A Three-phase Hybrid Times Series Modeling Framework for Improved Hospital Inventory Demand Forecast

Document Type: Research Paper


1 Department of Information Technology, Faculty of Engineering, Payam Noor University, Tehran, Iran

2 Department of Industrial Engineering, Iran University of Science and Technology, Tehran, Iran

3 Department of Industrial Engineering, Faculty of Engineering, Tarbiat Modares University, Tehran, Iran


Background and Objectives: Efficient cost management in hospitals’ pharmaceutical inventories have the potential to remarkably contribute to optimization of overall hospital expenditures. To this end, reliable forecasting models for accurate prediction of future pharmaceutical demands are instrumental. While the linear methods are frequently used for forecasting purposes chiefly due to their simplicity, they have serious deficiencies in capturing nonlinearities in real-world problems. On the other hand, real world time series data are rarely pure linear or nonlinear, calling for development of forecasting models accounting for both these features of the data. To help meeting this need in the health/healthcare domain, this study undertook development of a hybrid framework consisting of a linear and a nonlinear component to improve forecasting of operating rooms’ pharmaceutical demand. Methods: A hybrid modeling framework combining AutorRgressive Integrated Moving average (ARIMA) as the linear component, and Artificial Neural Network (ANN) as the nonlinear component was developed. The method encompasses three phases: 1) Fitting a linear ARIMA model to the targeted time series, (2) Building an ANN model based on the residuals of the ARIMA model, and (3) Build the hybrid model by combining ARIMA and ANN models for the final forecast. Using the pharmaceutical inventory database of the Iranian Mohem Hospital for fitting AMIRA model and training ANN model, the forecast performance of all three models was compared by calculating the corresponding mean squared error and mean absolute error values, and by superimposing the time series patterns of the operating rooms’ drug demand independently predicted by each model to the corresponding observed pattern. Findings: Both quantitative and intuitive comparisons demonstrated that our hybrid ARIMA-ANN framework outperforms forecasting capability of either ARIMA or ANN models. In particular, the hybrid model showed remarkably superior capability in capturing the nonlinear behavior of the operating rooms’ pharmaceutical demand time series. Conclusions: Our proposed framework sets a ground for developing mathematical and computational forecasting models with ever higher predictive accuracy and supports the promotion of using such forecasting models in practical cost optimization in health facilities.