467921 Algorithms for a Closed-Loop Artificial Pancreas: Challenges and Solutions Pertinent to Chemical Process Operations and Control
The development of a closed-loop artificial pancreas involves the efforts of many individuals, oversight committees and regulatory agencies. In this talk we concisely review the clinical trial process, including Investigational Device Exemption (IDE) applications to the US Food and Drug Administration (FDA), oversight by Institutional Review Boards (IRB), Data and Safety Monitoring Boards (DSMB), and clinical trial registration. Typically, short-term (1-3 day) studies to verify safety and conducted in an inpatient setting (hospital) are followed by hotel- or camp-based studies that approximate the home environment but allow additional safety through monitoring and rapid response of nursing and engineering staff. Finally, based on inpatient and hotel/camp results, longer-term (weeks, months) outpatient (at-home) studies are performed.
While we have been involved in many of the major steps towards an artificial pancreas, we focus on a multiple model probabilistic predictive control (MMPPC) strategy for a fully automated closed-loop artificial pancreas that does not require meal announcement (e.g. without feedforward control). We point out lessons learned and engineering and clinical protocol changes as we moved from simulation to inpatient and hotel/camp, and finally outpatient (at home) studies.
One of the major challenges in developing a fully closed-loop artificial pancreas is that subjects will neglect to provide a meal-related insulin bolus (feedforward control), which often results in a period of post-meal hyperglycemia (high blood glucose). Our MMPPC approach detects when a meal is being consumed, estimates the meal size, and begins to more aggressively inject more insulin. Included is knowledge of typical eating patterns; for example, if someone has just consumed a large meal they are unlikely to eat again for several hours. At each time step the controller delivers insulin to assure a low probability that an individual will become hypoglycemic (low blood glucose). Similarly, when a subject is likely to be sleeping (which can be inferred for the orientation of an accelerometer), the insulin delivery is less aggressive, but again targeting a low probability of hypoglycemia. The algorithm also accounts for the increased insulin sensitivity and reduced need for insulin during exercise.
A summary glucose plot for ten subjects (over 300 subject-hours) in an in-clinic study is shown in Figure 1. The overnight performance of the closed-loop strategy is clearly significantly improved compared to open-loop control. In the presentation we will also include results of 15 subjects in a hotel trial for 3 days and 2 nights. Further, we discuss studies currently being conducted in an outpatient (at-home) trial of 20 subjects, each undergoing two weeks of closed-loop control as well as two weeks of open-loop control, for a direct comparison of performance.
So what have we learned that can be generalized and applied to chemical process operations and control? The most direct application is to batch and semi-batch processes, where the dynamic behavior and objectives change during the course of the batch. Individuals have schedules/activities that change from day-to-day, imposing challenges that are similar to multi-product chemical plants. The major concern with hypoglycemia is analogous to chemical product purity and related specifications, where there can be severe economic penalties when these specifications are violated. The long time-scale pharmacodynamic action of insulin makes it important to keep track of “insulin on board” (IOB), since insulin injected two hours ago continues to have an effect on glucose for the next three hours or so; this is most analogous to an exothermic semi-batch reactor, where overfeeding can cause excess reactant to accumulate, possibly resulting in a large exotherm. Continuous glucose monitors used for the feedback signal create what is essentially an inferential control (state estimation) problem, since at least two reference glucose (fingerstick) measurements per day are required to keep the sensor in calibration; this is analogous to the combined use of continuous process measurements along with infrequent product composition measurements that have some level of uncertainty.
There are also process development analogies – for example, initial clinical trials are conducted in a hospital environment (inpatient) on a limited number of subjects, analogous to laboratory scale chemical process development. Like process development, lessons learned and data collected during the initial inpatient studies allow model development and a tailoring of the protocols for the next clinical trial stages. Similar to pilot plant testing where multiple scale reactors are studied, hotel and camp studies involve “scale-up” where many subjects, with different characteristics and behaviors, can be studied simultaneously as a transition between R&D and use at-home. Initial at-home studies in a sense represent the final pilot plant scale, with shorter-term (two week) studies followed by much larger trials of up to six months. A number of artificial pancreas groups throughout the world are now at the larger clinical trial stage, which will be followed by extensive trials for commercially available devices.