Emergency departments have always relied on experience and intuition to predict how busy their day will be. Thoughts like "Will today be hectic?" or "It seems quiet right now" cross the minds of staff daily—if not aloud, then at least in their own internal dialogue. But how accurate are these estimates compared to actual patient numbers? And can our forecast do better?
In this article, we take a deep dive into the alternatives to our short-term forecast, PraeSight— and reveal exactly what it offers that other methods fail to capture.
Historically, emergency department staff have had two ways to prepare for incoming patient flow:
The longer staff work in an emergency department, the better they should, in theory, become at predicting busy periods, managing sick leave, and calling in extra personnel to optimize resources. Yet, stress and burnout still lead to sick leave—and in some cases, staff shortages and high turnover. So, can we do better? Can we get ahead of these challenges?
One of Praemostro’s key products, the short-term forecast PraeSight, has been running in a major Danish emergency department for nearly three years. To test whether staff could predict patient flow accurately, we turned off the system for one month in 2024. The goal was to compare predictions from both the staff and our system against actual patient numbers. Twice a day—at 8 AM and 4 PM—we called the attending specialist doctor and the experienced nurse in charge of the department and asked them two questions:
Meanwhile, we tracked the actual number of patients arriving each hour and ran PraeSight in the background—without revealing its predictions to the staff.
We then compared the staff’s predictions to the actual patient numbers and calculated the difference for each 8-hour shift.
Figure 1: Deviations between staff forecasts and actual patient arrivals.
In Figure 1 above, you can see what we discovered. The orange bars show the difference in the number of arriving patients compared to the doctors’ estimates. The blue bars show the same discrepancy, but from the nurses’ perspective.
Specifically, for example, on March 1 during the day shift, the doctor overestimated by 7 patients, while the nurse underestimated by 1 patient compared to the actual arrivals. A notable outlier is the day shift on March 15, where the doctor and nurse disagreed by more than 40 patients, despite both stating that they expected a normal shift.
In the example above, we focus on how often staff predictions were correct within a margin of one patient per hour, measured over 8 hours (i.e., +/- 8 patients per shift). Over the three years that PraeSight has been in operation, the system has consistently predicted within +/- 1 patient per hour accurately (95% of the time measured over 8 hours, meaning +/- 8 patients per shift) in the given department.
The staff correctly predicted within +/- 8 patients per shift 58% of the time, while they were incorrect in 42% of shifts. In comparison, PraeSight was correct 95% of the time. Additionally, when PraeSight’s predictions were off, the deviation was smaller than when staff predictions were inaccurate.
Another common approach to forecasting patient arrivals is looking at last week’s numbers:For example, if 8 patients arrived between 1 PM and 2 PM last Monday—and there are no extraordinary circumstances—one might assume 8 patients will arrive again today.
While historical data is a useful reference, it doesn’t consider key factors such as epidemics, weather conditions, and special events—all of which PraeSight integrates.More importantly, PraeSight automatically analyzes how these factors interact with one another, significantly improving accuracy.
Using the Mean Absolute Scaled Error (MASE) metric , we found that PraeSight predicts patient arrivals 25-35% more accurately than using last week's numbers alone.For reference, research literature suggests that a MASE value below 0.7 is difficult to achieve in similar forecasting scenarios. In our test department, PraeSight consistently achieved MASE values between 0.65 and 0.75, meaning there is very little room for further improvement.
Some emergency departments might consider creating their own forecasting models using Excel spreadsheets. Basic forecasting methods like simple weighted averages, exponentially weighted averages, or linear regression could be used to estimate future patient numbers. However, we don’t see this as a viable alternative to PraeSight for a simple reason: it’s difficult to measure and compare accuracy reliably.
Technically, Excel could replicate some of our calculations, but the process would be slow, cumbersome, and overly simplified—reducing its effectiveness. To maintain transparency and credibility, we’ve chosen not to compare PraeSight directly with homemade Excel models.
While PraeSight is highly accurate, it is not 100% infallible.As outlined above, there is a 5% chance that forecast precision will drop below ±1 patient per hour.When discrepancies occur, we analyze the root cause and make adjustments where possible, but no forecasting tool can guarantee a perfect prediction every time.That’s why PraeSight—along with Praemostro’s other products—should be seen as a powerful decision-support tool, rather than an exact 1:1 predictor.
If you have questions about our forecasts, feel free to reach out to us..
Did you know that researchers have actually studied, whether saying "It’s quiet in here..." in an emergency department causes more patients to arrive? A randomized controlled trial was conducted to investigate this superstition—and the results? It makes no difference. So, in this case, superstitions can be put to rest.
