Abstract:
Diabetes management consists of two major tasks: forecasting the blood glucose trend and taking a therapeutic decision. Due to the advancement of sensor technologies, it becomes easy to obtain continuous glucose monitoring (CGM) and physical activity data along with logging of diet and injected insulin information. These data of the diabetic patient are being leveraged by applying machine learning (ML) strategies to obtain future trajectories of glucose level which contains no explanation.
This thesis is aimed to produce physiological explanation from ML-based forecasting of glucose concentration by operation research (OR). For producing forecasting, ML-based model is trained using CGM profile with diet and activity information of a type-2 diabetic patient. Due to unavailability in the literature, a constraint-based comprehensive glucose dynamics model integrated with other physiological models of external stimuli is also aimed to build for OR. An integrated physiological model consisting of glucose regulation and models of external stimuli is considered as a composition of several compartments separately connected with a common compartment named ‘plasma’. Plasma is the only accessible compartment and contains the state variables. Plasma variables are the integrated result of the net change in rates of metabolic processes and basal rates are influenced between two saturation constraints for an operating range of each variable. The influence of a plasma variable on a metabolic rate is represented using a form of the hyperbolic tangent function. Validation is done by fitting the model with clinical experiments and CGM data of a free-living environment. A feed-forward neural network (FFNN) being trained on CGM data along with diet and activity log is used to produce forecasting. The proposed constraint-based glucose regulation model of this thesis is optimized on the forecasting of FFNN with sequential quadratic programming using preestimated personalized constraints. The proposed integrated physiological model generates an average correlation coefficient of 0.84±0.12 on all simulated responses with the target in the fitting experiments. Besides this, the model can produce a spectrum of metabolic effects of plasma variables for showing more insight into glucose metabolism. Both OR response and ML forecasting are compared with real glucose profiles. Though increased RMSE is obtained for OR response in comparison to ML forecasting, an acceptable accuracy is found in Clarke Error Grid Analysis. ML-based forecasting of glucose profile is transformed into optimized glucose trend with physiological interpretation. The interpretation is visualized in a metabolic spectrum derived from a constraint-based comprehensive glucose regulation model. The adopted hybrid approach is capable of encapsulating both generalization of ML and the explanation of the physiological approach.
