We present a novel framework for hierarchical Bayesian modeling of adverse events in multicenter clinical trials, leveraging large-scale language models (LLMs). Unlike data augmentation approaches that generate synthetic data points, this study derives parametric prior distributions directly from the model. Using pre-trained LLMs, we systematically derive informative prior distributions for the hyperparameters of the hierarchical Bayesian model, directly incorporating external clinical expertise into Bayesian safety modeling. Comprehensive temperature sensitivity analyses and rigorous cross-validation on real-world clinical trial data demonstrate that LLM-derived prior distributions consistently improve predictive performance compared to existing meta-analytic approaches. This methodology paves the way for more efficient and expert-informed clinical trial design, significantly reducing the number of patients required to achieve robust safety assessments and potentially transforming drug safety monitoring and regulatory decision-making.
