Background and Aims: Individual responses to lifestyle factors, including physical activity, carbohydrate intake, and sleep, significantly affect 2 h glucose levels in people with type 2 diabetes (T2D). The substantial variation between individuals highlights the need for personal predictive tools. This study aimed to develop a mathematical model that quantifies the relationship between these key lifestyle factors and glucose levels and to evaluate the performance of different statistical modeling methodologies to achieve accurate, personal prediction. Methods: Data encompassing lifestyle factors and continuous glucose monitoring from 38 T2D participants were used. We initially employed a frequentist multilevel regression model to derive population estimates and individual random effects. These results were subsequently used to establish informed priors for a personal Bayesian model. Model performance, measured by the root mean squared error (RMSE) and mean difference, was compared against the population frequentist model and a personal Bayesian model initialized with weakly informative priors. Results: The personal Bayesian model utilizing informed priors exhibited better predictive accuracy. Specifically, the RMSE was significantly improved when compared against both the initial frequentist model and the personal model using weakly informative priors. Furthermore, substantial RMSE improvements were observed in several individuals, validating the efficacy of integrating population-level data to personalize subsequent modeling efforts. Conclusions: The use of multilevel regression estimates to inform personal Bayesian models enhances predictive performance for 2-h glucose levels. This methodological approach provides a robust framework to generate individual models, potentially enabling more targeted clinical management of T2D in the future.
Snel et al. (Thu,) studied this question.