Towards Defence In Depth In Diabetes Glucose Self-Management

Sammanfattning: Diabetes is a disease characterized by insufficient capacity to regulate the blood glucose level. In insulin-dependent diabetes, multiple daily injections of insulin have to be administered. In-between scheduled visits to the care provider, the patient has to manage the glucose control independently. Insulin dosing is a non-trivial task and many patients find it difficult. This is reflected in the health statistics, that indicate that a majority of patients with diabetes have poor metabolic control with associate risks of several short and long term complications. In this thesis, building blocks of a defence-in-depth approach to glucose self-management in insulin-dependent diabetes are investigated. Defence-in-depth is a concept where technical and administrative systems work in cohort to divert potentially dangerous conditions and events. In the context of insulin-dependent diabetes this amounts to avoiding low (hypoglycemia) and high (hyperglycemia) glucose values. Data from the European DIADvisor project and from a local trial conducted with patients from Skåne University Hospital were used in the thesis. A basis for improved glucose control is understanding and knowledge of the glucose-lowering effect of insulin, the insulin action, and the corresponding glucose-elevating effect produced by meal intake. Individualized models of these impacts, and methods to improve the predictive capacity of these models, were developed. Interesting properties, such as, time-variability and nonlinear effects, were found. The models allow for the glucose level to be predicted and different meal and bolus scenarios to be simulated. Using the models, the possibility to foresee and prevent nocturnal hypoglycemia was validated with good performance in a retrospective analysis on the collected data. Recent advances in sensor technology have allowed for commercial systems where the glucose level is measured with a high sampling rate in the interstitial fluid. However, a known deficiency with this approach is the measurement lag introduced by equilibrium dynamics between the blood and interstitial compartments. A Kalman filter based approach to resolving this issue was developed and successfully validated in a case study. Diabetes glucose dynamics is known to comprise both short and more long term time-variability. Merging different diversified models may prove to be a successful approach, as a means to improve performance and robustness under such conditions. A novel merging algorithm based in a Bayesian setting was developed. The suggested method admits for soft switching and interpolation between the different models based on an evaluation of the different predictors' recent performance, using a sliding data window, and by looking for data features identified to be correlated to switching. Different aspects of the merging approach were investigated, using a simulated dataset, and the concept was thereafter successfully validated, showing improved robustness to the prediction performance in comparison to relying on the individual prediction models. Meal impact models were estimated for 56 different meal types, and a clustering analysis showed that a majority of these models could be represented by three base models. Cross-validation confirmed good predictive capacity. The insulin action and meal impact models were further used to assess whether clinical recommendations on postprandial glucose levels, issued by international patient and professional organizations, are realistic and achievable. An important finding was that the postprandial excursion of meals with rapid postprandial response may be impossible to restrain within the recommended boundaries for even moderate meal sizes. This difficulty is exaggerated for persons with slower than normal insulin action. The above methods and models could contribute to improving already available technology in diabetes self-management such as, e.g., bolus dose guides in insulin pumps, warning systems in continuous glucose monitoring systems or in interpretation and implementation of postprandial recommendations. %A Bayesian method to allow several specialized prediction models to work in cohort was also developed. Validation on both simulated and real-world data confirmed that the prediction robustness increased. Finally, %Among these, the insulin model reconfirmed a previous result that the insulin action is heterogeneous across the glucose range, with elevated magnitude at low glucose values and reduced at high glucose values.