NOF1-BCD - a framework for causal estimates with N-of-1 Bayesian Digital Twins

Single-person self-experimental data, so-called N-of-1 trials, are the gold standard for inference on individual treatment effects; however, they often face data scarcity, limiting inference on effect estimates. Also, they are not always feasible, and often, only observational data is available. Here, we propose NOF1-BCD, a framework for individual-level causal inference for treatment effects that provides regularization through priors in a Bayesian Digital Twin. First, we leverage large language models to efficiently synthesize external data sources---such as population-level studies and domain knowledge---into informative priors. Second, these priors are used in a Bayesian Digital Twin applied to individual-level data. Third, the Digital Twin serves as a prediction model to estimate the Average Period Treatment Effect, a counterfactual contrast for each individual. We characterize the framework and demonstrate it using smartwatch-collected physical activity and sleep recovery data, assessing the potential of LLM-elicited priors to improve estimation efficiency without compromising causal inference. Applications of the framework can extend beyond healthcare.