A deep learning model for real-time mortality prediction in critically ill children.

Area of Science:

• Critical care medicine
• Machine learning in healthcare
• Pediatric intensive care

Background:

• Intensive care units (ICUs) generate vast amounts of data, offering opportunities for critical care advancements.
• A machine learning model, the Pediatric Risk of Mortality Prediction Tool (PROMPT), was developed for real-time prediction of mortality in pediatric ICUs.

Purpose of the Study:

• To develop and validate a machine learning-based tool for predicting all-cause mortality in pediatric intensive care units.
• To assess the predictive performance of the PROMPT tool using real-world data.

Main Methods:

• A convolutional neural network machine learning algorithm was used for model development and validation.
• Two retrospective observational cohorts were utilized: a development cohort (1445 patients) and a validation cohort (278 patients).
• Data included seven vital signs, patient age, and body weight upon ICU admission.

Main Results:

• PROMPT achieved high predictive accuracy for mortality 6 to 60 hours prior to death, with an area under the receiver operating characteristic curve of 0.89-0.97.
• The model demonstrated high sensitivity and specificity, outperforming the conventional Pediatric Index of Mortality scoring system.
• Model performance was consistent across both development and validation cohorts.

Conclusions:

• PROMPT is a deep model-based, data-driven early warning score.
• The tool can predict mortality in critically ill children.
• PROMPT may facilitate timely identification of deteriorating patients in pediatric ICUs.