Forecasting Hospital Room and Ward Occupancy Using Static and Dynamic Information Concurrently: Retrospective...

Area of Science:

• Healthcare Management
• Data Science in Medicine
• Hospital Operations Research

Background:

• Accurate prediction of hospital bed occupancy rate (BOR) is critical for resource management, budgeting, and patient care.
• While hospital-wide BOR prediction is important, granular predictions for specific wards and rooms offer greater practical utility for scheduling.

Purpose of the Study:

• To develop a web-based tool for hospital administrators to predict BOR at the ward and room level across different timeframes.
• To provide actionable insights for improving hospital bed management and resource allocation.

Main Methods:

• Time-series forecasting using Long Short-Term Memory (LSTM) networks trained on hourly bed data.
• Development of ward-level models with 7- and 30-day windows, and room-level models with 3- and 7-day windows.
• Integration of static (room-specific) and dynamic data to enhance prediction accuracy, utilizing both LSTM and Bidirectional LSTM (Bi-LSTM).

Main Results:

• Bidirectional LSTM (Bi-LSTM) models demonstrated superior performance over standard LSTM.
• Ward-level models achieved an R-squared of 0.544 (MAE: 0.067, MSE: 0.009, RMSE: 0.094).
• Room-level models incorporating static data achieved an R-squared of 0.600 (MAE: 0.129, MSE: 0.050, RMSE: 0.227).

Conclusions:

• High-performance predictive models for ward and room-level BOR were successfully developed.
• A web-based dashboard enables visualization of predictions, supporting efficient bed operation planning.
• The models contribute to optimizing hospital resources and reducing overall resource consumption.