Bed occupancy forecast for decision-making during periods of high hospital pressure

Sudden increases in hospital demand can degrade healthcare services and impact local populations, and are characterised by high bed occupancy. Adaptive policies involving weekly resource and staff allocation can mitigate these effects, but selecting adaptive strategies requires forecasting tools that are seamlessly integrated within hospital decision-making processes.

To develop and deploy a forecasting tool predicting imminent bed capacity saturation to support short-term adaptive policies.

We involved two major hospitals in the healthcare system centred around Bristol (England). We used routinely collected hospital-level aggregated and environmental data to predict the risk of bed capacity saturation up to seven days ahead. Risk was defined as the probability of crossing a predefined bed occupancy threshold for individual or aggregated days, and was computed from forecast ensembles, combining bed occupancy forecasts from different statistical and machine learning base models selected via cross-validation. We explored different weighting schemes to obtain forecast ensembles. The tool was deployed in a multi-centre operational setting scheduled to update predictions on a daily basis. We devised a dashboard condensing all relevant information to efficiently communicate risk predictions to decision-makers.

We identified hospital-specific subsets of base models for forecast ensembles. When evaluating risk prediction performance, ensembles were more reliable than individual base models. We did not find substantial performance differences between ensemble weighting schemes, hence we used a linear combination with equal weights during deployment. Feedback from senior managers suggests that our tool could help support adaptive policies.

We developed and deployed a tool to support short-term, high-level adaptive responses to impending hospital capacity saturation. The tool is efficient yet simple and will be extended to include more hospitals. This work contributes to improving data integration and analysis for efficient resource management and decision-making.