Deep learning auto-triages pediatric ER patients

Using a dataset of records from nearly 3 million pediatric patients, South Korean researchers have developed and validated a deep-learning algorithm that can tell emergency doctors which children will need to be admitted to critical-care units.

Reporting their findings online July 1 in Pediatric Emergency Care, the team noted that, in their country, ER overcrowding is “a national crisis in which pediatric patients are often prioritized at lower levels.”

The researchers work at Mediplex Sejong Hospital and Sejong General Hospital, and their study used data from the ERs at those institutions and at 149 others.  

For identifying patients as needing critical care, their deep-learning technique achieved an area under the ROC curve (AUC) of 0.908.

This put it handily ahead of conventional systems for scoring pediatric early warnings (0.812) and triaging patients on observed acuity (0.782).

In addition, the deep-learning algorithm bested two machine-learning methods, random forest (0.88) and logistic regression (0.851).

The deep-learning algorithm also proved much more accurate than all other approaches at identifying patients who needed non-critical hospitalization.

The authors stated they set out on the present project because predicting ER prognoses, especially for pediatric patients, is “important but difficult.”