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

Area of Science:

• Healthcare Operations Research
• Data Science
• Public Health

Background:

• Blood donation volumes exhibit high volatility due to external events like pandemics and holidays.
• Traditional statistical models struggle to account for event-driven shocks in blood supply.
• Unstructured text data from news and social media contains valuable information for predicting donation fluctuations.

Purpose of the Study:

• To develop a semantics-enhanced forecasting framework for blood donation.
• To integrate structured donation data with unstructured text data for improved prediction.
• To enhance the resilience of blood supply chains against unpredictable demand.

Main Methods:

• Benchmarked topic-modeling methods, selecting BERTopic for its performance.
• Aggregated document-level topic distributions, compressed using Principal Component Analysis (PCA), and transformed into lagged predictors.
• Combined semantic features with structured data to train Gradient Boosting, XGBoost, LightGBM, CatBoost, Random Forest, GAM, and SARIMAX models.

Main Results:

• Semantic features significantly improved predictive accuracy, with Gradient Boosting achieving the best performance.
• A 28% reduction in root mean square error (RMSE) and a 0.33 increase in R-squared were observed.
• SHapley Additive exPlanations (SHAP) identified key predictors including campaign indicators, mobile donation sites, donor demographics, and epidemic/disaster-related topics.

Conclusions:

• The "predict-then-optimize" pipeline enhances blood supply resilience by incorporating semantic insights.
• The framework is generalizable to other healthcare or perishable supply chains facing event-driven volatility.
• Accurate forecasting of blood donation volumes is crucial for efficient inventory management and service level optimization.