Summary of the diabetes dataset.

Accurate, reliable, and efficient estimation of blood glucose dynamics from real-world data is challenging due to the time-varying nature, high uncertainty, and nonlinear interplay of complex processes. In this study, we propose and investigate a stochastic representation of a virtual population by fitting a hierarchical Bayesian model. In total, we use 500 24h-long sequences, 50 from each of the 10 patients with type 1 diabetes on multiple daily injection therapy. We model uncertainty on multiple levels, in physiology and in self-reported events, and take into account intra- and interday variability, and the effect of physical activity as well. The root-mean-square error between the glucose measurements and the mean of the posterior predictive distribution using the fitted low-rank multivariate normal guide is 12.44 mg/dL. We show that the posterior distributions can be used to simulate realistic intra-, and interday variability in terms of the investigated patient cohort.