Emergency department surges affect patient flow, staffing, and access to care. Forecasting patient arrivals can provide an early view of changing demand. Probabilistic models estimate a range of possible arrivals, while constraint-aware staffing scenarios translate that range into operational options. Large language models can support the process by preparing those options for review. They do not replace clinical judgment or staffing authority.

Establishing the Forecasting Process

The forecasting process addresses data preparation, arrival prediction, workload estimation, and staffing review. It begins with reliable hourly arrival records and includes calendar features, operational conditions, and uncertainty estimates. The objective is not to predict a single number without qualification. It is to give emergency department leaders a usable view of likely demand and the range of outcomes that must be considered.

For a worked example, an emergency department has two years of hourly arrival data. The records include arrival time, triage category, day of week, and holiday status. The data team checks for missing hours, changes in operating policy, and periods when intake systems were unavailable. These checks help distinguish real demand changes from recording changes.

Modeling Patient Arrivals

Hourly patient arrivals are counts. A Poisson model provides a starting point, but arrival counts often vary more than a Poisson distribution allows. A negative binomial model can address this overdispersion and estimate both expected arrivals and prediction intervals.

The model uses hour of day, day of week, holiday status, recent arrival counts, and selected external indicators. Its output includes an expected count and a probability distribution for each forecast hour. For the next four hours, the worked example produces the following estimates:

  • Hour 1: 13 expected arrivals, with a 90% prediction interval of 8 to 19.
  • Hour 2: 16 expected arrivals, with a 90% prediction interval of 10 to 23.
  • Hour 3: 19 expected arrivals, with a 90% prediction interval of 12 to 27.
  • Hour 4: 17 expected arrivals, with a 90% prediction interval of 10 to 25.

The interval is operationally important. A plan based only on the expected count may not provide enough coverage when arrivals are higher than expected. A plan based on the upper bound for every hour may also use resources inefficiently. The forecasting process should therefore connect probability levels to explicit operating policies.

Translating Arrivals Into Workload

Arrival counts do not directly determine staffing needs. The department also considers acuity, expected service time, current occupancy, and the number of patients already waiting. In the example, historical data indicate that approximately 25% of arrivals require higher-acuity attention. The workload model combines predicted arrivals with this mix and with observed service times.

For Hour 3, the expected 19 arrivals include approximately five higher-acuity patients. The model estimates 28 nursing hours of demand across the hour and the following period, with a wider range under the 90% arrival scenario. This estimate is not a staffing order. It is an input to a capacity review that includes minimum safe coverage, role qualifications, shift rules, and available staff.

The workload estimate can be improved by tracking prediction error separately for arrival volume, acuity mix, and service time. This makes it possible to identify whether a forecast miss came from more arrivals than expected or from a change in the type of care required. The distinction supports more targeted model updates.

Generating Constraint-Aware Staffing Scenarios

The staffing process converts forecast demand into feasible options. It includes existing schedules, staff availability, credential requirements, break rules, overtime limits, and minimum coverage. A deterministic optimization model can evaluate these constraints and produce scenarios for review.

In the worked example, the scheduled team can support the expected workload through Hour 2. The Hour 3 forecast indicates a possible shortfall. The scenario engine evaluates several options:

  • Expected-demand scenario: maintain the current schedule and monitor arrivals at the next review point.
  • Elevated-demand scenario: request one qualified nurse for a four-hour period, subject to availability and approval.
  • High-demand scenario: request additional coverage and activate the department’s established surge procedures if arrivals or waiting-room measures cross defined thresholds.

Each scenario includes its assumptions, staffing impact, and trigger for review. The optimization process excludes options that violate credential, labor, or minimum-coverage requirements. Program managers can compare the cost and readiness of each option without treating a forecast as a directive.

Using an LLM to Prepare Scenarios

An LLM can support the workflow by summarizing forecast results and preparing staffing scenarios in a standard format. It receives structured inputs, such as forecast quantiles, current staffing, available personnel, and approved operating rules. It can then draft a concise explanation for the charge nurse or operations manager.

For example, the LLM may state that Hour 3 has the highest expected arrival count and that the 90% interval extends to 27 arrivals. It may describe the qualified coverage option returned by the optimization model and identify the trigger for reassessment. The language model does not create staff availability, waive a constraint, or approve a schedule.

To maintain control, the LLM should use only approved data fields and return a constrained output, such as a structured scenario with a summary, assumptions, and source values. A validation step checks that the narrative matches the forecast and optimization results. Human review remains part of the process, especially when data are incomplete or operating conditions change.

Evaluation and Operational Use

The system should be tested before operational use. Forecast evaluation includes count error, interval coverage, and performance by hour, day, and acuity group. If a nominal 90% interval contains actual arrivals only 70% of the time, the uncertainty estimates require recalibration. Staffing scenarios should also be reviewed for feasibility and usefulness by the people who will act on them.

Implementation can begin with a limited pilot. The team can compare forecasts with actual arrivals, record which scenarios were reviewed, and document whether the information supported timely decisions. These measures establish whether the system improves readiness, rather than only whether it produces accurate predictions.

Conclusion

Probabilistic patient-arrival models provide a range of expected emergency department demand. Constraint-aware optimization translates that range into feasible staffing scenarios, and LLM automation can present the results in a consistent form. Starting with a measured pilot, validated operating rules, and human review supports practical integration. Dills Analytics LLC can help organizations design this process and connect forecasting, computational modeling, and actionable AI solutions in high-consequence settings.

AI generated content label