We carried out two sequential, open label, randomised controlled crossover studies comparing the overnight closed loop delivery of insulin with conventional insulin pump (continuous subcutaneous insulin infusion) therapy (fig 1). The studies mimicked the two common scenarios of “eating in” and “eating out” in preparation for testing as an outpatient. In an initial study we evaluated closed loop delivery after a medium sized evening meal, mimicking an evening at home when closed loop delivery could be conveniently started with the meal. In a second, more challenging, study mimicking an evening out, the meal was larger, was consumed later, and was accompanied by wine. Closed loop delivery was started later and continued until noon the next day.
Fig 1Design of two studies comparing closed loop delivery of insulin with conventional insulin pump therapy after two meal scenarios: eating in and eating out
Setting and participants
The studies were carried out at the Wellcome Trust clinical research facility at Addenbrooke’s Hospital, Cambridge, between February and December 2009. Inclusion criteria were adults with type 1 diabetes (World Health Organization criteria or confirmed negative for C peptide), aged 18-65, and using insulin pump therapy for at least three months. Exclusion criteria were awareness of reduced hypoglycaemia and clinically significant nephropathy, neuropathy, or retinopathy. For the eating out scenario, we recruited participants who could tolerate the amount of alcohol used and excluded participants with insulin resistance (total daily insulin dose ≥1.4 U/kg), with poor glycaemic control (HbA1c ≥10%), and who were pregnant or breast feeding. Participants signed informed consent. Figure 2 shows the flow of participants through the study.
Fig 2Flow of participants through study comparing closed loop delivery of insulin with conventional insulin pump therapy after a medium sized evening meal (eating in scenario), and study comparing closed loop delivery with insulin pump therapy (more ...)
Eating in scenario
Of 32 adults invited to take part in the study, 13 agreed and were randomly assigned to be treated overnight with either closed loop delivery of insulin or conventional insulin pump therapy during two study nights, separated by an interval of one to three weeks. Before the first study visit we analysed data from continuous glucose monitoring for up to 120 hours to optimise the delivery of insulin by conventional pump therapy. On both visits participants ate an identical evening meal, comprising 60 g of carbohydrate, at 1900, accompanied by prandial insulin. We calculated the prandial boluses according to the participants’ own insulin to carbohydrate ratio and glucose values from finger stick testing. During the intervention visit, closed loop insulin delivery was applied from 1900 until 0800 the next day. During the control visit, participants applied their usual insulin pump settings over the same timeframe.
Eating out scenario
We approached a further 41 adults with type 1 diabetes and recruited 12 to a second randomised crossover study of closed loop delivery or conventional insulin pump therapy after a large evening meal accompanied by alcohol, depicting the scenario of “eating out.” On both visits, one to three weeks apart, participants ate an identical evening meal, comprising 100 g of carbohydrate, at 2030, accompanied by prandial insulin and dry white wine (Chenin Blanc, 13% vol, Ken Forrester Wines, South Africa). Participants drank the wine (7.2 mL/kg or 6.6 units per 70 kg participant) with or after their meal, completing the meal by 2200. During the intervention visit, insulin was delivered by closed loop from 2200 until 1200 the next day. During the control visit, participants applied their usual insulin pump settings over the same timeframe.
Continuous glucose monitoring and insulin delivery
To measure subcutaneous glucose in the eating in scenario we used the continuous glucose monitoring system FreeStyle Navigator (Abbott Diabetes Care, Alameda, CA) with a 10 hour run-in calibration period. In the eating out scenario we used FreeStyle Navigator with a one hour run-in calibration period. The systems were calibrated using capillary finger stick measurements as per manufacturer’s instructions. The accuracy of the sensor, evaluated as the median relative absolute difference between sensor glucose levels and paired plasma glucose levels divided by plasma glucose levels, was 8.0% (4.5-19.3%) in the eating in scenario and 12.0% (6.8-17.2%) in the eating out scenario.
When participants arrived at the research clinic, we replaced their insulin pump with a study pump (Deltec Cozmo; Smiths Medical, St Paul, MN). This pump was connected to the established infusion site, delivering rapid acting insulin analogue Aspart (Novo Nordisk, Bagsvaerd, Denmark).
Closed loop algorithm
We used an algorithm based on the model predictive control approach.18 19
Every 15 minutes a research nurse initiated a control cycle; the nurse inputted the sensor glucose value into the computer based algorithm and adjusted the insulin pump according to the basal infusion rate calculated by the algorithm. The calculations utilised a compartment model of glucose kinetics,20
describing the effect of rapid acting insulin and the carbohydrate content of meals on glucose excursions detected by the sensor. The algorithm was initialised using participant’s weight, total daily insulin dose, and basal insulin requirements. Additionally, the algorithm was provided with glucose levels measured by the sensor during a 30 minute period preceding the start of closed loop delivery, the carbohydrate content of the evening meal, and the prandial insulin bolus. The algorithm adapted itself to participants by updating two model variables: an endogenous glucose flux correcting for errors in model based predictions, and carbohydrate bioavailability. Several competing models differing in the absorption of subcutaneous insulin and oral carbohydrates ran in parallel.21
A combined model forecasted plasma glucose excursions over a 2.5 hour prediction horizon. The algorithm aimed to achieve glucose levels between 5.8 and 7.3 mmol/L and adjusted the actual level depending on fasting versus postprandial status, preceding glucose levels, and the accuracy of predictions made by the compartment model. Safety rules limited maximum insulin infusion and suspended insulin delivery when the sensor measured glucose at or below 4.3 mmol/L or when the sensor detected that glucose was decreasing rapidly. We used algorithm version 0.02.04 to 0.02.18.
Plasma glucose and plasma insulin sampling
We collected venous samples for glucose determination every 15 minutes and for insulin assay every 30 minutes. In the eating in scenario we collected samples between 1830 and 0800 and in the eating out scenario between 1930 and 1200. These data were not used to alter insulin infusion rates during the visits for closed loop delivery or control. In the eating out scenario we collected additional samples for determination of ethanol at 2030, 2200, and midnight and then three hourly until 1200 the next day.
We used YSI2300 STAT Plus Analyser (YSI, Lynchford House, Farnborough, UK) to determine plasma glucose levels, and an immunochemiluminometric assay (Invitron; Monmouth, UK) with an intra-assay coefficient of variation of 4.7% and an interassay coefficient of variation of 7.2% to 8.1% for plasma insulin assay. We determined plasma ethanol levels using the ethyl alcohol method (Dade Behring, Atterbury, UK), with an intra-assay coefficient of variation of 2.4% and an interassay coefficient of variation of 5.71%.
Based on previous studies18
we anticipated that overnight closed loop delivery would increase the time that plasma glucose concentrations were between 3.91 and 8.00 mmol/L by a mean 37% (SD 40%). We calculated that 12 participants would provide 80% power at the 5% level of significance to detect this difference between conventional and closed loop insulin delivery systems.
Randomisation and masking
For each study we placed a computer generated allocation sequence with permuted block four randomisation in sealed envelopes. Participants were randomised on recruitment. During the eating in scenario, investigators were blinded to the plasma glucose data. For safety reasons, during the eating out scenario investigators had access to plasma glucose levels. During both studies, participants did not have access to data on plasma or sensor glucose levels.
Senior investigators and study statisticians agreed on the analysis plan in advance. The primary outcome was the percentage time plasma glucose concentrations were in the target range (3.91-8.00 mmol/L) between 1900 and 0800 in the eating in scenario and between 2200 and 1200 in the eating out scenario. Secondary outcomes were the mean glucose concentration, time when glucose concentration was 3.9 mmol/L or less (hypoglycaemia), time when glucose concentration was greater than 8.0 mmol/L (hyperglycaemia), mean rate of insulin infusion, and mean plasma insulin concentration for both separate and pooled study data. We estimated glycaemic variability by mean amplitude of glycaemic excursions22
and standard deviation of glucose concentration. Secondary outcomes were calculated for both plasma and sensor glucose levels as well as for the period from midnight to end of closed loop control.
We used the low blood glucose index to assess the duration and extent of hypoglycaemia, calculated as an average of transformed glucose measurements progressively increasing at low glucose levels.23
Safety assessments examined plasma glucose levels below 3 mmol/L and above 16.7 mmol/L.
For each outcome we fit a repeated measures regression model based on the ranked normal transformation (except for the mean glucose concentration which was not transformed because it already had an approximate normal distribution) to compare the two treatments adjusting for plasma glucose level at the start of closed loop delivery. We carried out analyses using SAS software, version 9.1 and SPSS, version 15. Kernel density of plasma glucose was estimated using the package “np” (non-parametric kernel smoothing methods for mixed data types), version 0.40-1, adopting a bandwidth of 0.25 mmol/L and implemented in the R, version 2.11.1 (R Foundation for Statistical Computing).