Forecasting hospital bed occupancy: a time series approach with prophet

Area of Science:

• Healthcare Operations Research
• Time Series Analysis
• Machine Learning in Healthcare

Background:

• Accurate hospital bed occupancy forecasting is crucial for resource planning and patient flow.
• Complex machine learning models often present high maintenance costs and limited interpretability in healthcare.
• This study evaluates Prophet, a parsimonious time-series model, for mid-term hospital bed occupancy forecasting.

Purpose of the Study:

• To assess the performance and practicality of the Prophet model for mid-term hospital bed occupancy forecasting.
• To compare Prophet's accuracy and operational value against more complex forecasting models.
• To demonstrate the interpretability and utility of Prophet's components in a real-world healthcare setting.

Main Methods:

• Applied the Prophet model to daily bed occupancy data (2010-2023) from a university medical center.
• Incorporated public holidays and a COVID-19 indicator as exogenous regressors.
• Assessed forecast accuracy using rolling cross-validation for horizons of 30 to 180 days and implemented a production-ready pipeline.

Main Results:

• Prophet achieved low Mean Absolute Percentage Error (MAPE) values (3.21%-3.53%) with over 80% coverage across all forecast horizons.
• Accuracy was comparable to or better than complex models, with significantly lower computational and operational costs.
• Component analysis revealed interpretable patterns aligned with hospital operations, such as weekly/yearly cycles and holiday effects.

Conclusions:

• Prophet provides an accurate, interpretable, and practical solution for mid-term hospital bed occupancy forecasting.
• Simple models like Prophet can rival complex architectures in accuracy, reproducibility, and scalability for structured forecasting tasks.
• Prophet's minimal tuning, faster deployment, and clear insights offer significant operational value in healthcare settings.