A Bayesian adaptive digital twin design for avatar-driven cancer clinical trials

The premise of personalized medicine is that patients exhibit significant heterogeneity and respond differently to treatments. The best treatment for one patient may not be suitable for another. In this paper, we propose a Bayesian adaptive design for an avatar-driven clinical trial, using mouse or laboratory animal grafts to create personalized tumor models that guide individualized patient treatment while accounting for translational discrepancy between avatars and patients. We jointly model the tumor growth of avatars and the clinical outcomes of patients, including toxicity and progression-free survival, to capture the interplay between avatars and their source patients. This joint model enables us to leverage both avatar treatment responses and patient data to predict the treatment utility for each patient and inform the selection of personalized optimal treatments, thereby serving as a digital twin of the patient. To address the fact that avatars may only serve as valid surrogates for treatment effects in a subset of patients, we incorporate a latent-class model that accounts for this translational discrepancy. Through simulation studies, we demonstrate that the proposed design outperforms conventional clinical trials lacking avatar information. Under this design, patients are significantly more likely to receive personalized optimal treatments, leading to improved treatment outcomes.