COMPARATIVE STUDY OF ARIMA AND MACHINE LEARNING MODELS FOR BIOPROSTHETIC TISSUE HEART VALVE DEMAND FORECASTING AND ITS APPLICATION IN HIGH-VALUE MEDICAL SUPPLY INVENTORY MANAGEMENT

Abstract

This research aimed to (1) compare the forecasting performance of ARIMA and five Machine Learning (ML) models—Ridge Regression, Lasso Regression, Random Forest, Gradient Boosting, and XG-Boost together with two intermittent-demand benchmarks (Croston's Method and Syntetos-Boylan Approximation) for bioprosthetic tissue heart valve demand; (2) select the Best Model for nine valve items; and (3) develop an optimal inventory system and quantify financial savings at Queen Sirikit Heart Center of the Northeast. Monthly data for nine items (3 Porcine, 6 Bovine) over 36 months (1 January 2023-31 December 2025), totalling 324 data points, were purposively selected. Analyses employed ADF/KPSS stationarity tests, five-fold Walk-Forward Cross-Validation, Diebold-Mariano tests (baseline and pairwise), and MAE, RMSE, MASE metrics. Results indicated that all ML models significantly outperformed ARIMA (Pairwise DM Test, p < .05). Gradient Boosting achieved the lowest average MAE (0.354), whereas ARIMA recorded the highest (0.921). Per-SKU selection identified Ridge Regression as optimal for six items (66.67%), Gradient Boosting for two (22.22%), and XG-Boost for one (11.11%). At a 99% Service Level, optimal stock ranged from 36 to 59 units, reducing average inventory by 66.78% (143 to 47.5 units), releasing approximately 9.81 million baht in capital and saving an estimated 971,000 baht per year in holding costs. This study presents the first empirical Best Model Selection framework in Thailand for high-value medical supply forecasting under intermittent demand.

Keywords: Demand Forecasting, Bioprosthetic Tissue Heart Valve, Machine Learning, ARIMA, Inventory Management