ED Volume Forecasting for Better Resource Allocation & Enhanced Patient Care
Click here for code
.png)
Background
Emergency Departments (EDs) are unpredictable by nature. Patient volume can surge without warning or follow seasonal and event-driven trends such as flu season, holiday weekends, or large-scale local gatherings. These fluctuations create a challenge: how do you plan staffing and resource allocation when demand is constantly shifting? Before this project, our team lacked a reliable, data-driven way to predict patient volume across campuses. Decisions were often made based on historical averages or anecdotal experience, which left us vulnerable to both over-preparation and under-preparedness. I set out to change that with a forecasting system that could bring structure to the chaos — one that used advanced time series modeling to make our emergency departments smarter, faster, and better prepared.
Challenges
Each hospital location had distinct seasonal and operational rhythms. Yet, planning for these fluctuations was based on retrospective guesswork rather than a formal, scalable model. There was no consistent way to incorporate external drivers like holidays or local events into volume planning, which left operations unprepared during predictable surges. I saw an opportunity to build a robust forecasting system — one that could reflect the real-world complexity of emergency care while remaining operationally usable across five hospital sites.
Solution
I developed a dual-model forecasting framework using both ARIMA and ARIMAX methodologies to predict daily ED volumes across five campuses. First, I performed seasonal decomposition on each site’s historical patient visit data to isolate trend, seasonal, and residual components using Python’s statsmodels. From there, I built two separate, full-scale forecasting pipelines: one using traditional ARIMA modeling, and the other using ARIMAX, which incorporated exogenous variables such as holidays and major local events.
Each model was trained and tuned independently per hospital to account for differing autocorrelation patterns and seasonal effects. I then compared results across campuses using RMSE and MAE to evaluate accuracy, and ultimately deployed the forecasts into an operational dashboard used by leadership to guide staffing, resource allocation, and surge planning.
Key Features
-
Dual Forecasting Pipelines
Developed both ARIMA-based and ARIMAX-based models for each hospital to compare the influence of external variables and select the most performant solution. -
Seasonal Decomposition
Decomposed each time series into trend, seasonal, and residual components to better isolate and model patterns in ED volume. -
Exogenous Variable Integration
Incorporated external factors into the ARIMAX models, allowing forecasts to adjust dynamically based on known demand drivers. -
Location-Specific Modeling
Created and tuned models independently for all five hospital campuses to reflect their unique operational signatures.
Process
The project began with collecting and merging multiple datasets — including multi-year ED visit data, holiday calendars health outcome, and economic data. I cleaned and structured the data to ensure temporal consistency across sites, handling missing values, irregular time steps, and anomolies.
Next, I applied seasonal decomposition to break down each location’s volume history into trend, seasonal, and residual components. I built two comprehensive models per campus: one ARIMA-based and one ARIMAX-based. Each model was created by modeling each component of the time series directly and combining the results, with the ARIMAX models incorporating exogenous factors to account for volume shifts tied to external events.
Model development was done in Python using statsmodels and pmdarima, with parameter tuning based on AIC and BIC criteria. I validated model performance using standard metrics such as RMSE and MAE, then compared results between ARIMA and ARIMAX models to assess the added value of external variables. The final forecasts were integrated into Tableau dashboards that refresh on schedule and are used by operational leaders to anticipate daily patient volumes and plan staff coverage proactively.
Impact
This project empowered our EDs to shift from reactive, anecdotal planning to proactive, data-informed decision making. Forecasts are now used across five campuses to guide daily staffing, resource planning, and even long-term hiring strategies. The integration of external variables like holidays and events helped reduce the element of surprise, while the location-specific tuning ensured each forecast reflected on-the-ground reality.
Most importantly, the benefits extend directly to patients. With more consistent and accurate staffing aligned to forecasted demand, we’ve been able to reduce wait times, improve patient throughput, and support our clinical teams with better resource allocation during busy periods. This forecasting initiative laid the foundation for more scalable, intelligent operations — and set a new standard for how we prepare for seasonal care delivery challenges.