Characterizing Parameter Uncertainty in Biomedical Systems for Improved Estimation, Prediction and Control

Recent years have witnessed the vigorous research and development of Artificial Pancreas system (AP), which attempts to “close the loop” through communications with third party devices such as continuous glucose monitoring and insulin pump aimed at dispensing the patients with diabetes from the responsibility of insulin dosing. Physiological modeling is a pragmatic methodology for explaining observed dynamic effects, interpreting the experimental data and predicting system responses under certain stimuli. Given the complexity of the insulin-glucose dynamic system, lumping the parameter set is obligatory under most circumstances. The ubiquitous inter-individual and intra-individual variability across the population requires us to consider the necessary remedial action such as characterization of the parameter uncertainties.
Insulin sensitivity (SI), is one of the critical parameters that govern the insulin-glucose dynamics. We propose and validate a Kalman Filtering based technique that is capable of tracking SI in real time based on commonly available data measurements. We then apply the developed technique to estimate SI during the menstrual cycle. The results substantiate the hypothesis that a subset of premenopausal women with T1DM will experience a decrease in insulin sensitivity during the second half of the menstrual cycle (luteal phase). With the knowledge of SI, we optimize the predictive power of a dosing algorithm; short-term (up to 45 minutes) forecasting ability of BG is studied by exploring different structural designs (full model, feed-forward, with and without SI tracking). Finally, a model based decision support system is derived for insulin dosing; long-term (4 hours) forecasting characteristics of BG influenced by individualized parameters are studied both in-silico and in-vivo.