Integrating topic-distribution features into forecasting and inventory optimization: Evidence from Taiwan's blood supply system

BACKGROUND: Blood donation volumes fluctuate sharply with pandemics, disasters, holidays, and promotional campaigns, creating risks of shortages and waste. Traditional statistical models often fail to capture such event-driven shocks. This study proposes a semantics-enhanced forecasting framework that integrates structured donation data with unstructured news and social-media text. METHODS: We benchmark multiple topic-modeling methods-Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), Correlated Topic Model (CTM), and BERTopic-and select BERTopic for its superior coherence-diversity tradeoff. Document-level topic distributions are aggregated into daily vectors, compressed using Principal Component Analysis (PCA), and transformed into lagged predictors. These semantic features are combined with structured donation data and used to train several forecasting models, including Gradient Boosting, XGBoost, LightGBM, CatBoost, Random Forest, generalized additive models (GAM), and Seasonal ARIMA with exogenous regressors (SARIMAX). RESULTS: ). Robustness was confirmed using Diebold-Mariano tests, paired t-tests, Wilcoxon tests, and sign tests. SHapley Additive exPlanations (SHAP) analysis revealed stable influential predictors, including campaign indicators, mobile donation sites, donor age groups, and semantic topics related to epidemics, disasters, and blood-shortage appeals. Embedding forecasts into a perishable-inventory Mixed-Integer Linear Programming (MILP) model-using predicted daily capacities as constraints-achieved a 96.9 % service level, compared with 98.9 % under an Oracle benchmark with perfect foresight. CONCLUSIONS: The proposed "predict-then-optimize" pipeline enhances blood-supply resilience and can generalize to other healthcare or perishable-supply contexts characterized by event-driven volatility.