Hybrid LSTM-DGWO Framework for Forecasting Insulin Aspart Demand: A Case Study of NovoRapid in the Pharmaceutical Supply Chain

Abstract

Background: Accurate demand forecasting is critical for operational resilience and efficient resource allocation in pharmaceutical supply chains, particularly for life-critical medications such as Insulin Aspart (NovoRapid). However, the demand for these products is often characterized by nonlinearity and structural complexity, which frequently limits the predictive accuracy of conventional statistical methods. Objectives: This study aims to improve demand forecasting accuracy by developing and evaluating a hybrid long short-term memory (LSTM)-Dynamic Grey Wolf Optimizer (DGWO) framework. The objective is to leverage deep learning for temporal sequence modeling while employing metaheuristic optimization to systematically refine model hyperparameters. Methods: The proposed framework integrates an LSTM network to capture temporal dependencies and nonlinear demand patterns with a DGWO for hyperparameter tuning. The hybrid model’s forecasting performance was empirically evaluated against benchmark approaches, specifically an autoregressive integrated moving average (ARIMA) (1, 1, 1) model and a naive baseline, using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and the coefficient of determination (R2) as accuracy indicators. Results: The empirical results show that the proposed hybrid framework outperforms the benchmark models. The final model achieved an RMSE of 131,400, an MAE of 81,112, a MAPE of 10.59%, and an R2 of 0.6823. These findings indicate that the hybrid structure, which combines deep sequence learning with intelligent hyperparameter optimization, provides a more robust representation of demand dynamics than do traditional forecasting methods. Conclusions: Integrating deep learning with metaheuristic optimization provides a robust methodological foundation for pharmaceutical demand forecasting. The LSTM-DGWO framework serves as a practical decision-support tool for supply chain stakeholders, enabling improved inventory planning, optimized safety stock determination, and more accurate supply coordination. Although the model shows substantial promise, its generalizability is constrained by the 32-month dataset and the single-product focus. Future research should scale this framework across broader product portfolios and incorporate exogenous market variables to further strengthen predictive robustness.