Experimental Evaluation of Machine Learning Algorithms for Demand Forecasting of Medical Supplies in Natural Calamity Relief Operations

Purpose – This study evaluates machine learning algorithms for forecasting medical supply demand in natural calamity relief operations. It investigates forecasting accuracy, robustness, interpretability, and performance. Method – The Knowledge Discovery in Databases (KDD) process was adopted as the methodological framework. Thirteen algorithms were tested: linear models (Linear Regression, Ridge, Lasso, ElasticNet), tree-based models (Decision Tree, Random Forest, Extra Trees), boosting models (Gradient Boosting, XGBoost, LightGBM, CatBoost), and additional approaches (K-Nearest Neighbors, Support Vector Regressor). Their performance was assessed using RMSE, MAE, and R² metrics. Results – CatBoost achieved the highest baseline R² (0.9962), with XGBoost, Extra Trees, LightGBM, and Random Forest also performing strongly (> 0.988). Regression models, KNN, and SVR showed weaker robustness. After hyperparameter tuning with a randomized search and 5-fold cross-validation, LightGBM emerged as the top performer (0.9941), narrowly surpassing CatBoost and Gradient Boosting, underscoring the advantage of optimized boosting ensembles. Conclusion – LightGBM, CatBoost, and Gradient Boosting demonstrated superior accuracy and robustness with hyperparameter optimization, further enhancing results. Recommendations – Disaster response agencies should adopt ensemble models, particularly LightGBM, CatBoost, and Gradient Boosting, within their decision-support systems, while applying hyperparameter tuning and exploring real-time data integration for future applications. Research Implications – Findings reinforce boosting-based ensembles as reliable tools for disaster demand forecasting. Enhancing academic understanding and improving logistics by efficiency, response times, and resource allocation in relief operations.