Adaptive and Optimal Control based Artificial Pancreas for Type-1 Diabetes Mellitus Patients

Abstract

This research focuses on the development of adaptive and optimal control algorithms and their implementation to develop Artificial Pancreas (AP) systems for glucose regulation in Type-1 Diabetes Mellitus (T1DM) patients. Artificial Pancreas systems have two modes of operation: (i) Bolus mode, the insulin required to account for glucose rise due to food intake, and (ii) basal mode, the insulin required to regulate glucose levels throughout the day. In order to develop an AP system, a robust algorithm is required for both basal and bolus control to ensure glucose levels are within a safe range at all times.

The first portion of the thesis focuses on the development of an autonomous bolus control algorithm. A minimal mathematical model that captures the glucose-insulin interaction in T1DM patients during the meal cycle was selected after extensive literature studies. Next, the Mixed Meal Tolerance Tests (MMTT) were conducted on Indian T1DM patients in order to collect time-tagged blood samples from patients. The parameters were identified for T1DM patients through the blood glucose and blood insulin concentrations that were estimated from the blood samples. The parameters of the Indian patients were compared with the Caucasian population, which showed that Indian T1DM patients have lower insulin sensitivity. An implication of this variation is that Indian patients would require more insulin to lower glucose levels when compared to the Caucasian population. Therefore, this served as sufficient motivation to develop an Indian population-specific AP system. A model predictive control (MPC) algorithm is proposed to regulate glucose levels within normal ranges during the meal cycle. The proposed MPC strategy exploits the minimal models and is customized to individual patients by fitting the model to a patients blood glucose and blood insulin data from the MMTT. In order to account for day-to-day intra-patient variability, the MPC scheme is augmented with neuro-adaptive learning in the glucose state dynamics. However, the adaptive learning scheme cannot be used to estimate the parameter uncertainties in the other state dynamic equations. An unscented transform-based “Least Risk” MPC is proposed to handle the uncertainties in the parameters of the other state dynamics. A model-based algorithm such as MPC requires state-feedback information, for which an unscented Kalman filter (UKF) is synthesized to estimate the states. Note that the continuous glucose monitor (CGM) readings are delayed representations of blood glucose levels and are prone to errors. This measurement is corrected for errors first before being utilized in the UKF for estimating the states.

Next, the proposed closed-loop bolus control algorithm was implemented in an Android smartphone-based artificial pancreas system. The proposed AP system is realized using commercially available insulin pumps, CGMs, and additional communication devices. The proposed algorithms were integrated into an Android smartphone app. The algorithms were implemented through two software architectures that are proposed in this thesis: (i) A dual app architecture with a Java-based front-end App and a MATLAB/Simulink-based back-end App, and (ii) a single app architecture with a Java-based front-end and a Python-based back-end. The proposed android AP systems and the algorithms were tested on actual T1DM patients. The patients underwent the standard MMTT in order to identify the model parameters after obtaining their consent. The algorithm was customized to the individual patient parameters and tested for the morning bolus cycle in a clinical environment under medical supervision. The proposed AP system was shown to successfully control the glucose values autonomously in several clinical trials, thereby proving the efficacy and robustness of the AP system with the proposed bolus control algorithm.

Hence, to continue developing a complete AP system further, the basal model and control algorithm were developed using the T1DM simulator, which was approved by the Food and Drug Administration (FDA) of the United States of America. The second part of the thesis was only tested through in-silico simulations due to a lack of funds to conduct clinical trials. The basal modes are categorized into different sub-modes due to the variation of the glucose-insulin interaction throughout the day. As there are various modes of operation in an AP system, and the transition between various modes cannot be explicitly defined, an Interactive Multiple Model Filter is also proposed to address this issue. These models were utilized for the formulation of an MPC scheme to control the glucose levels throughout the day. The proposed algorithm was tested vigorously through in-silico tests to showcase its efficacy and robustness. In order to account for uncertainties and user errors, such as failing to declare a meal, a meal detection, estimation and meal compensation strategy is also proposed in this thesis. The proposed basal and bolus algorithm, with the other associated algorithms, was evaluated through vigorous robustness tests using the FDA-approved T1DM simulator. Therefore, the research proposed in this thesis can lead towards the development of a verifiable AP system that can benefit millions of T1DM patients. A brief note on current and future research activities is given in the Concluding chapter of this thesis.