Artificial pancreas model predictive control algorithm effective for maintaining glucose levels

NEW ORLEANS — A model predictive control algorithm for an artificial pancreas was superior to a proportional-integral-derivative algorithm for maintaining glucose levels within the safe range, according to results of a head-to-head clinical trial presented at the American Diabetes Association Scientific Session.

“There is a bewildering array of algorithms to use with the artificial pancreas,” Francis J. Doyle, III, PhD, dean of the Harvard Paulson School of engineering and applied sciences, said during his presentation. “This was our attempt at a capture-and-review paper, all of the possible variations on algorithmic design.”

Doyle and colleagues evaluated 20 adults with type 1 diabetes to compare model predictive control with proportional-integral-derivative control for an artificial pancreas. Both algorithms were compared in two supervised 27.5-hour closed-loop sessions.

The primary outcome was mean percent time in the glucose range of 70 mg/dL to 180 mg/dL.

The study challenges were designed to model the use of an artificial pancreas and also stress the glycemic-control algorithms. There was no prior optimization of insulin pump clinical parameters used to initialize the artificial pancreas included in the challenge. The challenges included overnight control after a 65-g carbohydrate dinner, response to a 50-g carbohydrate breakfast and an unannounced 65-g carbohydrate meal to determine a missed meal bolus scenario.

Compared with proportional-integral-derivative control (63.7%), model predictive control yielded a significantly greater time in the 70 mg/dL to 180 mg/dL range (74.4%; P = .02). Mean glucose during the entire trial (model predictive control, 138 mg/dL vs. proportional-integral-derivative control, 220 mg/dL; P = .012) and during the 5-hour period after the unannounced meal (181 mg/dL vs. 220 mg/dL; P = .019) were further reduced with model predictive control compared with proportional-integral-derivative control. Time in less than 70 mg/dL did not differ significantly between to two algorithms.

“We feel this is really the first balanced, even-protocol comparison of the two algorithms designed under standard conditions with no special enhancements to the algorithm,” Doyle said. “We can conclude that both [algorithms] provided safe and effective management of the patients’ blood sugar. There were statistically significant improvements in the [model predictive control] algorithm over the [proportional-integral-derivative], particularly on the primary endpoint of 70 mg/dL to 180 mg/dL time in range.” – by Amber Cox

Reference:

Lee JB, et al. 80-OR. Presented at: American Diabetes Association Scientific Sessions; June 10-14, 2016; New Orleans.

Disclosure: Doyle reports no relevant financial disclosures.