Bayesian mediation multi-state model elucidates the impact of cancer immunity on multiple oncology endpoints after chemoradiotherapy for esophageal cancer

Abstract Unraveling the complex relationships among interventions, intermediate health states, and survival outcomes is crucial to translating advances in clinical research into optimal cancer therapies. We propose a novel Bayesian multi-state mediation model to evaluate the direct and indirect effects of radiation modalities on overall survival (OS). Our model extends traditional mediation analysis by incorporating a multistate framework that accounts for intermediate outcomes, such as tumor response and progression or recurrence, providing a more detailed representation of treatment impact over time. Bayesian estimation and variable selection enhance model reliability and address uncertainty in parameters and structure, providing robust estimates of direct and mediated effects. The method characterizes dynamic relationships between endpoints and intermediate health states which may be time-varying mediators. Furthermore, our study highlights the clinical importance of considering intermediate outcomes in the evaluation of treatment efficacy and patient survival. The model is applied to a propensity-matched retrospective cohort of 510 esophageal cancer patients treated with proton beam therapy (PBT) or intensity-modulated radiation therapy (IMRT). Results reveal the survival benefit of PBT is significantly attributable to reduced risk of severe radiation-induced lymphopenia compared to IMRT (22.0% vs. 42.7%, respectively; $$P<0.001$$ ). At 72 months post-treatment, the mediated RR is 0.948 (95% credible interval [CrI]: 0.918–0.979) in favor of PBT. Future innovations in radiation therapy planning are needed to mitigate severe immunosuppression.