South Korean researchers build AI tool to flag which sick kids need emergency care

Researchers in South Korea are building an AI-powered smartphone app to help hospital staff decide whether a sick child needs emergency care. The tool reads free-text clinical notes from electronic medical records and predicts how urgently a patient needs treatment, before any lab results come back.

The project brings together researchers from Catholic University of Korea Seoul St. Mary's Hospital, Korea University's Department of Artificial Intelligence, Asan Medical Center and medical AI company VUNO. Their findings were published in Scientific Reports.

The timing matters. South Korea is dealing with serious overcrowding in emergency departments and a growing shortage of pediatric specialists. Getting triage decisions right, and getting them quickly, has become a practical problem with real consequences for patients and staff alike.

How does it work?

The AI model uses natural language processing to read symptom descriptions and treatment details written by medical staff in free-text clinical notes. The key point is that it only uses information available before test results are returned, which is exactly when triage decisions are hardest.

The team trained the model on data from around 87,759 patients under 18 who visited a pediatric emergency department between 2012 and 2021. Rather than using the standard five-level triage scale, they classified patients based on the care they actually received:

• Emergency cases: patients who had at least one blood test, urine test, intravenous fluid therapy, inhalation therapy, emergency medication or hospital admission
• Non-emergency cases: patients discharged with only an oral medication prescription and no testing or further treatment

The model is built on Korean Medical-BERT, a local version of Google's BERT language model trained on Korean medical text. It was further trained using masked language model pre-training on clinical notes.

The resulting model, called KM-BERT with MLM, achieved an AUROC score of 84% and an AUPRC score of 88%, outperforming other machine learning models tested in the study. It also showed stronger predictive accuracy than the Korean Triage and Acuity Scale (KTAS), the classification system currently used in Korean emergency departments.

Why does it matter?

Triage is difficult under pressure. The KTAS system, while widely used, relies on a relatively coarse scale and is vulnerable to evaluator subjectivity. As Dr Changhee Lee, co-corresponding author and assistant professor at Korea University Medical Center, put it: "When KTAS scores alone are used to predict which patients require an ER visit, the discriminative power is not sufficient."

The researchers argue that free-text clinical notes contain early signals about a child's condition that structured data like vital signs often miss. By reading those notes automatically, the AI model can pick up on patterns that reflect how experienced emergency medicine specialists actually assess patients.

Dr Woori Bae, who led the study and directs the Pediatric Emergency Medical Center at Seoul St. Mary's Hospital, said the model could support more efficient use of resources and improve patient safety. The practical goal is a smartphone app that gives medical staff a prediction they can factor into their decision-making.

The team is also planning a multi-center validation study using external datasets, and preparing additional validation with data collected after the COVID-19 pandemic to check how well the model holds up in more recent conditions.

The context

Pediatric emergency care has a specific problem that makes triage harder than it is for adults: children often cannot describe their own symptoms clearly. That makes early, accurate assessment more difficult and increases the risk of either under-treating a serious case or sending families to the emergency room unnecessarily.

Unnecessary ER visits are a real issue. As Dr Lee noted, "there are many cases in which children are brought to the ER even when emergency care is not actually required, contributing to overload of emergency medical resources and a potential decline in the quality of care."

South Korea has already started looking for technology-driven solutions. Last year, the Ministry of Health and Welfare piloted a 24-hour app-based pediatric counseling service called Ai Ansim Talk, which connects parents of children under 12 to pediatric and emergency specialists from three major hospitals for home care guidance. The AI triage model described here is a different approach to the same underlying problem: too many patients, not enough specialists, and not enough time to make the right call.

Emergency department overcrowding is a global issue, not just a Korean one. AI tools that can support clinical decision-making at the point of triage, using data that already exists in hospital records, are attracting serious research attention. This study is one of the more concrete examples of that work moving toward a real clinical tool.