Pharmaceutical stock forecasting in healthcare supply chain management: Comparative analysis of deep learning architectures

Ensuring uninterrupted treatment processes, constant medication accessibility, and preventing wastage due to expired overstock, thereby avoiding additional costs, require a well-planned supply chain. To enhance planning accuracy within the healthcare supply chain, recent approaches have increasingly harnessed not only traditional methods but also the power of deep learning architectures. Although previous studies have generated forecasts using past stock and demand data, these methods often fail to fully capture the dynamic nature and abrupt changes inherent to drug inventories. In this study, the objective is to improve pharmaceutical stock forecasting by correlating diagnosis counts with stock levels. A two-stage approach is employed: the first stage focuses on forecasting diagnosis counts, while the second stage develops models that utilize both diagnosis and stock data. Various deep learning models such as TCN, LSTM, GRU, hybrid GRU, and hybrid LSTM are comprehensively analyzed. Model performance is evaluated using error metrics such as MSE, MAE, RMSE, MAPE, and R². Notably, tests conducted on a limited dataset show that a simple GRU architecture with 6 neurons achieves one of the best outcomes in the stock forecasting stage, obtaining an MSE of 0.2071 and an R² of 0.8686. These results are promising for achieving more accurate and predictable stock planning in the healthcare supply chain, while also contributing significantly to the efficient utilization of resources and the reduction of unnecessary costs.