AI for ED Staffing: Forecasts Are Not Shift Plans

The staffing huddle starts with a forecast: next Saturday may bring more emergency department arrivals than usual. The schedule is already posted. Should the charge nurse add coverage, protect the fast track, or ask hospital leaders to prepare for slower bed turnover? The number is useful. It is not an answer.

AI can help an emergency department see demand coming. But a forecast of arrivals is not a forecast of workload, and it is not proof that a staffing change will improve care. The safe role for a model is to inform a decision that a named clinician or operations leader still owns.

What an arrival forecast can tell you

A 2024 study tested machine-learning methods for forecasting daily emergency department arrivals across 11 departments in Australia, the United States, and the Netherlands. The researchers used historical data from 2014 through 2016 and forecast horizons of 7 and 45 days. Their strongest models varied across sites, and the study compared prediction error; it did not test whether changing staffing in response to a forecast reduced waits, crowding, or harm (Porto and Fogliatto, 2024).

That is a useful result, and a limited one. It shows that historical arrival patterns can support forecasts. It does not prove that a specific model will work in your ED today, on your shift, or after your patient mix and local processes change. The study’s data ended in 2016. A hospital considering a forecast should validate it on its own recent data before it changes a schedule.

The forecast horizon matters too. A 45-day estimate may help a leader review staffing templates or planned leave. A 7-day estimate may inform next week’s coverage. Neither is the same as a reliable hour-by-hour forecast for tonight’s resuscitation area. The study evaluated daily arrivals at those horizons; it did not establish real-time staffing recommendations (Porto and Fogliatto, 2024).

Volume is not workload

An arrival count treats unlike shifts as though they were the same. The same number of patients can require very different work: one shift may have more high-acuity arrivals, several patients awaiting admission, and fewer open treatment spaces. Another may have a steady stream of lower-acuity visits and reliable inpatient bed turnover. An arrival forecast alone does not distinguish those conditions.

Crowding is not measured in one standard way. An overview of 13 systematic reviews found that measures varied, causes fell across patient, staff, and system levels, and boarding of admitted patients was identified as a major contributor (Pearce et al., 2023). That matters for AI. A model can predict one part of pressure while leaving the rest unseen.

Before using an arrival forecast, pair it with local measures that describe the work the department must absorb. Depending on the decision, that may include arrival acuity, waiting-room volume, patients boarding for inpatient beds, open treatment spaces, staffing already on duty, and the expected skill mix. These are not interchangeable numbers. A useful dashboard should show which were forecast, which were observed, and which were not available.

There is another boundary to keep clear: an ED forecast cannot solve a hospital-wide bed shortage by itself. In 2024, the American College of Emergency Physicians asked CMS to require hospitals to plan and act when ED boarding passes a defined threshold, framing boarding as a hospital-level problem rather than a failure of ED efficiency alone (ACEP, 2024). Forecasting may give teams more time to coordinate. It cannot create staffed inpatient beds.

Put the forecast inside a human decision

The first question is not “How accurate is the model?” It is “What decision will this forecast change?” If the answer is unclear, do not put the score in front of a busy charge nurse and hope it helps.

For each use, write down the action, the person who may take it, and the conditions that should stop or override it. A forecast could prompt an early staffing review or a check-in with bed management. It should not silently cut scheduled coverage, deny a patient an assessment, or treat a predicted quiet shift as permission to ignore a worsening department.

Keep the forecast’s uncertainty visible. If the model reports a range, display the range and its recent local error, not only a single number. If it produces a point estimate, show how often actual arrivals have been materially higher or lower. Ask whether the error is small enough for the decision at hand. A one-patient miss may be acceptable for a long-range template review and unacceptable when deciding whether one clinician can safely cover an area.

Do not turn an arrival forecast into a clinician performance score. Forecasts are estimates about demand, not judgments about the quality or pace of an individual’s work. When reality differs from the prediction, the first response should be to understand the mismatch: was there an unusual event, a change in access, a model failure, or an operational constraint that the model never saw?

Test the workflow before relying on it

Start in silent review. Generate forecasts without using them to change coverage. Compare each prediction with actual arrivals, and record whether the forecast would have altered a real decision. This is a proposed safety step, not an intervention proven by the arrival-forecast study: that study evaluated prediction performance, not a live staffing workflow (Porto and Fogliatto, 2024).

Then run a small, time-limited pilot with a named operational owner. Choose a few outcomes before launch: forecast error, time to a clinician, waiting-room accumulation, boarding, overtime, and staff reports of workload. Review misses as well as apparent successes. A model that predicts average volume accurately but misses the rare surge may not be safe for the decision you want it to support.

Set a stop rule. Pause the forecast if the data feed is stale, the model changes without review, its error rises beyond a local limit, or staff cannot tell what the alert means. Keep a manual staffing process available. The forecast should make the team’s plan easier to adjust, not harder to question.

Finally, make the review cross-departmental. If the forecast points to pressure from boarding, invite the people who manage inpatient beds and hospital flow into the response plan. That matches the system-level framing of ACEP’s 2024 boarding proposal; an ED-only staffing change cannot address every source of delay (ACEP, 2024).

Dr. Chet's Take

I have spent more than 25 years in emergency medicine, and I know what an understaffed shift feels like from the bedside and the operations side. A useful forecast can buy a department time. It can help a leader ask the next question before the waiting room fills. But the model sees only the data we gave it. It does not see the patient who is about to get sicker, the boarded patient who needs a bed, or the nurse who is already covering two rooms.

That being said, I would not reject a forecast because it is imperfect. I would reject the idea that a prediction owns the schedule. The honest answer is that a model can be statistically good and still fail the local workflow. Put it in silent review first. Show the misses. Ask the charge nurses whether the alert would have changed a decision they could actually make. Then measure whether the change helped patients and staff, not whether the software produced a confident number.

If you are leading an emergency department, pick one decision the forecast could improve and name the person accountable for it. Set a stop rule before the first live alert. The schedule stays a human decision.

Key Takeaways

• Arrival forecasts can help teams prepare, but the published 2024 multi-site study measured predictions, not whether forecast-driven staffing improved care (Porto and Fogliatto, 2024).
• Arrival volume is only one part of ED pressure; crowding reviews describe varied measures and identify boarding as a major contributor (Pearce et al., 2023).
• Validate a model on current local data, show its uncertainty, and keep a named human responsible for every schedule change.
• Test the forecast in silent review before it influences coverage. Keep a manual process and a clear stop rule.

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download it here: AI in EM Survival Guide.

FAQ

Can AI predict emergency department arrivals?

Machine-learning models have been studied for daily ED arrival forecasts. A 2024 study tested data from 11 emergency departments and forecast horizons of 7 and 45 days, but it did not test a forecast-driven staffing intervention (Porto and Fogliatto, 2024).

Should an ED automatically change staffing based on an AI forecast?

No. Use a forecast as one input to a decision with a named human owner. Check local performance, display uncertainty, and keep a manual process available.

Why can an accurate arrival forecast still fail to help an ED?

It may predict patient counts without describing acuity, boarding, available beds, staffing mix, or other sources of crowding. Reviews report that ED crowding measures and causes are varied, with boarding a major system-level contributor (Pearce et al., 2023).

What should an ED measure during a staffing-forecast pilot?

Track local forecast error and the outcomes tied to the decision, such as waiting-room accumulation, time to clinician, boarding, overtime, and staff feedback. Decide in advance when to pause the tool.

Sources

• Bruno Matos Porto and Flavio Sanson Fogliatto. Enhanced forecasting of emergency department patient arrivals using feature engineering approach and machine learning. BMC Medical Informatics and Decision Making. 2024.
• Sabrina Pearce, Tyara Marchand, Tara Shannon, Heather Ganshorn, and Eddy Lang. Emergency department crowding: an overview of reviews describing measures, causes, and harms. Internal and Emergency Medicine. 2023.
• American College of Emergency Physicians. ACEP to CMS: Require Hospital Plans to Address Boarding in Emergency Departments. September 16, 2024.
• Related simulation training: EM-Sim.
• Books by Dr. Shermer.