Authors:
Mattis Hartwig
1
;
Simon Schiff
2
;
Sebastian Wolfrum
3
and
Ralf Möller
2
Affiliations:
1
singularIT GmbH, 04109 Leipzig, Germany
;
2
German Research Center for Artificial Intelligence, Ratzeburger Allee 160, 23562 Lübeck, Germany
;
3
University Medical Center Schleswig–Holstein, Campus Lübeck, Ratzeburger Allee 160, 23538 Lübeck, Germany
Keyword(s):
Bed Occupancy Prediction, Emergency Department, MIMIC-IV, CatBoost Architecture.
Abstract:
This paper addresses the important issue of optimizing hospital bed management by integrating machine learning-based length of stay (LoS) predictions with bed occupancy forecasting. The study primarily utilizes the MIMIC-IV dataset to compare actual bed occupancy against predictions derived from estimated LoS. A novel approach is adopted to translate individual patient LoS predictions into bed occupancy forecasts for the entire hospital. Through various simulations, the paper evaluates the effects of different error margins and patterns in LoS predictions on bed occupancy forecasting accuracy. Key findings reveal that a more symmetric error distribution in LoS predictions significantly enhances the accuracy of bed occupancy forecasts compared to merely reducing the overall prediction error. The paper makes significant contributions to the field. The paper introduces a practical translation scheme from LoS prediction to bed occupancy, which is crucial for hospital administrators in re
source planning and management. Also the paper illuminates how various improvements in state-of-the-art LoS prediction models can directly impact the accuracy of bed occupancy forecasts, thereby setting clear objectives for future machine learning research.
(More)