Predicting risk before it becomes harm in the pediatric ED - Children's National

In a pediatric emergency department (ED), crowding is not just an inconvenience. It can become a safety issue, especially when families leave before a clinician can evaluate a seriously ill child. Historically, "left without being seen" has been more common among lower-acuity visits, but recent staffing shortages and inpatient boarding have changed that reality. High-acuity children are increasingly affected, raising the stakes for ED teams trying to manage flow in real time.

The question our teams set out to answer was practical and operational: Can we use machine learning to predict when the risk of high-acuity patients leaving without being seen is about to rise, early enough to intervene?

Building a model for real-world decisions

In new open-access research, investigators at Children's National Hospital and Virginia Tech analyzed 512,616 emergency department visits from 2018 to 2024. Instead of predicting whether a specific patient might leave, the team focused on system-level forecasting. The goal was to estimate how many high-acuity children (Emergency Severity Index 1-3) would arrive and then leave without being seen over the next eight hours.

That eight-hour window was intentional. It aligns with typical shift schedules and staffing decisions, making the predictions actionable for ED leaders. The models drew on routinely available operational data, including arrivals, current census, waiting times, boarding burden and components of the National Emergency Department Overcrowding Score (NEDOCS).

What worked best

Several approaches were tested, from a simple model using ED census alone to more advanced machine learning techniques. Two stood out. Gradient boosting (XGBoost) and temporal fusion transformers - strong prediction methods that work by building lots of small decision trees, where each new tree focuses on fixing the mistakes made by the earlier ones - consistently showed the strongest performance, accurately identifying periods when two or more high-acuity patients were likely to leave without being seen in the next eight hours.

These models also performed well during the busiest part of the day, from noon to 8:00 p.m., when staffing flexibility is often greatest. Importantly, combining multiple models into an ensemble did not improve performance, reinforcing the value of solutions that are both strong and deployable.

Feature analysis showed that recent changes in census, overall crowding scores, time of day and day of the week were among the most influential predictors. These are factors ED teams already track, but the model brings them together in a consistent, data-driven way that goes beyond operational instinct alone.

Why it matters

This work demonstrates how applied machine learning can support safer, more proactive care. By forecasting when risk is rising, ED leaders can better target interventions such as surge staffing, split-flow pathways or other crowding mitigation strategies before families feel pressure to leave.

Future work will focus on prospective validation, monitoring performance over time and understanding how prediction-driven interventions affect patient outcomes and operational costs. But the takeaway is clear. Predicting risk in advance gives teams a chance to act and that can make a meaningful difference for children who need care the most.

This research, A Machine Learning Strategy to Predict the Number of High-Acuity Children Who Leave Without Being Seen From the Emergency Department, was published in the Journal of the American College of Emergency Physicians (JACEP) Open. Authors from Children's National include Brandon Kappy, MD, MPP; Sarah Isbey, MD; James M. Chamberlain, MD; Kenneth McKinley, MD; Trang Ha, PhD and Marius George Linguraru, DPhil, MA, MSc, in collaboration with colleagues from Virginia Tech.