Model uncertainty due to differing assumptions or unknown system mechanisms is often overlooked when applying mathematical models in biology and medicine. In diabetes diagnostics, mathematical models have long been used to make inferences about a patient’s metabolic health using available clinical data such as blood glucose measurements over time. These approaches often rely on a phenomenological model to approximate the physiological system, ignoring possible uncertainty in the model structure. However, there are usually several possible phenomenological models, each of which uses different formulations to represent the same biological processes. Given a family of phenomenological models, one typically chooses a single model based on a priori assumptions. In this work, we instead focus on leveraging the whole family of models to develop robust predictors in the face of uncertainty in the models describing the biological process. This talk outlines several approaches to average the predictions from all available models, including Bayesian model averaging and probability distribution fusion. As a test case, we chose the prediction of beta cell insulin regulation and associated diagnostic metrics from blood glucose measurements. Our results show that working with a family of models instead of a single model improves the certainty of modeling-based predictions, reduces biases associated with selecting one model, and explicitly accounts for model uncertainty. 

The video of this presentation is available here.