What is it about?
Emergency departments need to know how many patients are likely to arrive in the coming hours so they can plan staffing, beds, and other resources. We studied whether machine learning could improve the hospital’s existing method for predicting hourly patient arrivals. Using data from more than 35,000 emergency department visits at a Norwegian hospital, we compared several machine learning models over prediction periods ranging from 1 to 24 hours. Machine learning improved predictions at short time horizons, but some models performed poorly when predicting further ahead. A model designed to predict all 24 future hours at once avoided this problem and provided useful predictions up to 24 hours ahead. These results suggest that machine learning could support more proactive planning of emergency department capacity.
Featured Image
Photo by camilo jimenez on Unsplash
Why is it important?
What is particularly timely about this work is that emergency departments face increasing pressure to manage unpredictable patient demand, while staffing and capacity decisions must often be made hours in advance. We show that machine learning can improve hourly arrival forecasts compared with the hospital’s existing calendar-based approach, but also identify an important limitation of commonly used recurrent models: their performance can deteriorate substantially at longer forecasting horizons. Our results demonstrate that predicting the full 24-hour trajectory directly can overcome this problem and provide useful forecasts for longer-term planning. This offers a practical approach for integrating machine learning into existing hospital planning systems and could support more proactive management of emergency department capacity.
Perspectives
For us, the most rewarding aspect of this work was seeing a practical hospital problem become a machine learning problem with a clear operational purpose. Rather than focusing only on achieving the best possible model performance, we wanted to understand whether these methods could actually improve the forecasting system already used at the hospital. We found the discovery of the “frozen-feature problem” particularly interesting, because it shows how seemingly small design choices can have major consequences when models are used for longer-term forecasting. We hope this work encourages others to look beyond headline performance metrics and consider how machine learning models behave when they are taken from the research setting into real-world healthcare operations.
Mr Bjørn-Jostein Singstad
Akershus University Hospital
Read the Original
This page is a summary of: Machine Learning Approaches for Hourly Emergency Department Patient Arrival Forecasting: A Multi-Horizon Comparison with Operational Benchmarking at a Norwegian Hospital, August 2026, Springer Science + Business Media,
DOI: 10.21203/rs.3.rs-10805235/v1.
You can read the full text:
Contributors
The following have contributed to this page
