RNN's predictive performance with the addition of extraneous variables to EMR variables were negligible.

Viewing the trajectory of a patient as a dynamical system, a recurrent neural network was developed to learn the course of patient encounters in the Pediatric Intensive Care Unit (PICU) of a major tertiary care center. Data extracted from Electronic Medical Records (EMR) of about 12000 patients who were admitted to the PICU over a period of more than 10 years were leveraged. The RNN model ingests a sequence of measurements which include physiologic observations, laboratory results, administered drugs and interventions, and generates temporally dynamic predictions for in-ICU mortality at user-specified times. The RNN's ICU mortality predictions offer significant improvements over those from two clinically-used scores and static machine learning algorithms.