Definition of model parameters and functions.

Accurate vaccine demand forecasting is crucial for minimizing wastage and ensuring efficient immunization programs. In this study, we introduce an Adaptive Large Language Model for Vaccine Prediction (ALLMVP) that integrates large language model (LLM) architectures with an adaptive value correction mechanism. Using vaccination record data from Xuhui District, Shanghai, China (2014–2022), we conducted a comparative analysis of ALLMVP against seven other models, including standard machine learning methods (logistic regression, random forest, Long Short-Term Memory) and their enhanced versions with the same adaptive value correction mechanism. Our findings indicate that traditional models encountered significant challenges in attaining high predictive accuracy, while frameworks based on LLMs markedly enhanced their forecasting capabilities. Notably, ALLMVP achieved an acceptance ratio of 71.43%, which is considerably superior to that of the other models under consideration. Yearly and cumulative evaluations across seven vaccines from the National Immunization Program demonstrated that ALLMVP consistently delivered more precise estimates, aligning closely with actual vaccine demand even under challenging conditions, such as the post-COVID era. These results highlight the potential of adaptive LLM-driven forecasting tools to fullfill stringent prediction accuracy standards set by governments and to aid in data-informed vaccination strategizing. The AI infrastructure underpinning ALLMVP holds the promise of being generalized and deployed across a range of forecasting applications and at a significantly larger scale.