Use of Microdialysis-Based Continuous Glucose Monitoring to Drive Real-Time Semi-Closed-Loop Insulin Infusion

Abstract

Continuous glucose monitoring (CGM) and automated insulin delivery may make diabetes management substantially easier, if the quality of the resulting therapy remains adequate. In this study, a semi-closed-loop control algorithm was used to drive insulin therapy and its quality was compared to that of subject-directed therapy. Twelve subjects stayed at the study site for approximately 70 hours and were provided with the investigational Automated Pancreas System Test Stand (APS-TS), which was used to calculate insulin dosage recommendations automatically. These recommendations were based on microdialysis CGM values and common diabetes therapy parameters. For the first half of their stay, the subjects directed their diabetes therapy themselves, whereas for the second half, the insulin recommendations were delivered by the APS-TS (so-called algorithm-driven therapy). During subject-directed therapy, the mean glucose was 114 mg/dl compared to 125 mg/dl during algorithm-driven therapy. Time in target (90 to 150 mg/dl) was approximately 46% during subject-directed therapy and approximately 58% during algorithm-driven therapy. When subjects directed their therapy, approximately 2 times more hypoglycemia interventions (oral administration of carbohydrates) were required than during algorithm-driven therapy. No hyperglycemia interventions (delivery of addition insulin) were necessary during subject-directed therapy, while during algorithm-driven therapy, 2 hyperglycemia interventions were necessary. The APS-TS was able to adequately control glucose concentrations in the subjects. Time in target was at least comparable or moderately higher during closed-loop control and markedly fewer hypoglycemia interventions were required, thus increasing patient safety.

People with type 1 diabetes mellitus require lifelong insulin treatment to control their blood glucose (BG) concentrations sufficiently well and to avoid long-term complications.^1-3 Continuous glucose monitoring (CGM) and automated insulin delivery, also known as closed-loop control or artificial pancreas (AP), may make diabetes management substantially easier. The first AP systems were invented independently by Albisser and colleagues⁴ and Pfeiffer and colleagues.⁵ These AP systems were large machines that required intravenous (IV) access for glucose monitoring and insulin delivery and, thus, were primarily used in hospital settings. With the introduction of subcutaneously applied CGM systems and continuous subcutaneous insulin infusion (CSII), the technical means were created for smaller-scale AP systems. However, high CGM accuracy is required to achieve optimal control with these new AP systems.^6,7 While the prolonged insulin action of subcutaneously (SC) infused insulin can be compensated in the absence of meals,^6,7 the delayed pharmacodynamic insulin response hinders adequate automated control after meals. The core of AP systems is the control algorithm that calculates insulin requirements. Many of the currently used algorithms are based on model predictive control,^8-11 while others incorporate the proportional-integral-derivative approach.^12-15

The objective of this study was to use semi-closed-loop control based on microdialysis CGM and an empirical algorithm implementing a simple glucose control model based on common insulin therapy parameters and simple meal and insulin action models. Study results of a similar empirical algorithm are published.¹⁶ In this context, the term “semi-closed-loop” relates to the fact that the system had, in principle, closed-loop properties, but also had open-loop characteristics since meals were announced to the algorithm. Thus, the control loop is closed between meals and opened for the announcement of meals.

Methods

The study protocol for this single-center, exploratory study was approved by the ethics committee at the University of Ulm, Ulm, Germany. The study was performed in compliance with the Good Clinical Practice guidelines, ISO 14511, the Declaration of Helsinki and the German Medical Devices Act. The study was performed between March and May 2003 and took place at the Institut für Diabetes-Technologie Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm, Ulm, Germany.

Twelve subjects provided written consent and were enrolled in this study. Only type 1 diabetes patients were enrolled and they had to be on CSII therapy for at least 3 months prior to study enrolment.

Microdialysis Continuous Glucose Monitoring

In this study, a microdialysis CGM system was used.^17,18 A push-pull flow generated by a peristaltic micropump resulted in a perfusion of the microdialysis membrane with 0.3 μl/min, then glucose oxidase was mixed with the dialysate, which passed over the ex-vivo sensor. Isotonic saline solution was used as perfusion solution. The generated current was averaged every 60 seconds. The fluid transportation from the catheter site to the extracorporeal sensor and the following enzymatic reaction resulted in a technical measurement delay of approximately 30 minutes. The electrical current generated in the enzymatic reaction was calibrated against BG values obtained with conventional BG meters every 24 hours. For calibration, input of 2 BG meter results was required.

Study Device: Automated Pancreas System Test Stand

The subjects were provided with the Automated Pancreas System Test Stand (APS-TS), which consisted of 2 microdialysis CGM systems (SCGM, Roche Diagnostics GmbH, Mannheim, Germany; investigational device; 1 CGM system was designated as primary sensor for CGM-based algorithm predictions while the second CGM system served as backup),¹⁸ a modified insulin infusion pump (D-Tron, Disetronic Medical Systems AG, Burgdorf, Switzerland, now Roche Diabetes Care AG), a personal digital assistant (PDA) and an interface that connected these devices to a laptop computer on which the control algorithm created by Roche Diagnostics Operations, Inc (Indianapolis, IN) was running. The laptop computer communicated wirelessly with a transmitter-receiver unit that, in turn, was connected to the PDA, the CGM systems and the insulin infusion pump. The PDA and the CGM systems had a wired connection to this unit, while the insulin infusion pump used a wireless infrared connection. For safety purposes, IV BG values obtained with a HemoCue B-Glucose meter (HemoCue AB, Ängelholm, Sweden) were entered into the control software (once every 10 minutes between 07:30 and 22:00, once every hour between 22:00 and 7:00). Two Accu-Chek® Active meters (Roche Diagnostics GmbH, Mannheim, Germany) were used to obtain capillary BG measurement values for calibration of the microdialysis CGM systems. All BG values were reported as whole-blood values.

The control algorithm provided insulin dosage recommendations every 10 minutes that were based on the previous glucose readings and the active insulin. The algorithm used common insulin therapy parameters: a glucose target value (set at 120 mg/dl for all subjects) and a glucose target range (set at 90 to 150 mg/dl for all subjects) as well as the following subject-specific parameter: 4 insulin-to-carbohydrate ratios (for breakfasts, lunches, dinners, and snacks), an insulin sensitivity factor, a basal rate, and an insulin action time. The subject-specific parameters were determined based on logbook data from an intensive monitoring phase the subjects performed at home. For safety purposes, insulin dosage recommendations exceeding 1 unit had to be approved manually by a physician. Basal rate delivery consisted of up to 6 miniboluses per hour (1 minibolus every 10 minutes). If the algorithm predicted hypoglycemia, the actual basal rate was automatically reduced to 50% of the basal rate setting.

The algorithm was able to work in 3 modes, an observation mode and 2 control modes. In the observation mode, BG values and insulin delivery data were stored, but no insulin dosage calculations took place and the patients followed their own therapy rules. In the 2 control modes, the algorithm calculated the required insulin dosages based on the available glucose and insulin delivery data. One control mode was designed to exclusively use CGM values for insulin dosage calculation, while the other was designed to exclusively use IV BG values. In this study, the second control mode was not used; all insulin dosage calculations in this study were based in microdialysis CGM values. In the CGM-based control mode, 1 of the CGM systems served as primary sensor while the second system was used only in case technical problems arose with the first system. For a CGM system to qualify as primary sensor after the run-in phase of approximately 2 hours 1) the predictive residual error sum of squares^19,20 needed to be < 30%, 2) the bias between CGM values and the IV BG values needed to be between −5% and +15% for BG values ≥ 100 mg/dl or between −5 mg/dl and +15 mg/dl for average BG values < 100 mg/dl and 3) the deviation between individual CGM values and IV BG values after correction for average bias needed to be < ±15% for BG values ≥ 100 mg/dl or < ±15 mg/dl for BG values < 100 mg/dl. The second criterion addressed the systematic difference between CGM and BG results, whereas the third criterion addressed the distribution of individual CGM results with regard to the BG results. Qualification was performed 4 times during the complete study to ensure constant performance on an acceptable level. While CGM values were being used, the algorithm compensated for the expected technological time delay of approximately 30 minutes by predicting the actual glucose concentration based on the glucose value, which reflected the subject’s glucose approximately 30 minutes earlier, and the remaining active insulin. In both control modes, a glucose change prediction within the insulin action time was calculated based on the remaining active insulin, the basal requirement up to the specific time and the subject’s insulin sensitivity factor. The recommended insulin bolus was then calculated based on the current glucose concentration (based on CGM values or obtained via IV BG), the predicted glucose change, the target value, the subject’s insulin sensitivity factor, and the basal requirement. Although the algorithm was able to recommend insulin dosages based on IV BG values, in the study presented here only the CGM-based control mode was used. The carbohydrate content of meals had to be provided manually to the algorithm that calculated the required meal insulin dose based on the insulin-to-carbohydrate factor. A schematic of the algorithm is displayed in Figure 1. More details about this algorithm can be found in the US patent no. 6,544,212.

Figure 1.

Schematic of the algorithm used in the APS-TS.

Study Protocol

Each subject remained onsite for the 4 days of the study (total duration approximately 70 hours). The subjects arrived on the morning of the first day and they were connected to the study device, the APS-TS. The observation mode was started at approximately 12:00. Subjects followed their own insulin therapy until 22:00 on the second day. At 9:00 on the second day, an IV catheter was placed to allow the measurement of venous BG. Between 22:00 on the second day and 06:00 on the fourth day, therapy was driven by the control algorithm based on CGM data. At approximately 10:00 on the fourth day, the APS-TS was removed from the subjects who left the study site after a physical examination (see Figure 2). While wearing the APS-TS, all subjects used insulin lispro.

Figure 2.

Study procedure. Green: subject arrival (day 1) and departure (day 4). Red: subject-directed therapy. Blue: algorithm-driven therapy.

During subject-directed therapy, subjects estimated the CHO content of their meals, whereas during algorithm-driven therapy, the CHO content of meals was calculated based on nutritional information and weight of the meals or their components.

Data Preparation

Data from all 12 subjects were included in the analysis. Both observation mode and control mode were used for more than 30 hours each. However, to avoid influences from run-in effects, the last 24 hours of observation mode (22:00 on day 1 to 22:00 on day 2) were compared to the last 24 hours of control mode (06:00 on day 3 to 06:00 on day 4). For 2 subjects, control mode data were only available between approximately 07:00 on day 3 to 06:00 on day 4, because technical complications did not allow for starting the control mode at 22:00 on day 2. All calculations referring to glucose results are based on the primary CGM sensor.

Safety Assessment

Safety was assessed by capturing hypo- and hyperglycemia interventions. Hypoglycemia interventions, that is, consumption of 18 g carbohydrates, were performed at BG concentrations below 50 mg/dl (measured in whole blood) or in presence of hypoglycemia symptoms. Hyperglycemia interventions, that is, delivery of additional insulin, were performed if BG concentrations were above 250 mg/dl for more than 1 hour, if BG concentrations exceeded 350 mg/dl at all, in the case of a positive ketone test or in presence of hyperglycemia symptoms when BG concentrations exceeded 200 mg/dl. In some cases, especially at night, hypo- and hyperglycemic CGM values did lead to additional BG measurements to confirm hypo- and hyperglycemia.

Results

All of the 12 subjects (5 female, 7 male, 31 to 54 years of age, mean ± standard deviation (SD): 40 ± 7 years) included in this study had type 1 diabetes mellitus that was diagnosed 22 ± 8 years earlier (range: 11 to 34 years) and the subjects were under adequate glycemic control with a glycated hemoglobin (HbA1c) level of 6.9 ± 0.9% (range: 5.7 to 8.6%). The subjects’ mean body mass index was 23.8 ± 3.4 kg/m² (range: 18.5 to 28.8 kg/m²).

During algorithm-driven therapy, that is, in control mode, all insulin dosage recommendations made by the algorithm exceeding 1 unit were approved by the physicians, no recommendation was rejected.

The mean glucose during subject-directed therapy (observation mode) was 113.5 ± 33.3 mg/dl compared to 125.0 ± 17.2 mg/dl during algorithm-driven therapy (control mode), see also Table 1. Mean glucose traces are shown in Figure 3.

Table 1.

Results From Subject-Directed Therapy and Algorithm-Driven Therapy.

	Subject-directed therapy	Algorithm-driven therapy	P value
Blood glucose in mg/dl	113.5 ± 33.3	125.0 ± 17.2	.1078
Percentage of time in target range (90-150 mg/dl)	46.1 ± 12.7	57.6 ± 13.5	.0326
Percentage of insulin recommendations > 1 IU/h	31.5 ± 29.0	30.2 ± 15.1	.8454
Number of patients with hypoglycemic events
<70 mg/dl	11	11	—
<60 mg/dl	10	9	—
<50 mg/dl	9	3	—

Results are displayed as mean ± standard deviation. P values were calculated using paired t tests. Differences with P values < .05 were considered statistically significant.

Figure 3.

Glucose traces for the subject-directed therapy (observation mode) and algorithm-driven therapy (control mode). One-sided error bars represent 1 standard deviation. Data for the algorithm-driven therapy were obtained between 6:00 on day 3 and 6:00 on day 4 (blue diamonds); data for the subject-directed therapy were obtained between 22:00 on day 1 and 22:00 on day 2 (red diamonds).

The time in target (90 to 150 mg/dl) was approximately 46% during subject-directed therapy and approximately 58% during algorithm-directed therapy. Time below 60 mg/dl and time above 250 mg/dl were higher during subject-directed therapy (see Figure 4).

Figure 4.

Percentage of time within specific glucose ranges for the subject-directed therapy (observation mode) versus algorithm-driven therapy (control mode). Target range was 90 to 150 mg/dl. Error bars show +1 standard deviation.

On average, subjects required slightly more insulin during subject-directed therapy (48.2 ± 15.5 units per day) than during algorithm-driven therapy (46.0 ± 16.6 units per day) (see also Table 1). During subject-directed therapy, 38 hypoglycemia interventions and no hyperglycemia interventions were required as opposed to 12 hypoglycemia interventions and 2 hyperglycemia interventions during algorithm-driven therapy. Information about hypo- and hyperglycemia interventions can also be found in Table 1 and Table 2.

Table 2.

Comparison of the Number of Subjects With Hypoglycemic Events Below 70 mg/dl, 60 mg/dl, and 50 mg/dl.

		Algorithm-driven therapy
<70 mg/dl		Yes	No	Total
Subject-directed therapy	Yes	10 (83.3%)	1 (8.3%)	11 (91.7%)
	No	1 (8.3%)	0 (0.0%)	1 (8.3%)
	Total	11 (91.7%)	1 (8.3%)	12
	P value	1.0000
<60 mg/dl		Yes	No	Total
Subject-directed therapy	Yes	7 (58.3%)	3 (25.0%)	10 (83.3%)
	No	2 (16.7%)	0 (0.0%)	2 (16.7%)
	Total	9 (75.0%)	3 (25.0%)	12
	P value	.6547
<50 mg/dl		Yes	No	Total
Subject-directed therapy	Yes	3 (25.0%)	6 (50.0%)	9 (75.0%)
	No	0 (0.0%)	3 (25.0%)	3 (25.0%)
	Total	3 (25.0%)	9 (75.0%)	12
	P value	.0143

Displayed are 2 × 2 tables showing the numbers of subjects who had hypoglycemic events (“yes”) and who did not have hypoglycemic events (“no”). P values were calculated using McNemar’s test. P values < .05 were considered to show statistically significant differences.

Equipment failures were captured throughout the study. Out of 45 equipment failures in total, 24 related to the APS-TS. For 2 subjects, the algorithm initially used an inappropriate CGM lag time, which was corrected. For these 2 subjects, control mode started at 07:00 on day 3 instead of 22:00 on day 2. In 6 cases, the algorithm showed unanticipated error messages that lead to a computer restart in 2 cases and a APS-TS replacement in 1 case. The APS-TS insulin pump caused 2 failures, 1 battery failure (the battery was then replaced) and 1 error that could not be resolved and which lead to the replacement of the pump. Two times, the battery of the PDA had to be changed. In 12 cases, the CGM devices did not function properly after sensor application and the specific sensor units were replaced.

Discussion

Overall, the control algorithm used in this study improved the quality of the subjects’ diabetes therapy by increasing time in target and simultaneously decreasing time below 60 mg/dl and above 250 mg/dl. During algorithm-driven therapy, fewer hypoglycemia interventions were required as compared to during subject-directed therapy. The algorithm used in this study was a so-called empirical algorithm, which modeled basal-bolus insulin therapy based on empirically gathered therapy parameters, and compared to model predictive controllers,^9,21-23 it was relatively simple yet provided adequate control. One major drawback of the empirical algorithm was that the subjects’ therapy parameters had to be known with adequate accuracy. In some cases, inadequate subject-directed therapy was reflected by a correspondingly inadequate algorithm-driven therapy. Recently, data from a large, international, multicenter study were published²⁴ in which closed-loop control (using 2 separate model predictive control algorithms) and open-loop control were compared. In that study, time in target was at approximately 60% during 23 hours of closed-loop control which included 3 meal challenges (with announced meals) and an exercise challenge.²⁴ The target range was 3.9 to 8.0 mmol/l (approximately 70 to 145 mg/dl) except for the first 3 hours after meals, when it was 3.9 to 10.0 mmol/l (approximately 70 to 180 mg/dl).²⁴ Both target ranges of the multicenter study thus allowed a broader spectrum of glucose concentrations than the target range in this study.

The microdialysis CGM system used in this study had a technological measurement delay of approximately 30 minutes, which is substantially higher than the measurement delay of other microdialysis CGM systems or needle-type CGM systems.^25-28 However, the algorithm seemed to be able to compensate for this delay by adequately predicting the current glucose concentration. While needle-type CGM systems are easier to use than microdialysis CGM systems, the microdialysis CGM systems were reported to provide higher performance.^26,29

The APS-TS might be most suitable for clinical settings because in clinical settings the usage of multiple devices may pose less of a problem than in everyday life, and, in addition, in some cases IV data may be a feasible alternative to CGM data. In clinical settings, exact carbohydrate contents of meals may be easier to provide to the algorithm as opposed to diabetes patients’ carbohydrate estimations at home.³⁰ Although meal detection is possible,³¹ the additional time lag it introduces might possibly be too large for adequate glucose control. The APS-TS could also be used to provide adequate overnight glucose control. Patte and colleagues³² presented data on overnight control from a study, in which another version of the APS-TS was used. In hospital settings, the control algorithm could run on a central computer, thus enabling hospital staff to supervise multiple patients simultaneously.

With the technological advances in recent years, it could be possible for an updated version of the APS-TS to run on smart phones or tablet computers. This could, in turn, replace the PDA and the computer laptop option, and make use of needle-type CGM systems, whose measurement performance has improved over the last years.^33-37 Both updates would serve the miniaturization of the APS-TS and likely increase the ease of use.

Conclusions

In this study, semi-closed-loop algorithm-driven therapy was compared to subject-directed therapy in a clinical setting. Microdialysis CGM values were provided to the control algorithm, which then provided insulin dosage recommendations. Despite using a relatively simple algorithm based on empirically gathered patient therapy parameters, the algorithm provided adequate quality of diabetes therapy during this study. Time in target was increased and fewer hypoglycemia interventions were required during closed-loop control as opposed to subject-directed therapy.