Feasibility of automated insulin delivery guided by continuous glucose monitoring, in extremely preterm infants

One Sentence Summary Closed loop systems based on subcutaneous glucose measurements could provide an efficacious and safe means of optimizing glucose control in preterm infants while reducing resources required including time of bedside personnel. Abstract Closed loop systems have been used to optimise insulin delivery in children with diabetes, but they have not been tested in neonatal intensive care. Extremely preterm infants are prone to life-threating hyperglycaemia and hypoglycaemia; both of which have been associated with adverse outcomes. Insulin delivery is notoriously variable in these babies and time-consuming, with management requiring frequent changes of dextrose-containing fluids and careful monitoring. We aimed to evaluate the feasibility of closed loop management of glucose control in preterm infants in a single centre feasibility study with a randomised parallel design. Eligibility criteria included birth weight <1200g and <48hours of age. All infants had subcutaneous continuous glucose monitoring for the first week of life, with those in the intervention group receiving closed loop insulin delivery between 48 and 72hours of age. The primary outcome was percentage of time in target (sensor glucose 4-8mmol/l). The mean (SD) gestational age and birth weight of intervention and control study arms were 27.0(2.4) weeks, 962(164) g and 27.5(2.8) weeks, 823(282) g respectively. The time in target was dramatically increased from median (IQR) 26%(6,64) with paper guidance to 91%(78, 99) during closed loop (p<0.001), without increasing hypoglycaemia. There were no serious adverse events and no difference in total insulin infused. Closed loop glucose control based on subcutaneous glucose measurements is feasible and appears to provide an efficacious means of optimising glucose control in extremely preterm infants.


Introduction
Preterm infants are at high risk of both hyperglycaemia and hypoglycaemia, predominantly related to deficit of insulin production and glycogen stores, with additive effects of parenteral nutrition, inotropic drug infusions and sepsis related insulin resistance. (1,2) Hyperglycaemia is observed in 80% of preterm infants, and glucose variability is associated with increased mortality and morbidity. (2)(3)(4)(5)(6)) Moreover, at a practical level glucose control is difficult to achieve in an extremely low birth weight preterm infant; often this requires multiple changes on intravenous infusions and insulin dosing requiring extra attention of bedside staff and general expense. The use of sliding scale insulin therapy is widespread but considered suboptimal as the desire to minimize blood sampling, , (7) combined with the extremely variable response to insulin,(1) puts these babies at risk from hypoglycaemia. (8) This results in babies often being managed with a reduction in parenteral nutrition, and potentially inadequate nutritional support, at a critical time of growth and development.
Continuous glucose monitoring has been used in neonatal intensive care to identify hypoglycaemia and is considered sufficiently accurate to support clinical management. (8,9) (10) However, the wide variation in individual insulin sensitivity and the limited staff resources make it challenging for the full potential of continuous glucose monitoring to be realized. Adaptive computerized algorithms utilizing hourly to four hourly blood glucose measurements have been evaluated in adults (11)(12)(13) and neonates undergoing intensive care. (10) (14) The addition of frequent glucose levels obtained by continuous glucose monitoring allows the development of closed loop systems as investigated in adult intensive care patients documenting its safety and efficacy. (15,16) The present study hypothesized that closed loop based on subcutaneous continuous glucose monitoring can be similarly effective in informing insulin delivery and targeting glucose control in extremely preterm infants compared with continuous glucose monitoring with insulin therapy guided by a paper algorithm. .

Study Population
Ten babies were randomly assigned to closed loop intervention and ten babies were randomly assigned to continuous glucose monitoring with insulin therapy guided by a paper algorithm.
Baseline characteristics of the two groups were similar (Table 1).
All 20 babies remained in the study throughout the intervention period from 48 to 72h with comparable amount of sensor glucose data available for analyses in each group [median (IQR) for both study groups 24h (23.75, 24.00)]. Control algorithm directed insulin therapy was followed at all times during the pre-specified 48-72h period. The maximum period of sensor signal loss during the closed loop was 3.5h during which hourly blood glucose values were used by the control algorithm. Data are presented as mean (SD)

Glucose control, insulin and dextrose administration between 48 and 72 hours
There was no difference in the baseline mean sensor glucose at 48h between study groups (  The summative glucose profiles for each study group as well as insulin infused are provided in Figure 1. Nine out of the ten babies in the closed loop group had received insulin prior to the 24 hour intervention with the one remaining baby being started on insulin during the 24 hour closed loop. This compared to four babies in the control arm having received insulin prior to, and eight babies receiving insulin during the intervention period. Four babies in the closed loop study arm received additional 20% dextrose for short periods during the intervention period (up to 3.5 hours). The mean infusion rate in these babies ranged from 0.13ml/kg/hour to 0.53ml/kg/hour.
The highest rate being infused in a baby who was hypoglycaemia prior to the start of the closed loop intervention. There was no statistical difference in the total amount of insulin infused or nutritional intake between study groups during the 24 hour intervention period.

Glucose control, insulin and dextrose administration between 72 and 160 hours
In the post intervention period after 72 hours, there was a trend of increased time in both glucose target ranges (4.0-8.0mmo/l and 2.6-10.0mmol/l) in the closed loop group compared to the control group (Table 3), but these differences did not reach statistical significance.

Nutrition and clinical care
All babies received parenteral and enteral nutrition according to the standard local neonatal unit protocol. During the closed loop intervention period between 48 and 72 hours four babies in each study group were receiving minimal amounts of trophic feeds, and in the post intervention period after 72 hours there was no difference in the volumes of milk received between the two study groups (Table 3).

Safety
There were no reported concerns about the sensor site in terms of inflammation or infection either during the study or after removal. In the closed loop study group there were two babies who had documented episodes of hypoglycaemia. One episode occurred when checking the baseline blood glucose prior to the onset of closed loop, when maintenance fluids were being changed, and when no insulin was being infused. The algorithm advised 20% dextrose that was infused. The second episode was on day 6, again associated with a change of maintenance fluids.
There were a further two babies who had periods (after the 72 hour closed loop) when sensor glucose fell to below 2.6mmol/l but the blood glucose at this time was documented above 2·6mmol/l, one baby had two episodes each lasting 10 minutes and one baby had a single episode lasting 25 minutes. In the control study group no babies had a documented blood glucose value less than 2.6 mmol/l. One baby in the control group had an episode lasting 205 minutes when the sensor glucose fell to less than 2.6 mmol/l (blood glucose was not checked at this time despite this being a study recommendation). None of the babies were on insulin, and there were no clinical concerns about hypoglycaemia in these babies during these episodes.

Discussion
This study is the first to show that a The frequency of blood glucose sampling in preterm infants is typically much lower than in adults and children in intensive care and previous studies have explored the potential for the use of subcutaneous continuous glucose monitoring (10) . These studies though remained dependent on staff responding to trends or alarms in sensor glucose before intervening. (10) This contrasts with the present study in which the targeting of glucose control is proactively driven by the closed loop algorithm, which was responding to frequently sampled sensor glucose data. This study is unique in exploring a control approach belonging to the family of model predictive control algorithms and optimized on a validated computer simulations environment (22) prior to study commencement to ensure a favorable outcome.
This is the first randomized study to evaluate the feasibility of a closed loop control based on continuous glucose monitoring in preterm infants to guide insulin delivery to support glucose control. The strengths of our study are the randomized controlled study design and comparability of the study groups and nutritional intakes. The study limitations include a small sample size, and a short study duration as well as a single center study design, which limit the generalizability but do not affect the main study outcome. Further studies are required to explore the impact of a fully automated system with infusion pumps providing insulin and 20% dextrose under fully automated computer control. Thus, closed loop insulin delivery based on subcutaneous continuous glucose monitoring appears a potentially safe, feasible and efficacious approach for targeting glucose control in preterm infants requiring intensive care.

Study design and participants
Babies were recruited from the neonatal intensive care unit at Addenbrooke's Hospital, Cambridge, UK. The study applied a randomized, open-label, single-center, two-arm, parallel design. Ethics approval was obtained prior to start of study recruitment. Eligibility criteria included birth weight less than 1200g and age less than 48h. Babies were excluded if they had a major congenital malformation or an underlying metabolic disorder, or mothers had had pregnancies complicated with diabetes. Informed written parental consent was obtained prior to study procedures.
All babies had real-time continuous glucose monitoring inserted within 48h post birth, which remained in situ for up to seven days. The paper-based guideline advised on the use of insulin or additional dextrose support. For a pre-specified period of 24h, between 48 and 72h post birth, a closed loop system controlled glucose in babies during the closed loop intervention, whereas babies in the control group continued to use real-time continuous glucose monitoring alongside the paper guideline to direct insulin therapy to maintain glucose control.

Randomization
Babies were randomized within 48h of birth to closed loop or real time continuous glucose monitoring with a paper based guideline for insulin therapy. Randomization applied the minimization methods using the Minim randomization software.(23) Randomization was stratified according to gestational age and birth weight to ensure balance between the two groups.

Common study procedures
Apart from glucose control over the pre-specified period between 48 and 72 hours, all other aspects of care including nutritional management were identical between treatment groups.

Paper algorithm
The paper algorithm for insulin delivery was developed for the purposes of studies using realtime continuous glucose monitoring to optimize glucose control in preterm babies. It provided guidance based on the absolute sensor glucose value as well as glucose trends. If sensor glucose levels were outside of the target range or demonstrated a persistent trend the advice given was to review the clinical care and consider the need to check blood glucose level, review lines and nutritional intake and to consider modifying insulin delivery dose or providing additional dextrose. The bedside nurse could initiate a physician prescribed alteration in insulin delivery based on the paper algorithm. The insulin and dextrose were delivered by Alaris pumps (Carefusion, San Diego, CA, USA).

Closed loop glucose control between 48 and 72 hours
Babies randomized to closed loop therapy were treated between 48 to 72 hours post birth using a

Assessments and data collection
Demographic and clinical characteristics were collected at study initiation. Antenatal variables were defined as: antenatal steroids as having received at least one dose prior to delivery, prolonged rupture of membranes as rupture >24 hours prior to delivery, maternal smoking included mothers who smoked at any time during pregnancy and hypertension and chorioamnionitis were based on diagnoses recorded in the maternal medical file. All blood glucose measurements, insulin administration, type and volume of enteral and parenteral nutrition and additional intravenous glucose administration were recorded from the time of randomization to the end of continuous glucose monitoring.

Statistical analysis
Investigators agreed on the outcome measures and the statistical analysis plan in advance. The primary outcome was the time spent in target glucose range between 4.0 and 8.0mmol/l as recorded by sensor glucose measurements. This data was compared between study arms. Secondary outcomes were time spent with sensor glucose levels between 2.6 and 10.0 mmol/, prevalence of hyperglycaemia (percent time sensor glucose >10·0 mmol/l), mean and standard deviation of sensor glucose. Safety endpoints included frequency of significant hypoglycaemia (any blood glucose <2.6 mmol/l) and other adverse events. As this was a feasibility study, no formal power calculations were performed. All analyses were performed on intention to treat basis. Unpaired t-test was used to compare normally distributed variables. Non-normally distributed variables were compared using Mann-Whitney U-test. Calculations were carried out using SPSS Version 23 (IBM Software, Hampshire, UK). Values are given as mean, standard deviation (SD) or median (interquartile range). P value less than 0.05 was considered statistically significant.