Forecasting patient census at a university hospital using the Prophet model

Abstract

Abstract The widespread emergence of machine learning-based solutions provides most fields with a range of available tools. Utilizing these tools requires resources unavailable to many institutions. Off-the-shelf methods could fill this gap, as they usually require little implementational effort, and are easily customizable to fit a wide range of tasks. To assess how easily off-the-shelf methods can be adapted to clinical tasks so that clinicians would be provided actionable predictions, we adopted the time series forecasting model Prophet to forecast patient census time series 3 days from the day of prediction. The underlying time series used for model training consisted of day-to-day census data collected in 2021 and 2022. Performance of the model had been evaluated daily against a baseline and the manual predictions of a business subject matter expert. The Prophet model outperformed the other methods and provided accurate predictions for several time series; it predicted adult intensive care unit census, overall adult census, and emergency department census with 2.1%, 3.2%, and 7.6% mean absolute percentage error, respectively. To acquire these predictions, necessary steps were installing Python, the library containing Prophet, and writing a script around it to train it on the time series of interest and generate actual predictions. The model has been operationalized in a large academic hospital. We conclude that medical facilities which accumulate time series data (patient census, resources, etc.) can benefit from advances in machine learning, or more specifically time series forecasting by adopting off-the-shelf methods to generate actionable predictions to enhance patient care.

Ähnliche Arbeiten