Search papers, labs, and topics across Lattice.
This paper introduces a two-step data-driven framework for vaccine demand forecasting and inventory management in hospital travel clinics. The framework uses XGBoost and LightGBM for multi-day demand forecasting, with bootstrapping to quantify uncertainty, and then employs Monte Carlo simulation with a custom heuristic to optimize inventory policies. A case study demonstrates that the framework, particularly with dynamic reorder points, effectively balances stock-out risk and inventory efficiency, outperforming benchmark forecasting algorithms like CatBoost, LSTM, and Prophet.
Stop guessing and start planning: a new data-driven framework slashes vaccine stockouts while optimizing inventory, even with unpredictable demand and strict shelf-life constraints.
Forecasting vaccine demand and determining inventory policies are critical challenges in healthcare supply chains, where uncertainty poses significant operational risks. This study proposes a two-step data-driven framework to support vaccine planning under uncertainty. The first step leverages machine learning models鈥擷GBoost and LightGBM鈥攆or daily demand forecasting using a recursive multi-day strategy, with model deviations generated via bootstrapping to characterize uncertainty. Forecast accuracy is evaluated using a sliding-window Mean Cumulative Absolute Percentage Error to capture cumulative deviations relevant to operational planning. The second step employs a stochastic Monte Carlo simulation and a custom performance-based heuristic to determine proper policy parameters. A key feature is the implementation of dynamic reorder points and order quantities that adapt to forecasted demand and volatility to ensure responsiveness. By incorporating data-driven forecast distributions, the simulation evaluates trade-offs between stock-out risk and inventory efficiency using Value-at-Risk metrics. A case study examining vaccines in a hospital travel clinic confirms the framework鈥檚 real-world applicability and the effectiveness of this hybrid approach. Results reveal that XGBoost performs better for seasonal or volatile demand, while LightGBM excels with smoother profiles. Notably, both algorithms outperform benchmark algorithms including CatBoost, LSTM, and Prophet. Furthermore, the proposed heuristic identifies effective policy parameters for each vaccine within a computationally efficient timeframe. Inventory results show that the proposed method maintains inventory days within hospital targets to maintain vaccine potency while simultaneously minimizing the risk of stockout. This is particularly advantageous for travel clinics that manage diverse vaccine portfolios with unpredictable demand and strict shelf-life constraints.