Engineering data-driven forecasting models for intelligent blood inventory decision-support in healthcare systems

Résumé fourni par la source

Abstract Background Forecasting an adequate supply of blood components by anticipating donation volumes ensure efficient control of blood inventories in modern healthcare systems. Analyzing donor demographics and long-term donation patterns allow blood centres support informed data-driven planning, especially in the presence of seasonal fluctuations and post-pandemic changes. The aim of our study is to mitigate supply uncertainty of blood centres by constructing a forecasting framework that helps in the integration of theoretical modeling approaches with practical clinical demands. Methods Daily donation volumes for three years (2021–2023) were collected directly from the blood bank repository of a leading tertiary healthcare institution in Coastal Karnataka, India. Seasonality assessment and formal statistical testing were performed on the donation units evaluated on the basis of daily, weekly, and monthly donation levels. A structured comparison of donation forecasting across Seasonal AutoRegressive Integrated Moving Average (SARIMA), Facebook Prophet, eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and hybrid ensemble approaches was performed employing a rolling-origin backtesting framework in order to maintain methodological rigor and reproducibility. Results Our study showed donors’ demographics with 93.41% male and 6.59% female donor population, most aged between 18 and 45 years. Significant weekly and monthly variation among blood groups A and O, and significant monthly variation for blood groups AB and B was revealed through statistical testing. Multi-resolution analysis demonstrated that the Hybrid A (SARIMA+Prophet+LSTM) ensemble achieved the highest forecasting accuracy at weekly scale (s = 52), while standalone models such as SARIMA, XGBoost and LSTM performed competitively at daily (s = 7) and monthly (s = 12) horizons. In general, the hybrid architectures reduced forecasting error, with the lowest Mean Absolute Error values of 11.46, 10.52, 3.00, and 20.02 for blood groups A, B, AB, and O, respectively. Conclusion The need for comprehensive forecasting frameworks compared to standalone models addresses the issues of stochastic variability inherent in blood donation. This study establishes the importance of hybrid methodologies in proactive decision-making by integrating linear seasonal models with nonlinear learning techniques via inverse-variance weighting and evaluating them through rolling-origin validation on historical operational data. The proposed hybrid framework supports Sustainable Supply Chain Management by enhancing supply-side visibility, providing a reliable predictive foundation for resource planning and responsive, donor-centred services.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.