Diabetes Mellitus Glucose Prediction by Linear and Bayesian Ensemble Modeling

Diabetes Mellitus is a chronic disease of impaired blood glucose control due to degraded or absent bodily-specific insulin production, or utilization. To the affected, this in many cases implies relying on insulin injections and blood glucose measurements, in order to keep the blood glucose level within acceptable limits. Risks of developing short- and long-term complications, due to both too high and too low blood glucose concentrations are severalfold, and, generally, the glucose dynamics are not easy too fully comprehend for the affected individual—resulting in poor glucose control. To reduce the burden this implies to the patient and society, in terms of physiological and monetary costs, different technical solutions, based on closed or semi-closed loop blood glucose control, have been suggested. To this end, this thesis investigates simplified linear and merged models of glucose dynamics for the purpose of short-term prediction, developed within the EU FP7 DIAdvisor project. These models could, e.g., be used, in a decision support system, to alert the user of future low and high glucose levels, and, when implemented in a control framework, to suggest proactive actions. The simplified models were evaluated on 47 patient data records from the first DIAdvisor trial. Qualitatively physiological correct responses were imposed, and model-based prediction, up to two hours ahead, and specifically for low blood glucose detection, was evaluated. The glucose raising, and lowering effect of meals and insulin were estimated, together with the clinically relevant carbohydrate-to-insulin ratio. The model was further expanded to include the blood-to-interstitial lag, and tested for one patient data set. Finally, a novel algorithm for merging of multiple prediction models was developed and validated on both artificial data and 12 datasets from the second DIAdvisor trial.