Diabetes is a disease characterized by deficiency of insulin secretion or deficiency of the body to respond to insulin normally (insulin resistance), or both. This imbalance of insulin, a hormone produced by the beta-cells of the islets of Langerhans in the pancreas, impairs the metabolism of glucose and other nutrients in the bloodstream. Patients with diabetes, especially with type 1 diabetes who have no endogenous insulin, control their blood glucose concentrations by administering exogenous insulin and balancing their diet. With the current therapy, insulin is delivered through multiple daily injections (3 to 5 injections usually taken before meals) or a manually controlled subcutaneous insulin pump. Patients adjust their bolus (meal-related) insulin dose based on their pre-meal blood glucose test, estimated carbohydrate content of the meal and planned post-meal activity levels. Success rate at achieving normoglycemia with current insulin therapies has been very low, and patients may experience prolonged hyperglycemic or hypoglycemic episodes. Models predicting a subject's future glucose concentrations can be used for preventing such conditions by providing early alarms.
Reliable glucose prediction models have the potential to significantly improve the management of diabetes. Patients will benefit from predicted glucose concentrations for their insulin dose decisions. More importantly, future glucose levels can be used to provide an early alarm that predicts hypoglycemia or hyperglycemia, allowing the necessary time for the patient to prevent the incident. In this research, a subject-specific glucose prediction model is developed using measurements from a glucose sensor and metabolic, physical activity and lifestyle information from a multi-sensor armband.
The model developed is subject-specific and dynamically captures the changes in the glucose metabolism of the subject. Time-series approach is used for development of a linear low-order model. Adaptive system identification is proposed to estimate model parameters at each sampling step. It consists of an on-line parameter identification using the weighted recursive least squares method and a change detection strategy that monitors variation in the model parameters. A variable forgetting factor is implemented through the change detection strategy which ensures good data tracking not only during slow frequency changes in glucose dynamics but also during drastic and sudden variations. The modeling algorithm estimates future glucose levels using subject's glucose and physiological measurements, and does not require any prior experimental data, tuning for each subject, or disturbance information. Therefore, it can easily be implemented for any subject using a CGM sensor and a multi-sensor body monitor.
The prediction algorithm is evaluated on glucose concentration data and physiological signals collected from subjects with diabetes under ambulatory conditions. A continuous glucose monitor is used to gather the subject's glucose concentration every five minutes. Additionally, metabolic, physical activity and lifestyle information is collected by a multi-sensor body monitor (armband). The armband provides total of seven signals to describe the subject's activity and emotional conditions: energy expenditure, average longitudinal acceleration, transverse acceleration peaks, transverse acceleration mean of absolute difference, near-body temperature, heat flux and galvanic skin response.
Univariate models developed from subject's continuous glucose measurements are compared to multivariate models that are enhanced with continuous metabolic, physical activity and lifestyle information from the armband. Results show that physiological signals supplement the continuous glucose information and enhance the prediction accuracy of the glucose predicting models. Prediction performance analysis demonstrate that error in predictions are significantly reduced with additional measurements from the armband (multivariate model), when compared to predictions done solely on glucose measurements (univariate model). The multivariate modeling algorithm is further evaluated for predicting hypoglycemia in advance, with early alarms. The proposed multivariate algorithm provides 30 min ahead predicted glucose values that closely follow the sensor data, which improves the alarm performance (sensitivity).