Performance of a closed-loop glucose control system, comprising a continuous glucose monitoring system and an AI-based controller in swine during severe hypo- and hyperglycemic provocations

Abstract

Intensive care unit (ICU) patients develop stress induced insulin resistance causing hyperglycemia, large glucose variability and hypoglycemia. These glucose metrics have all been associated with increased rates of morbidity and mortality. The only way to achieve safe glucose control at a lower glucose range (e.g., 4.4-6.6 mmol/L) will be through use of an autonomous closed loop glucose control system (artificial pancreas). Our goal with the present study was to assess the safety and performance of an artificial pancreas system, composed of the EIRUS (Maquet Critical Care AB) continuous glucose monitor (CGM) and novel artificial intelligence-based glucose control software, in a swine model using unannounced hypo- and hyperglycemia challenges. Fourteen piglets (6 control, 8 treated) underwent sequential unannounced hypoglycemic and hyperglycemic challenges with 3 IU of NovoRapid and a glucose infusion at 17 mg/kg/min over the course of 5 h. In the Control animals an experienced ICU physician used every 30-min blood glucose values to maintain control to a range of 4.4-9 mmol/L. In the Treated group the artificial pancreas system attempted to maintain blood glucose control to a range of 4.4-6.6 mmol/L. Five of six Control animals and none of eight Treated animals experienced severe hypoglycemia (< 2.22 mmol/L). The area under the curve 3.5 mmol/L was 28.9 (21.1-54.2) for Control and 4.8 (3.1-5.2) for the Treated animals. The total percent time within tight glucose control range, 4.4-6.6 mmol/L, was 32.8% (32.4-47.1) for Controls and 55.4% (52.9-59.4) for Treated (p < 0.034). Data are median and quartiles. The artificial pancreas system abolished severe hypoglycemia and outperformed the experienced ICU physician in avoiding clinically significant hypoglycemic excursions.

Keywords: Artificial intelligence; Closed loop glucose control system artificial pancreas; Continuous glucose monitor; Hypoglycemia; Intensive care unit.

Conflict of interest statement

Author MN has declared receiving financial support from Maquet Critical Care AB to perform this study. Author MW works for Maquet Critical Care AB, which was responsible for funding this study. Authors LD and JD hold stock in Ideal Medical Technologies, Inc.

Figures

Fig. 1

Protocol format. Dashed line represents expected glucose response to study interventions. Severe hypoglycemia was expected at point of time 100 min based on insulin injection (3 units) given at point 20 min. A hyperglycemic excursion greater than 10 mmol/L was induced by the glucose infusion of 17 mg/kg/min given between point 110 and 140 min with an expected maximum at point 150 min. A rebound hypoglycemia was anticipated about point 240 min based on the insulin therapy used to treat the hyperglycemic excursion. The formal study starts at point of time 0 min. Total study duration was 300 min. Observe that the target range for the AI-controller (4.4–6.6 mmol/L) is different and narrower than the target range for the conventional treatment (4.4–9.0)

Fig. 2

a Plot of median ± interquartile range (25–75). Blood glucose trend for the artificial pancreas system (Treated) is presented in blue. The trend for manual control groups (Control) is red. The insulin injection (3 units) given at point 20 min caused a more severe hypoglycemia at time 50 min in Controls compared to Treated. All indices for hypoglycemia were significantly different between groups and in advantage to the subjects controlled by the artificial pancreas system (Treated) (Table 2). b For clarity are median glucose values trended separately and without interquartile ranges. To be noticed is that severe hypoglycemia did not occur in the “Treated”-group. The Treated group had a higher peak glucose excursion at time 150 min due to the design of the AI algorithm used in the present study