Towards a Safe Artificial Pancreas: Meal Detection and the Intraperitoneal Route
Abstract
Patients with diabetes mellitus type 1 have insufficient insulin production. If untreated, the insulin deficiency causes diabetic ketoacidosis and death. Long-term elevated blood glucose levels lead to a number of complications. On the other hand, insulin must be carefully dosed to avoid life-threatening hypoglycemia. Food, exercise, as well as daily and seasonal hormonal variations act as disturbances on the glucose homoeostasis and render the treatment a challenge.
The aim of this thesis was to develop safety functions that may enable fully automated closed-loop glucose control systems in diabetes mellitus type 1. Safety is defined as the absence of risks. Risks in the context of glucose regulation arise from hyperglycemia and hypoglycemia. A risk analysis of a conceptual artificial pancreas by means of failure modes and effects analysis elucidated that a large variety of system faults may occur. The risk analysis revealed among others that: (1) the glucose homoeostasis is most frequently disturbed by meals, (2) the subcutaneous tissue as the typical route for glucose sensing and insulin infusion is associated with slow dynamics and several other challenges.
The achievable postprandial regulation by insulin bolusing upon meal detection was investigated in simulations. Published methods were used that trigger a meal detection when the glucose concentration or its first two derivatives exceed thresholds. Even with the surprisingly fast dynamics that had been reported for glucose sensing in the peritoneal cavity, the postprandial regulation was insufficient. Furthermore, the simulation study recommended to individualise thresholds which is hard to realise in outpatient settings.
A non-individualised meal detection based on CGM data was proposed. The method exploits the patterns in horizons of CGM measurements by classifying them either directly or after processing by moving horizon estimation. The classification by linear discriminant analysis was tested in simulations and validated on a clinical data set. Higher sensitivity and fewer false alarms were demonstrated compared to the threshold-based methods.
Abdominal sounds were studied as a redundant or alternative information source of food intake. The pilot study investigated whether abdominal sounds change fast enough for an early meal detection in the context of an artificial pancreas. The study showed that the distribution of power in frequency intervals increases within a few minutes after meal start. This enables meal detection by pattern recognition in the early postprandial period. Moreover, a data-driven filtering method was proposed that segments bowel sounds from background noise.
The challenges associated with subcutaneous glucose sensing and insulin infusion hamper the development of an artificial pancreas. The peritoneal cavity was explored in pre-clinical trials as alternative route for glucose sensing. Glucose measurements were fitted to a differential model to quantify the dynamics by means of time delay and time constant. Intraperitoneal glucose sensing was less rapid than previously reported but still faster than subcutaneous glucose sensing.
This thesis has contributed towards the goal of a safe artificial pancreas by assessing and developing methods for meal detection and insights into intraperitoneal glucose sensing.