Hospital Bed Occupancy and Admissions Forecasting Dashboard

The problem hospitals face with bed pressure

When a hospital runs out of beds, patients wait in the emergency department, planned operations are cancelled and staff spend hours on the phone finding space. Much of this is predictable: Mondays are busy, winter is worse, discharges bunch up in the late afternoon. Yet in many hospitals, especially mid-sized private and public hospitals in developing countries, bed status lives on a whiteboard or in a spreadsheet updated twice a day. Nobody looks ahead more than a shift.

The raw data usually exists in the hospital information system: admission, discharge and transfer events. It is rarely modelled into a daily census, let alone forecast.

What already exists and the gap

• TeleTracking and Qventus sell patient-flow and capacity platforms to large hospital systems, mainly in the US and UK.
• Epic and other large electronic health record vendors include capacity dashboards for their own customers.
• Published research shows simple models work: a 2022 study of an NHS trust in Bristol found ARIMA models with day-of-week, holiday and weather inputs forecast admissions more accurately than the trust's existing method.

The gap: the research proves the idea, but mid-sized hospitals without an enterprise system have no affordable tool that turns their event data into occupancy history and a short-term forecast. This project builds that tool on open and synthetic data first, designed to connect to a real hospital system later.

What you build

Roles: bed manager (reads the forecast daily), ward in-charge (sees their ward), analyst (manages data and models).

Flows:

1. Admission, transfer and discharge events are loaded daily into staging tables.
2. dbt models build a census: which patient was in which ward at each hour.
3. Metrics are computed and shown on the dashboard.
4. Forecast models retrain weekly and publish 14-day predictions with intervals.
5. The bed manager sees red days and tests "what if 30% of discharges happened before noon".

The technically interesting part

Two parts. Building a correct census from events is harder than it looks: transfers create overlapping intervals, events arrive late or out of order, and a patient discharged and readmitted the same day must not be double-counted. Write tests that check the census against hand-worked examples.

Forecasting with honest evaluation: use rolling-origin backtesting, compare every model against a seasonal naive baseline, and report interval coverage (do 80% intervals contain the actual value about 80% of the time?). Explain why a small, explainable model may beat a complex one here, which is what the NHS studies found.

How to evaluate it

• Census correctness: match hand-calculated occupancy on 20 sampled ward-days exactly.
• Forecast accuracy: MAE and MAPE for admissions and occupancy at 1, 7 and 14 days ahead, versus seasonal naive and ARIMA.
• Intervals: coverage of 80% and 95% prediction intervals.
• Red-day alerts: precision and recall of days where occupancy passed the threshold.
• Users: one or more clinicians or managers review the dashboard; record what they would act on.

Suggested plan

1. Weeks 1 to 3: talk to a bed manager or nurse, read the published forecasting studies, design the event schema.
2. Weeks 4 to 8: synthetic data generation and calibration to published aggregates.
3. Weeks 9 to 12: census models and tests, operational metrics.
4. Weeks 13 to 19: baselines, forecasting models, backtesting framework.
5. Weeks 20 to 24: dashboard, red-day view, discharge what-if.
6. Weeks 25 to 28: user feedback, evaluation report, write-up.

Path to a product

The first customer is a mid-sized private hospital or a hospital group with a basic information system but no capacity tool. Sell an annual licence per hospital with set-up, including a connector for their system. Health departments running several public hospitals are a second market for a regional view.

Scope warning

Do not build an electronic health record or clinical decision support. Do not use real patient data without formal approval; synthetic and aggregate open data is enough for the project. Do not forecast individual patients' discharge dates in the core scope.

Learn what you need on Sefism

• Data Analytics roadmap : SQL, metrics and time-series analysis in order.
• Machine Learning roadmap : model evaluation habits for the forecasting part.
• dbt Documentation : tested census and metric models.
• OpenIntro Statistics : intervals and error measures explained simply.
• Storytelling with Data : a dashboard a busy bed manager reads in seconds.

Get help with this idea

The key decisions are how to generate believable synthetic data and which baselines to beat. If you want a second opinion on scope, stack or how to evaluate it before you pitch it to your supervisor, a mentor can help you shape it. You still build it; the guidance is about making the plan realistic and the results defensible. New here? create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session