Emergency-department crowding is not only an operational problem. It affects waiting, staffing, bed availability and, ultimately, patient outcomes. This project explored whether a hospital could see demand early enough to act on it.
I used the electronic health record at Addenbrooke’s Hospital to construct a minute-by-minute dataset spanning two years. From the preceding 96 hours of activity, a deep neural network forecast emergency-department occupancy, admissions and breach performance for every hour in the following day.
The resolution of the dataset mattered. Demand does not move in a simple straight line: it reflects time of day, day of week, recent arrivals, hospital flow and other interacting signals. The model was designed to learn those non-linear patterns and was evaluated against established statistical forecasting methods.
The most useful lesson was that a prediction is only valuable if it arrives in time to change a decision. A forecast of likely admissions can support bed planning; a forecast of crowding can inform staffing and escalation. That connection between model output and a real workflow has remained central to how I approach clinical AI.
What I built
- A two-year, minute-level dataset of emergency-department activity.
- A multi-output forecasting model for occupancy, admissions and performance.
- A comparison with conventional statistical approaches, focused on operational usefulness.