Integrating Digital Twins into Randomized Controlled Trials: Recent Advances in AI and Statistical Methods

Randomized controlled trials (RCTs) remain the gold standard for causal inference in clinical and biomedical research, but they often face limited external validity, inadequate statistical power, unrepresentative enrollment, and ethical constraints. Digital twins (DTs), including emerging intervention-aware patient world models, offer new ways to represent longitudinal patient states, forecast treatment responses, and construct virtual trial populations from heterogeneous biomedical data. This paper reviews recent progress in health-domain DTs and synthesizes how DT-derived information may support RCT design and analysis. We focus on three statistical uses: calibration of DT-generated outcomes, selective borrowing in covariate regions where DT and trial distributions align, and DT-based prognostic covariates for improving precision. We also examine how multimodal foundation models expand the feasibility of patient-level DTs while leaving unresolved challenges in missing-data identification, temporal consistency, transportability, uncertainty quantification, and validation. Rather than proposing a new DT architecture or causal estimator, we provide a technical synthesis of the assumptions and safeguards required for DTs to complement randomized evidence. Under appropriate calibration and validation, these approaches may improve trial efficiency and generalizability without treating simulated outcomes as substitutes for observed randomized evidence.