Demand forecasting for platelet usage: From univariate time series to multivariable models.

Area of Science:

• Hematology
• Data Science
• Health Services Management

Background:

• Platelet products are costly with limited shelf lives, necessitating efficient demand and supply management.
• Variable platelet usage rates present significant challenges for inventory control and distribution.

Purpose of the Study:

• To develop and evaluate an efficient forecasting model for platelet demand at Canadian Blood Services (CBS).
• To compare the performance of statistical time series models against data-driven regression and machine learning techniques for platelet demand forecasting.

Main Methods:

• Utilized five forecasting methods: ARIMA, Prophet, Lasso regression, Random Forest, and LSTM networks.
• Employed a rolling window method for model evaluation.
• Analyzed a comprehensive clinical dataset of daily platelet transfusions (2010-2018) including product specifications, recipient characteristics, and laboratory results.

Main Results:

• Multivariable approaches generally demonstrated the highest forecasting accuracy.
• Simpler time series models, like ARIMA, proved sufficient when ample historical data was available.
• Identified key clinical predictors for improving multivariable forecasting models.

Conclusions:

• Advanced forecasting models, particularly multivariable ones, can significantly improve platelet demand prediction.
• The choice of forecasting method should consider data availability, with simpler models being effective for long historical datasets.
• This research provides a framework for optimizing platelet inventory management through data-driven insights.