Machine Learning to Predict Cardiac Death Within 1 Hour After Terminal Extubation

Area of Science:

• Pediatric critical care medicine
• Artificial intelligence in healthcare
• Organ donation and transplantation

Background:

• Predicting time to death after withdrawal of life-sustaining therapies is crucial for family support and organ donation.
• Current methods lack precision, impacting end-of-life care and donor identification.

Purpose of the Study:

• To develop and validate a long short-term memory (LSTM) model for predicting cardiac death within 1 hour post-extubation.
• To assess the model's positive predictive value and number needed to alert for organ donation.
• To identify clinical predictors of rapid death after extubation.

Main Methods:

• Retrospective cohort study of 237 pediatric patients (0-21 years) in PICU and cardiothoracic ICU.
• Trained an LSTM model to predict death within 1 hour of terminal extubation.
• Utilized Cox regression to analyze patient characteristics and physiological variables.

Main Results:

• 70% of patients died within 1 hour of extubation; median time to death was 0.3 hours.
• LSTM model achieved an AUC of 0.85 and a positive predictive value of 0.81 at 94% sensitivity.
• Identified 93% of potential organ donors with a number needed to alert of 1.08, minimizing OR preparation.

Conclusions:

• The LSTM model accurately predicts death within 1 hour of extubation in children.
• This predictive tool can enhance family counseling and optimize organ donor identification.
• Low Glasgow Coma Score, high Pao2-to-Fio2 ratio, low pulse oximetry, and low serum bicarbonate independently predicted shorter time to death.