What does it REALLY MEAN to close the loop from the CGMS to the insulin pump?

In a feedback control system, huge time delays are bad. Yet there is one clearly overwhelming fact about closing the loop between CGMS and the insulin pump: when insulin is infused subcutaneously (under the skin), it then takes the insulin another 15 minutes to reach the bloodstream. The day will come (not real soon) when the loop is closed between a venous-implanted glucose sensor that detects blood glucose with no delay and an insulin pump that delivers insulin intravenously with no delay. That future closed loop without the horrible 15 minute delay will perform well, like a real pancreas.

Yet the fact remains -- we need to close the loop now between continuous subcutaneous insulin infusion and continuous glucose monitoring systems such as they are. So a key decision has been made for us: the algorithm we need to develop now to marry CGMS with CSII must be predictive.

What is the best predictive algorithm for the insulin pump in a closed loop?
There is already a good predictive algorithm. It's the one that insulin pumps are already using. In addition, we can use adaptive filtering to achieve expert control so that the predictive algorithm can learn what is optimum for each pump user individually. It will be as if you have your own certified diabetes educator built in to your pump looking at your basal rates and C/I ratios every day and keeping them adjusted for you. Sometimes, it will actually control the basal rate in response to your actual blood sugar just like a real pancreas. In the future, it may even control the delivery of amylin and glucagon.

Capabilities of the first generation closed loop
We will close the loop on a predictive algorithm. This means that there will occasionally be times when the feedback loop can be used to control the basal rate and actually compensate for hypoglycemia and hyperglycemia automatically, and other times when it cannot. At night or other times when the insulin pump would typically have no "insulin on board," then the basal rate can be both adjusted temporarily to adjust the blood sugar and also can be "learned" with adaptive variables and "profiled" as a function of the time of day so that they match the user's actual basal insulin demand.

During the other times when the CGMS feedback cannot be used to control the basal rate (right after meal boluses, and therefore much of the day), the closed loop can still be used for other essential tasks such as:
 * adjusting the basal delivery to compensate for an imbalance of insulin relative to the BG level
 * interrupting the basal delivery to compensate for a hypoglycemic trends
 * halting the basal delivery for severe hypoglycemia
 * evaluating and adapting the insulin-to-carb ratio as a function of the time of day (i.e. "learning" the breakfast lunch and dinner ratio, automatically)
 * evaluating and adapting insulin sensitivity factors as a function of time of day
 * evaluating and adapting insulin sensitivity factors as a function of hyperglycemia magnitude and duration

Insulin pump users can compensate for the 15 minute infusion delay by bolusing manually 15 minutes before the expected meal. This way the insulin is hitting the bloodstream at about the same time as the sugar is hitting the blood, avoiding the worst of the spike in blood sugar values as when a pump user consumes simple carbohydrates. This pre-bolus and spike-avoidance will ease the task of the closed loop, improving its performance. As the technology advances, the prebolus may include pramlintide.

But for now, until a new set of implanted (or other fast) sensors that are available which truly will allow the loop to be closed in a classical control systems sense, we must realize that the predictive algorithm is the best approach. The predictive algorithm will use feedback from CGMS to control the basal delivery in an immediate sense when it makes sense to do so, but moreover the feedback will allow the insulin pump to learn the insulin pump variables expertly such that the goals of the artificial pancreas can be met now, with the existing insulin pumps and CGMS systems that are available right now.

How will states and modes be used to control an insulin pump?