Control algorithms are, by definition, designed and tuned based on a model of how a system works, ranging from the simple (knowledge of whether a manipulated input increases or decreases the output) to the complex (sets of nonlinear partial differential equations). This range trades off ease of design and implementation and possibly robustness to uncertainty with performance and ability to fine-tune and learn. These trade-offs become increasingly challenging when delays separate action and effect. Currently available insulin pumps utilize simple algorithms to incorporate current glucose levels into suggestions for bolus doses (the “bolus calculator” or “bolus wizard” features), with the availability of glucose trend from continuous glucose sensors, more sophisticated algorithms can be developed.
The simplest form of a “partial” closed-loop system would be for the delivery of insulin to be suspended when the patient is hypoglycemic and not responding to alarms. Minimed has developed such a system, the Veo® Pump which is currently only available in Europe. When the glucose is below the hypoglycemic threshold (determined by the patient) and the patient does not respond to the alarm, insulin delivery is stopped for 2 hours.
The next step would be to stop insulin delivery based on predictive alarms, i.e. the subject would not need to be hypoglycemic before basal insulin delivery is attenuated or stopped. This would be particularly important overnight, when subjects fail to respond to over 70% of alarms41
. In a clinical research center setting where basal insulin infusion rates were increased to induce hypoglycemia, predictive algorithms which triggered a suspension of basal insulin prevented hypoglycemia 75% of the nights when hypoglycemia was predicted to occur42, 43
. These algorithms could trigger a pump shut off without wakening the patient, thereby also decreasing the incidence of sleep disruption due to alarms. In our reviews of nighttime CGM monitoring during which a seizure occurred during the night, hypoglycemia was recorded on the sensor for a minimum of 2 ¼ hours prior to a seizure44
, so with this safety window, a pump shut off over the night should prevent most episodes of nocturnal seizures and dead-in-bed (4 hours of nocturnal hypoglycemia 45
), unless a large dose of insulin had been given prior to bed. Another retrospective approach would be to have a computer program review 3–6 days of CGM and pump data looking for patterns. This can be done by dividing the day into 3 hour windows with windows beginning when a meal bolus has been given. Time blocks beyond the meal blocks can be used for adjustment of basal insulin infusion rates. For a patient is using an insulin infusion pump this can be accomplished by downloading both the sensor and pump information into a common file. If there is a consistent trend seen over multiple days, this could generate a recommendation to the patient to change either a basal rate or a carbohydrate to insulin ratio for a particular meal. These suggested doses would be more accurate than what physicians initially calculate and would allow for testing of algorithms before fully closing the loop. A third partial approach to closing the loop would be to have an algorithm incorporated into the insulin infusion pump which includes glucose rate of change information as well as insulin action profiles into the bolus calculator. This would allow adjustment of meal bolus doses and basal infusion rates based on glucose trend analysis as well as glycemic targets, but the final decision on insulin delivery is done by the user.
Another partial approach to a closed-loop system would be to have a “control-to-range” algorithm that would only be active when blood glucose levels are projected to be above a user defined upper target (perhaps 160 to180 mg/dl) or below a lower target (perhaps 70–80 mg/dl). The JDRF artificial pancreas consortium is planning studies in the next year on a “control-to-range” algorithm in a clinical research center setting, and the JDRF has also signed an agreement with Animas to bring such an algorithm to the market in 4 years.
To create a fully functional artificial pancreas there must be an algorithm which determines insulin delivery. Several algorithms have been proposed including a proportional-intergral-derivative (PID) algorithm46, 46
, model predictive control47–49
, and adaptive neural networks50
. The first of these models to be tested in humans has been the proportional-integral-derivative (PID) algorithm 3
. At each point in time the controller assesses how far the current glucose is from the desired glucose (proportional), the rate of change in glucose (derivative), and how long the glucose has remained above or below target (integral). In these CRC studies on 10 subjects with type 1 diabetes who were on the artificial pancreas for 30 hours, the PID controller achieved excellent control overnight, but there was mild hyperglycemia following meals, particularly breakfast, and a tendency for hypoglycemia 4–6 hours following meal insulin delivery3
. These issues can be partly addressed by using a feed-forward algorithm where a partial meal bolus is given 5–15 minutes before the meal, i.e. a “hybrid” closed-loop. This approach was initially tested at Yale and resulted in a significant improvement in post-prandial hyperglycemia 51
. The basal rate can also be decreased several hours after a meal to compensate for the insulin onboard from the meal bolus. With extended hyperglycemia the integral component can become significant, and can only be decreased by a corresponding area under the curve below the target. To prevent this from happening, constraints can be placed on the insulin infusion rates using techniques such as “reset windup” 52
In model predictive control (MPC) the controller has a model of expected glucose values and responses to insulin which may vary by time of day (dawn phenomenon), meal events, changes in insulin sensitivity. At each point in time the model compares the predicted glucose with the actual glucose and the model is then updated with a new prediction horizon. At each step the model takes into account the previous history of glucose measurements and insulin delivery and model may be updated to learn from discrepancies between actual and predicted values, and then the optimization is repeated. How to best update the model to correct for model mismatch is one of the major challenges to MPC. MPC has been utilized in a simulated patient47
, and there are some short term studies in humans53, 54
. It should be noted that MPC is a basic strategy or concept, but any number of model types can be used, with many different methods of performing the optimization. Classic MPC uses a fixed linear model, but there have been many formulations using nonlinear models55
, including artificial neural networks56
. A nice feature of an optimization-based approach is that different weighting on the control objective can be used depending on whether the glucose is entering hyperglycemia or hypoglycemia conditions. Also, multi-objective optimization techniques can be used to rank order the important objective; for example, the highest ranked objective might be to avoid hypoglycemia.