Continuous glucose monitoring in extremely preterm infants in intensive care: the REACT RCT and pilot study of 'closed-loop' technology

Hyperglycaemia and hypoglycaemia in the preterm infant

Hyperglycaemia and hypoglycaemia have both been associated with increased mortality and morbidity for preterm infants. 9,10 Hyperglycaemia can lead to acute problems of persistent osmotic diuresis and metabolic acidosis, which can be difficult to control, and has been associated with increased risk of intraventricular haemorrhage. 11 Hyperglycaemia has also been associated with increased long-term morbidity, including increased risk of retinopathy of prematurity (ROP). 9,12,13 Pivotal single-centre studies in adult intensive care have demonstrated that tight glycaemic control can reduce both mortality and morbidity. 14 However, these results have been difficult to replicate because of the risk of hypoglycaemia. 15 A trial in paediatric intensive care, targeted to reduce hyperglycaemia, demonstrated a reduction in length of intensive care admission and mortality, but a significant increase in hypoglycaemia. 16 This is of particular concern as attempts to reduce the prevalence of hyperglycaemia may increase the risk of hypoglycaemia in preterm infants because of their variable sensitivity to insulin. In addition, the developing brain appears to be particularly vulnerable to both hyperglycaemic17 and hypoglycaemic insults. A Cochrane review9 has highlighted the need for further studies into the impact of interventions to improve glucose control in these infants.

Current glucose monitoring in neonatal intensive care

Although glucose monitoring is implemented in all very preterm infants, it is currently limited to intermittent blood sampling, with long periods when glucose levels are unknown. The reason for the intermittent nature of glucose monitoring is that current methodologies for measurement of glucose levels are dependent on blood sampling either from a central arterial line or by heel prick. The very small circulating volumes of blood in preterm infants mean that it is also important to minimise the number of blood samples and volume of blood sampled in this way. Furthermore, current practice in neonatal intensive care units (NICUs) aims to reduce the frequency of handling of infants, as this has been shown to improve outcomes. 18 By contrast, other physiological parameters, such as oxygen saturation, blood pressure and heart rate, are all monitored continuously to prevent wide fluctuations. It is increasingly thought that fluctuations in glucose levels may also have a significant impact on long-term outcomes. 19

If clinical interventions to optimise glucose control are to be safe and effective in the intensive care setting, robust methods of monitoring glucose levels in real time need to be in place. Furthermore, to fully understand the clinical significance of hyperglycaemia and hypoglycaemia, it is imperative that robust data are available regarding their true prevalence throughout the period of intensive care. Current follow-up studies are dependent on infrequent measurements of blood glucose, typically taken for clinical reasons and with the inherent associated bias. Continuous glucose monitoring (CGM) therefore has the potential to both safely guide acute clinical management to optimise control and provide robust data on which to assess the long-term impact.

Continuous glucose monitoring

A range of technological methods have been tested for their abilities to support CGM. Currently, the only CGM method used in clinical practice for the management of patients with diabetes involves measurement of interstitial glucose levels, which are calibrated against blood glucose measurements. CGM devices comprise a disposable subcutaneous oxidase-based platinum electrode that catalyses interstitial glucose, generating an electrical signal that is transmitted to a monitor for recording or display. Three companies have devices on the market for CGM: Dexcom Inc. (San Diego, CA, USA), Abbott Laboratories UK (Berkshire, UK) and Medtronic plc (Watford, UK). However, because of the size of extremely preterm infants and issues of insertion and attachment to these small infants, only the Medtronic plc devices are currently suitable for use in the extremely preterm infant.

Although this is a rapidly changing field, early CGM devices were used to collect data in real time, but these data could be downloaded only retrospectively to be reviewed and used to guide future clinical treatment. This is useful (1) in the setting of research, (2) to provide advice on long-term behavioural or medical interventions and (3) in the case of patients with stable diabetes, to review patterns of glucose control over time. These devices have also been used in preterm infants without side effects and identified significant periods of both hyperglycaemia and hypoglycaemia that were undetected clinically. 20–22 However, they are not suitable for providing data that is useful for acute clinical management. Newer CGM devices allow the data to be viewed in real time, providing information on glucose trends, with the potential to identify episodes of hyperglycaemia and hypoglycaemia, thus allowing the possibility of earlier intervention and prevention.

Development of continuous glucose monitors

The development of CGM devices provides the opportunity for data on glucose levels to be recorded every 5 minutes and viewed continuously. A number of different devices are available for clinical use and they are increasingly found to help reduce glycated haemoglobin and glycaemic variability in patients with type 1 diabetes without increasing the risk of hypoglycaemia. These devices allow patients to monitor their glucose levels by displaying absolute glucose concentrations as well as trends. Some systems depend on intravenous access, some use microdialysis and others measure interstitial glucose levels. Their use has been trialled in adults in intensive care and in patients requiring cardiac surgery. 23 Although these studies have been limited, they have shown that the devices are accurate and safe in cardiac surgical patients and can reduce the risk of hypoglycaemia in adults in intensive care. 24 Some have raised concerns about the lack of absolute sensitivity of individual readings at glucose thresholds, emphasising the potential importance of CGM as an early-warning system for both hyperglycaemia and hypoglycaemia. 25

These studies have highlighted the need for methodologies that can provide real-time data and allow adjustment of clinical management in the setting of a rapidly changing clinical picture of intensive care. 26,27 These benefits have been seen in the adult intensive care setting, where blood glucose levels are routinely measured much more frequently than in NICUs. Therefore, the potential for benefits in the NICU setting, where the frequency of blood glucose measurements is much lower, is likely to be more clinically significant. Key recent developments in these devices include extended sensor life, such that sensors can now remain in situ for 6 days (previously 72 hours), and changes to the sensor construction and the calibration algorithms, which have led to improved overall accuracy. 28 This is particularly important in terms of sensitivities to detect threshold levels of hypoglycaemia and hyperglycaemia.

Previous use of continuous glucose monitors in neonates

There is limited use of CGM in NICUs. We have used blinded CGM devices as part of an international multicentre trial in preterm infants and assessed the accuracy of these devices compared with current clinical practice. 29,30 These studies suggested that the value of CGM could be in providing early warning of fluctuations in glucose levels to guide the need for blood glucose assessment. In addition, by providing a continuous readout, episodes of hyperglycaemia and hypoglycaemia could be anticipated and therefore avoided. We have previously collected ‘blinded’ CGM data as part of the Neonatal Insulin Therapy [Nirture (Neonatal Insulin Replacement Therapy in Europe Study)] trial using a CGM (Medtronic plc)21 and in a small feasibility study using the iPro™2 (Medtronic plc). In the small feasibility study of 10 infants with a birthweight ≤ 1200 g monitored from birth for up to 7 days, infants had a median gestational age at birth of 28.4 weeks and a median birthweight of 840 g. The median time in the target glucose range of 2.6–10 mmol/l was 58% and the median time in the target glucose range of 4–8 mmol/l was 40%.

Study design

This was a single-centre study, with infants recruited from the NICU at Addenbrooke’s Hospital, Cambridge, UK. Ethics and trust research and development approval were obtained prior to the start of study recruitment, and informed parental consent was obtained prior to any study procedures. Inclusion criteria were infants with a birthweight ≤ 1200 g and aged ≤ 48 hours for whom informed parental consent had been received. Infants were excluded if they had a major congenital malformation or an underlying metabolic disorder, or if the mothers had a pregnancy complicated with diabetes. The intervention lasted for up to 7 days. Data regarding glucose levels, clinical condition and nutritional intake, as well as insulin use, were collected prospectively.

Continuous glucose monitoring and management

Real-time CGM was performed using the Paradigm^® Veo™ system (Medtronic plc) (Figure 1). Enlite™ sensors (Medtronic plc) were linked to the Paradigm^® Veo™. The sensor glucose data could be viewed in real time and were used in conjunction with the guideline to support clinical management. The guideline provided simple guidance, was not a rigid guideline and had not undergone formal in silico testing. The nurses recorded the sensory glucose value, alongside standard hourly clinical observation, using it to guide the need for blood glucose testing. The guideline prompted review and intervention based on both absolute glucose levels and change. The CGM devices were calibrated at least twice daily using a blood glucose measured on the point-of-care (POC) StatStrip^® meter (Nova Biomedical, Waltham, MA, USA). The StatStrip meter was chosen because it has been validated for accuracy in the newborn and intensive care settings.

FIGURE 1.

Continuous glucose monitoring system. (a) Enlite sensor; (b) Paradigm Veo and MiniLink^® REAL-Time Transmitter (Medtronic plc) attached to a sensor; and (c) preterm infant with device in situ. Reproduced with permission from Thomson et al. 1 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Thomson et al. 1 This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Blood glucose monitoring

Blood glucose levels were measured using a combination of arterial, venous and capillary samples, and were tested on the blood gas analyser (cobas b 221; Roche Diagnostics, Welwyn Garden City, UK) and the Nova StatStrip.

Analyses

Recruitment

Twenty-one infants were successfully recruited to the feasibility study. One infant was withdrawn because of failure of sensor insertion. The baseline demographic details of the remaining 20 infants with CGM data are shown in Table 1. The recruitment rate fluctuated widely from month to month and was delayed by two factors: (1) availability of eligible infants to approach and (2) availability of appropriately trained staff to undertake the study procedures. During the recruitment period, the NICU underwent a period of reduced admissions associated with the introduction of a new electronic medical records system, resulting in fewer infants being admitted and available for recruitment. In addition, sensor insertion was initially limited to key study staff to facilitate a core of experience within the team. However, the unpredictability of preterm deliveries, combined with the need for many mothers to be transferred from other hospitals, resulted in a very small window for recruitment and meant that staff with the key skills for sensor insertion were not always available.

Accuracy of continuous glucose monitors

Comparative data providing more than two glucose measurements (blood glucose or sensor glucose) taken within 5 minutes of each other were available at 247 time points. Bland–Altman analyses provided a mean bias between POC and CGM of –0.27 [95% confidence interval (CI) –0.35 to –0.19] (Figure 2). Error grid analyses (comparing sensor glucose with either blood glucose methodology) demonstrated that 98% of values lay within area A or B (see Figure 2). Performance in relation to the ISO 200331 and ISO 201332 standards is provided in Table 2.

FIGURE 2.

Feasibility study: Bland–Altman and error grid plots. 33 (a) Bland–Altman comparison of CGM sensor glucose with StatStrip meter; (b) error grid plot comparing CGM sensor glucose with StatStrip meter or blood gas (cobas b 221) glucose values; (c) Bland–Altman comparison of POC StatStrip meter with blood gas (cobas b 221) glucose values; (d) error grid plot comparing POC StatStrip meter with blood gas (cobas b 221) glucose values; (e) Bland–Altman comparison of CGM sensor glucose with blood gas (cobas b 221) glucose values; and (f) error grid plot comparing CGM sensor glucose with blood gas (cobas b 221) glucose values. BG, blood glucose; SG, sensor glucose. Orange dots represent outliers whose results were reviewed and were deemed to be secondary to pre-analytical error in blood glucose sampling. The reference lines show the estimate of the mean difference, 95% CIs around the mean estimate and the predictive interval, the region in which a new observation would be expected, with 95% confidence, to be observed. Reproduced with permission from Thomson et al. 1 This is an open access article distributed in accordance with the Creative Commons Attribution (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Thomson et al. 1 This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Safety

There were no serious adverse device effects (SADEs). One infant died prior to discharge from hospital, but after the intervention period. Death was due to prematurity and felt to be unrelated to the study intervention. One infant was withdrawn from the study early as the mother became acutely unwell and, in the light of the mother’s mental health, it was felt appropriate to withdraw the infant from the study. One infant was withdrawn because of failure of sensor insertion. This was the second infant recruited to the study. The Enlite sensor failed to insert on two attempts and remained within the needle guard. There were no concerns about infection or trauma at the site of sensor insertion and the CGM device did not interfere with clinical care.

There were three infants who had clinically documented episodes of hypoglycaemia, defined as any blood glucose measurement < 2.6 mmol/l. In each of these infants, the falling sensor glucose level had prompted action to be taken to check the blood glucose level. There were no concerns about the infants during these episodes and without the falling sensor glucose level the blood glucose level would not have been checked. After confirming a blood glucose level < 2.6 mmol/l, these infants were treated with intravenous dextrose. In two cases, the falling sensor glucose level also revealed failure of central access. In two of these episodes, sensor glucose values were also < 2.6 mmol/l (recorded on the CGM devices as episodes lasting 40 and 25 minutes, respectively). In the other case, the lowest sensor glucose value was 2.7 mmol/l. There was one further episode when the sensor glucose fell to < 2.6 mmol/l for 70 minutes, but this was not documented clinically because of a protocol deviation, as blood glucose was not measured at this time. There were no concerns about any of the infants during any of these episodes and the infants were not receiving any insulin.

Utility

The overall assessment of the CGM devices by research and clinical staff was that they were easy to use in terms of insertion and calibration. There were occasions when there was loss of connectivity between the sensor and the monitor (e.g. if an infant was moved out of an incubator for kangaroo care or someone stood between the infant and the monitor). On such occasions, it was evident that the transmitter range was limited if transmission was blocked by objects between the sensor and the monitor. This could be resolved by moving the monitor closer to the infant. Staff found the use of both predictive and threshold alarms challenging, considering them an unnecessary addition to recording sensor glucose levels hourly. As a result of frequent requests to silence these alarms, they were subsequently turned off in the early stages of the study to ensure continued staff engagement.

With a need for twice-daily calibrations, the median number of blood samples taken for glucose measurement per 24-hour period did not appear to be excessive or significantly different from the number of samples taken in other infants of similar gestational and postnatal age on the unit. The median of the mean number of samples per day per infant was 4.1 (range 3.3–4.3).

Staff expectations varied greatly, with some nurses excited about the potential to limit the need for frequent blood glucose sampling and avoiding ‘hurting the baby’; however, others had concerns about increased workload, sensor insertion, risk of ‘tissue damage’ or difficulties in positioning an infant following sensor insertion. Comments varied from ‘exciting’ to ‘how much extra time will it take up?’. After caring for an infant using CGM, there was an over-riding view that the intervention improved the quality of care (Figure 3). Comments included ‘found the monitor to be useful when the infant was on insulin’, ‘the chart . . . is really useful’, ‘I think this is the best treatment’ and ‘it is not extra workload’. Nursing staff reported positively about the impact on care in terms of feeling more empowered to manage the infant’s glucose control. They did not report any significant negative effects for the infant or the clinical/nursing team caring for the infant.

FIGURE 3.

Feasibility study: staff assessment of the use of continuous glucose monitoring. 1 (a) Do you think the baby is distressed?; (b) How did you find the monitor to read and calibrate?; (c) Do you think the device has interfered with the nursing care?; (d) Did having the baby on the monitor significantly affect your workload?; and (e) Did you think that being able to monitor glucose continuously had an impact on clinical care? Data are presented as a per cent of the responses received. Reproduced with permission from Thomson et al. 1 This is an open access article distributed in accordance with the Creative Commons Attribution (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Thomson et al. 1 This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Efficacy

A total of 2626 hours of CGM data were collected from 20 infants, with a median for individual infants of 144 (range 7–166) hours. There was wide variability in glucose control in this population. Data demonstrating the percentage of time within different target thresholds are provided in Table 3. The median percentage of time in the target range of 2.6–10 mmol/l was 78% (IQR 59–94%), with values > 10 mmol/l recorded 22% of the time.

Study design

This study was an international open-label parallel-group RCT, with infants recruited between July 2016 and January 2019 from 13 NICUs in the UK, Spain and the Netherlands. The protocol has been published. 2

Eligibility criteria

Infants were eligible if they were within 24 hours of birth, weighed ≤ 1200 g and were born at ≤ 36⁺⁶ weeks gestation, and if written parental consent had been obtained. Exclusion criteria included any lethal congenital malformations and any congenital metabolic disorder known at trial entry. Infants were also ineligible if, in the opinion of the treating clinical consultant, they had no realistic prospect of survival.

Recruitment procedure

All infants were recruited within 24 hours of birth at one of the NICUs that had been approved for study participation as a level 3 unit with capacity and capability to recruit. Owing to the short time frame from birth to recruitment, potentially eligible infants were identified in a number of ways: (1) liaison with the obstetric team to highlight mothers at risk of preterm delivery, (2) liaison with the neonatal transport team to identify infants who had been born in local units and being transferred to a study centre and (3) liaison with the NICU clinical team of study centres. Screening of eligible infants was undertaken in collaboration with the clinical team and families were approached for consent only if they were considered eligible. Screening logs were reviewed regularly by the co-ordinating centre to identify any issues around recruitment. The investigator or a suitably qualified person designated by the principal investigator was able to receive written informed consent from the infant’s parent/legally acceptable representative before any trial-specific activity was performed.

Randomisation and blinding

Randomisation to the standard-care and CGM arms of the study in a 1 : 1 ratio was undertaken centrally using a web randomisation system, Trans European Network ALEA software (TENALEA; ALEA Clinical, Abcoude, The Netherlands). This used block randomisation, with blocks of random size (i.e. 4, 6, 8) stratifying by recruiting centre and gestation (< 26 or ≥ 26 weeks’ gestation). The management of glucose control and use of CGM from recruitment at aged ≤ 24 hours until day 7 were predetermined in the protocol for each study arm. 2 Clinical outcome data were collected until 36 weeks’ corrected gestational age. No other aspects of concomitant care were prohibited during the trial.

Intervention

All infants were fitted with a sensor (i.e. the Enlite glucose sensor) that was linked to a MiniMed™ 640G System (Medtronic plc). Sensors were inserted subcutaneously (into the thigh) by hand and not using the standard insertion device, thereby ensuring that the sensor was inserted into the subcutaneous tissue. The sensors are soft and flexible, approximately 8.75 mm in length, and are mounted inside a hollow needle to allow for subcutaneous insertion. Once the sensor was inserted, the introducer needle was withdrawn and the sensor attached to a small Guardian™ 2 Link transmitter [Medtronic plc; Conformité Européenne (CE) Mark 2013; certificate number 8858] for data transfer to the MiniMed 640G System for data viewing. The sensors were secured with a clear occlusive dressing so that the insertion site could be inspected daily. A blood sample was required every 12 hours for calibration. For consistency across sites, all units were provided with Nova StatStrip meters for calibration of the CGM.

Efficacy

Efficacy was assessed by comparing the data collected from the CGM in the CGM and standard-care arms of the study.

Safety

Safety was assessed in three areas: (1) incidence of hypoglycaemia, measured as part of clinical care (blood glucose levels) and after review of sensor glucose data; (2) device safety through adverse device effect (ADE) reporting; and (3) acute mortality and morbidity outcomes as part of the CRF.

Utility

Nursing staff and parents caring for the infants in the CGM arm of the study were asked to complete a study-specific questionnaire.

Health economic evaluation

Data were collected on the health service resources used in the treatment of infants during the period between randomisation and 36 weeks’ gestational age, and based on the British Association of Perinatal Medicine standard criteria for level of care,38 as well as neonatal complications. Current UK unit costs were applied to each resource item and a per diem cost for each level of neonatal care was based on Department of Health and Social Care reference costs calculated on a full absorption costing basis. 39

Primary outcome

The primary outcome was defined as the percentage of time that the sensor glucose level was in the target range of 2.6–10 mmol/l. This was selected as the internationally most widely accepted clinical target range for glucose control in this population.

Secondary outcomes

Health economic outcomes

Cost-effectiveness was expressed in terms of incremental cost per additional case of adequate glucose control between 2.6 and 10 mmol/l.

Exploratory outcomes

Exploratory outcomes included mortality before 36 weeks’ corrected gestational age, maximum severity of ROP, bronchopulmonary dysplasia (BPD) (i.e. the need for supplemental oxygen or respiratory support at 36 weeks’ corrected gestational age), microbiologically confirmed or clinically suspected late-onset invasive infection, necrotising enterocolitis (NEC), patent ductus arteriosus (PDA) requiring medical or surgical treatment, maximum grade of intracranial haemorrhage (Papile and Burstein grading), growth (weight, length and head circumference at the end of week 1 and at 36 weeks’ corrected gestational age), nutritional intake during the first week of life (carbohydrate, protein and lipid), and use of insulin during the first and second week of life.

Efficacy analyses

Analyses were predetermined to be undertaken based on ‘full analysis’ (primary outcome) and ‘per-protocol’ populations to compare the CGM group with the standard-care group and were prespecified in the statistical analysis plan. Linear regression was used to estimate the absolute difference in the percentage of time that the sensor glucose level was in the target range of 2.6–10 mmol/l, adjusting for baseline variables (i.e. centre and gestation). Estimates of treatment effect, with 95% CIs and p-values, were calculated. Secondary end points that were continuous variables were analysed in a similar fashion and binary/ordinal variables were analysed using logistic/ordinal logistic regression.

To control the study-wide significance level to 5%, a Benjamini–Hochberg multiple testing procedure was applied to three end points of equal importance (i.e. the percentage of time that the sensor glucose level was in the target range of 2.6–10 mmol/l, the percentage of time that the sensor glucose level was in the target range of 4–8 mmol/l and mean sensor glucose level), followed by a gatekeeping sequence applied in the following order: sensor glucose variability, as assessed by within-infant SD; and percentage of time that glucose levels were in the hyperglycaemic range (i.e. sensor glucose level > 15 mmol/l). Nominal p-values and CIs are presented, unless otherwise stated.

Predefined subgroup analyses were undertaken by estimating, using the regression framework in an exploratory, non-confirmatory manner, treatment interaction effects with the following baseline variables: centre, sex, corrected gestational age, birthweight standard deviation score (SDS), use of antenatal steroids, maternal chorioamnionitis, maternal diabetes and first sensor glucose level.

Utility analyses

Acceptability of the intervention was assessed by a specifically designed staff questionnaire, which was completed anonymously by staff caring for an infant in the CGM arm on days 3 and 7, and by a similar parent questionnaire (completed on day 7). Compliance in the CGM group was assessed as a percentage of time that sensor glucose readings were recorded on the CRF and when the insulin rate of infusion was recorded if the sensor glucose > 8 mmol/l. In the standard-care group, non-compliance was defined as 12 or more CGM readings available during the time between the tamper tag being broken and resealed on at least one occasion.

Cost-effectiveness analyses

An incremental cost-effectiveness analysis was performed, details of which are presented in Chapter 5.

Regulatory requirements

Ethics and regulatory approvals were obtained and the protocol is in the public domain (ISRCTN12793535),2 with a list of protocol amendments provided in Appendix 1. A DMEC and TSC were appointed and reviewed the data in accordance with their formal charters (see Appendices 2 and 3). Research teams at each site were required to have up-to-date good clinical practice training and have undertaken training in study procedures, including the use of the CGMs and Nova Biomedical POC devices. Paper and online resources, as well as a call line, were available to support the research teams.

Clinical trial authorisation was granted by the Medicines and Healthcare products Regulatory Agency (MHRA) (reference CI/2016/0011). Written approvals were received from individual hospital sites prior to recruitment. Approvals were also obtained from the Medical Ethics Review Committee (Vrije Universiteit University Medical Centre, Amsterdam, the Netherlands) and Heath Care Inspectorate (Heerlen, the Netherlands) (reference 2017-1 398434/VlO 14949), the Research Ethics Committee, Sant Joan de Déu Research Foundation (Barcelona, Spain), and the Ministry of Health Social Services and Equality (reference 591/16/EC) (Madrid, Spain). The trial was conducted in accordance with the principles of the Declaration of Helsinki40 and in conformity with the European Union medical devices regulations and any relevant amendments. 41

Monitoring

Data returns were continually monitored by the central team for the completeness and timeliness of all data returned. Compliance with intervention strategy in each study arm was also reviewed to ensure that there was no ‘crossover’ between study arms. A monitoring plan was in place, determining frequency and scope of site monitoring based on continuing risk review. Face-to-face monitoring visits were initially undertaken within the first 6 months; the interval between visits was then adjusted following assessment of recruitment rate and the number of data queries and serious adverse event (SAE) reports. The study sites were provided with direct access to all trial-related source data and reports for the purpose of monitoring and auditing by the central study team, sponsor and regulatory authorities, as required.

Data and safety monitoring

The DMEC chairperson was responsible for safeguarding the interests of the trial participants and making recommendations to the TSC. The REACT DMEC roles and responsibilities and operating procedures are defined in the REACT DMEC Charter (see Appendices 2 and 3). The DMEC was composed of three independent multidisciplinary experts who were not involved in the conduct of the trial in any way. They met prior to the initiation of enrolment and determined a plan to review the protocol, compliance, safety and adverse events (AEs), and outcome data after a prespecified number of infants had been recruited. The TSC was composed of six independent members, whose roles and responsibilities were defined by a charter. The TSC provided advice, through its chairperson, to the chief investigator and reported to the trial sponsor and trial funder.

Safety was assessed continuously during each infant’s stay in NICU. The frequency of AEs and SAEs that would normally require reporting within a clinical trial, as defined by the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use,42 was anticipated to be high in the population being studied, despite the low risk of the study intervention. Following discussions with the MHRA and in accordance with regulatory guidance that allows for exceptions in such circumstances, a modified reporting plan was agreed.

Any ADE was recorded and reported to the co-ordinating centre. All device deficiencies that might have led to a SADE if suitable action had not been taken, the intervention had not been made or if circumstances had been less fortunate were reported to the sponsor, as for SAEs and SADEs. AEs were recorded in the notes and captured as exploratory outcomes as part of the CRF. SAEs are known to be common in this population and, therefore, the MHRA requested the following expectations for safety reporting.

Efficacy

Secondary outcome analyses showed that time in the target range of 4–8 mmol/ was 12 percentage-points higher in the CGM arm than in the standard-care arm. The mean sensor glucose level, sensor glucose variability and percentage of time sensor glucose level was > 15 mmol/l were all lower in the CGM arm than in the standard-care arm, but none of these differences reached statistical significance. Results are provided with multiple testing procedure data (see Appendix 4, Table 27).

Safety

Continuous glucose monitor data showed that more infants in the standard-care arm (15%) than in the CGM arm (6%) were exposed to at least one episode of sensor glucose < 2.6 mmol/l for > 1 hour. Safety outcome data are provided in Table 7. AEs and ADEs are reported in Tables 8 and 9, respectively. The tables show that episodes of blood glucose levels ≤ 2.2 mmol/l and 2.2–2.6 mmol/l occurred in 7% and 12%, respectively, of the standard-care group, and in 13% and 15%, respectively, of the CGM group. Collection of the blood glucose data was biased by the result of a low sensor glucose level in the CGM arm, prompting blood glucose to be checked, and is reflected in the higher percentage of infants in the standard-care arm who were exposed to at least one episode of hypoglycaemia lasting > 1 hour. Owing to the small numbers and interinfant variability, none of these differences was statistically significant.

The mean percentage of time for which sensor glucose level was < 2.6 mmol/l was 1% (SD 3.2%) in the standard-care group and 1% (SD 5.3%) in the CGM group. The percentage of time that sensor glucose level was < 2.6 mmol/l depended on the number of CGM data available, which explains why it is similar in both groups, despite the total length of time that sensor glucose level was < 2.6 mmol/l being greater in the standard-care arm. Further details of SAEs and SADEs are provided in Tables 8 and 9, and details of withdrawals from the intervention or study, and deaths are shown in Appendix 4, Tables 28–30.

Utility

Table 10 shows summary statistics for the number of sensors used, the individual sensor duration and the number of CGM data collected (hours). We received CGM data from 155 infants, with loss of data due to reported failure of sensor insertion, calibration failure and lack of data at download. Insertion of the first sensor failed in 18% of infants and, therefore, the average number of sensors needed per infant was 1.2. Details of withdrawals are provided in Appendix 4, Tables 28 and 29. The success of sensor insertion varied markedly between centres. Difficulties in calibrating the sensor with the blood glucose level were reported in 22% of infants and episodes where data could not be linked between the transmitter and monitor occurred in 7% of infants. ADEs reported included bruising at the sensor site in 3% of infants, and significant discrepancy between sensor glucose and blood glucose in seven (4%) infants in the CGM study arm. In three infants, the sensor became displaced from under the skin (fell out).

The staff and parent questionnaire results are presented in Figures 8 and 9, respectively. Most staff found the devices easy to calibrate (86%) and did not think that the CGM device interfered with the infant’s nursing care (87%) or caused any distress (94%). Seventy per cent of staff felt that the use of CGM improved clinical care. The majority of parents reported that the CGM did not interfere with the infant’s nursing care (96%) or cause any distress (89%), and 80% felt that the use of CGM improved clinical care.

FIGURE 8.

Randomised controlled trial: staff questionnaires of utility. Stacked bar chart showing proportion of response for each Likert question. Reproduced with permission from Beardsall et al. 4 in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Beardsall et al. 4 in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

FIGURE 9.

Randomised controlled trial: parent questionnaire of acceptability. Stacked bar chart showing proportion of response for each Likert question. Reproduced with permission from Beardsall et al. 4 in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Beardsall et al. 4 in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Measurement of resource use and costs

Health economic data were collected on:

• resource use and costs associated with CGM (i.e. direct intervention costs)

• broader health service use during the 36-week follow-up period.

All costs were presented in Great British pounds and valued in 2016–17 prices. If necessary, costs were inflated or deflated to 2016–17 prices using the Hospital and Community Health Services Pay and Price Inflation Index. 45

Measuring broader resource use

The different types of resource categories for broader health-care costs and the sources of their unit costs are presented in Table 15.

Broader resource use data were collected for each infant using data extraction sheets that were completed by research nurses in each clinical centre. The nurses extracted the relevant data directly from routine hospital systems and patient records following the final follow-up assessments at 36 weeks’ corrected gestational age. The CRFs recorded details of length of inpatient stays by intensity of care, investigations performed [e.g. ultrasound scan, magnetic resonance imaging (MRI), electrocardiography (ECG), computed tomography (CT) scan, X-ray, cranial ultrasound scan and electroencephalography (EEG)], medications and procedures performed (e.g. lumbar punctures, ventricular taps, long lines, surgical procedures and interventions for ROP treatment). The CRFs also captured information surrounding transfers between hospitals, and the modes of transport used for hospital transfers and at final hospital discharge. Distances associated with transfers between hospitals were calculated using Google Maps (Google Inc., Mountain View, CA, USA).

Valuation of resource use

Unit cost data were mainly based on NHS reference costs. 47

Costs for sensor insertion, sensor monitoring and sensor removal were calculated by multiplying the duration of each activity, based on clinical advice, by the unit cost extracted from the NHS Agenda for Change pay scales (2016/17). 46 The costs of medications for infants were estimated based on the doses and frequencies reported on the CRF, with medication costs obtained from the British National Formulary. 48 The costs of blood transfusion products, such as platelets, red blood cells and frozen plasma, were obtained from NHS Blood and Transplant Price List 201749 and multiplied by the volume of each product used.

When infants were transferred to another hospital, the distance (miles) between the study site and transferred location was calculated using Google Maps and fuel costs were calculated using the advisory fuel rates obtained from the UK government. 50 The fuel costs applied to the distances of hospital transfers in clinical sites located outside the UK were estimated in the same manner.

Measurement of outcomes

The primary clinical outcome measure for this study was the percentage of time sensor glucose readings were within the target range of 2.6–10 mmol/l during the first 6 days of life. Cost-effectiveness was primarily expressed in terms of incremental cost per additional case of adequate glucose control during the first 6 days of life. For the primary cost-effectiveness analysis, adequate glucose control was defined as at least 80% of readings falling within the target range. However, to assess the robustness of the cost-effectiveness results, a number of sensitivity analyses were also performed that varied the threshold for number of readings falling within the target range of 2.6–10 mmol/l. Therefore, adequate glucose control was re-estimated at alternative thresholds of 60%, 70% and 90% of readings falling within the target range. Given the nature of the outcome measure used in this project and the methodological constraints surrounding utility measurement in newborn infants, cost–utility analysis using a preference-based measure, such as quality-adjusted life-years, was not performed.

The secondary clinical outcome measures selected for secondary expressions of cost-effectiveness were NEC and BPD. Cost-effectiveness analyses using these secondary clinical outcome measures were expressed using the incremental cost-effectiveness ratio (ICER), and expressed in terms of incremental cost per case of NEC averted and incremental cost per case of BPD averted, using imputed costs for the entire period extending to 36 weeks’ corrected gestation.

Cost-effectiveness analysis methods

Sensitivity and secondary outcomes analyses

Several sensitivity analyses were carried out to assess the robustness of the cost-effectiveness estimates. These were conducted under the following scenarios: (1) restricting the analyses to complete cases (i.e. those with complete cost and outcome data for the 7 days or up to 36 weeks’ corrected gestation); and (2) estimating the incremental cost-effectiveness using different thresholds of 60%, 70% and 90% for adequate glucose control. The same sensitivity analyses described above were repeated over alternative follow-up periods of the first 7 days of life or extending to 36 weeks’ corrected gestational age.

Two sets of prespecified subgroup analyses were conducted to explore the heterogeneity in the cost-effectiveness results. Subgroup analysis was performed for each study site and for two gestational age groups: extremely preterm (i.e. < 28 weeks’ gestation) and very preterm (i.e. 28–32 weeks’ gestation). A post hoc subgroup analysis was subsequently carried out using sex as an additional subgroup category. All subgroup analyses were based on cases with imputed cost and outcomes data at 7 days or 36 weeks’ corrected gestational age.

Health-care resource use

Table 16 shows resource use values for participants by trial group allocation and resource use category over the 7-day intervention period and the entire follow-up period for complete cases. The resource values are presented for subcategories of resource use, including hospital inpatient medications, investigation, medical treatment and surgical procedures. Use of Enlite sensors and glucose monitoring was reported for the 7-day intervention period only, as infants used the devices for this period only. Clinical recording of sensor glucose values was performed, on average, 121 times in infants in the CGM arm, but was not performed in infants in the standard-care arm, as the clinical team was masked to the sensor data.

Economic costs

Table 17 summarises the total NHS and PSS costs associated with resource use during days 1–7, day 8 to 36 weeks’ corrected gestational age and the entire study period for complete cases. Table 18 breaks these costs down by subcategory. The mean direct intervention costs over the first 7 days were £11,198 in the CGM arm and £9934 for the standard-care arm; the mean difference was not statistically significant at the 5% level. The cost of using the Enlite sensor and glucose monitoring was not included in the standard-care arm, as it is not part of standard care. The total cost during the period covering day 8 to 36 weeks’ corrected gestational age was higher in the standard-care arm (£61,065 vs. £65,464), but the difference was not statistically significant at the 5% level. The main driver for the higher cost in the standard-care arm during the follow-up period was the higher cost of neonatal care. Likewise, the mean total NHS and PSS cost throughout the entire follow-up period was higher in the standard-care group (£75,348) than in the CGM group (£71,909), but the mean between-group cost difference of –£3439 was not statistically significant at the 5% level.

Cost-effectiveness results

The cost-effectiveness results are presented in Tables 19 and 20 for days 1–7 and the entire study period, respectively. The associated CEACs and cost-effectiveness scatterplots are graphically represented in Figures 16–19.

FIGURE 16.

Randomised controlled trial: (a) CEAC and (b) scatterplot at intervention period for 80% threshold of adequate glucose control, imputed and covariate adjusted.

FIGURE 17.

Randomised controlled trial (a) CEAC and (b) scatterplot at intervention period for 80% threshold of adequate glucose control, complete case and covariate adjusted.

FIGURE 18.

Randomised controlled trial: (a) CEAC and (b) scatterplot over entire follow-up period for 80% threshold of adequate glucose control, imputed and covariate adjusted.

FIGURE 19.

Randomised controlled trial: (a) CEAC and (b) scatterplot at entire follow-up period over 80% threshold of adequate glucose control, complete case and covariate adjusted.

Study design

This was a single-centre feasibility study in infants recruited from the NICU at the Cambridge University Hospitals NHS Foundation Trust. All parents of eligible infants admitted to the unit within 48 hours of birth were approached (if a member of the research team was available to do so). The study applied a randomised open-label parallel design, with infants randomised to CGM alone supported by a guideline or to CGM with an additional intervention period of closed-loop CGM. Research Ethics Committee approval and informed written parental consent were obtained prior to study procedures. Eligibility criteria included birthweight ≤ 1200 g and aged ≤ 48 hours. Infants were excluded if they had a major congenital malformation or an underlying metabolic disorder, or if the mother’s pregnancy had been complicated by diabetes. Criteria were chosen based on previous data that identified these infants as at most risk of glucose dysregulation, while avoiding potential bias. 21

A real-time CGM device was inserted in all infants and remained in situ for up to 7 days. The guideline advised on the use of insulin or additional dextrose support. For a prespecified period of 24 hours, between 48 and 72 hours post birth, a closed-loop system controlled glucose in infants during the closed-loop intervention, whereas infants in the control group continued to receive CGM alongside the guideline. This window for closed-loop intervention was preselected to be a standard time from birth in all infants to allow time for informed parental consent to be obtained and because previous data (from masked CGMs) had shown this to be the time when hyperglycaemia is most prevalent. 21 Randomisation applied the minimisation methods using the Minim randomisation software (version 1, London School of Hygiene and Tropical Medicine, London, UK), with stratification according to gestational age and birthweight.

Common study procedures

With the exception of glucose control over the prespecified period, all other aspects of care were identical between treatment groups. Blood glucose monitoring was undertaken on the POC glucose StatStrip meter. Actrapid^® insulin (Novo Nordisk, Bagsværd, Denmark) at a concentration of 25 units/kg in 50 ml of 0.9% saline was used in both treatment groups. For CGM, an Enlite sensor was inserted by hand into the lateral thigh of each infant and linked to the Paradigm Veo for display of the sensor glucose concentration. Outside the intervention period (48–72 hours), CGM data were used in combination with the guideline by the clinical team to guide glucose control in all infants. Nurses calibrated the CGM device at least once every 12 hours, with a blood glucose measurement taken on the POC glucose StatStrip meter.

Paper guideline

The guideline for insulin delivery has previously been described. 1 It provided guidance based on the absolute sensor glucose value, as well as glucose trends. If sensor glucose levels are outside the target range, the guideline recommends review of clinical care and the dextrose infusion rate, and that modification of insulin delivery be considered, which could be instigated by the nurse at cotside. The insulin and dextrose were delivered by Alaris™ pumps (CareFusion, San Diego, CA, USA).

Closed-loop glucose control between 48 and 72 hours

The closed-loop system comprised (1) an Enlite sensor, (2) a laptop computer running a model predictive control algorithm and (3) two Alaris syringe pumps. We used a control algorithm based on the model predictive control approach,71,72 optimised and tuned in silico using a computer simulation environment validated for glucose control in the critically ill,73 aiming to keep sensor glucose levels between 4 and 8 mmol/l. The algorithm calculated insulin requirements or, at low glucose values, 20% dextrose infusion rates based on real-time sensor glucose values. A study nurse entered sensor glucose values into the laptop and modified insulin and dextrose pumps, as directed by the control algorithm, every 15 minutes. During the closed-loop intervention, Actrapid insulin in a concentration of 25 units/kg in 50 ml of 0.9% saline was used to allow for finer-dose titration in accordance with the algorithm. The algorithm calculations utilised a compartment model of glucose kinetics,74 describing the effect of insulin on sensor glucose excursions. The algorithm was initialised using an infant’s weight and adapted itself to a particular infant by updating two model parameters: (1) a rapidly changing glucose flux, correcting for errors in model-based predictions; and (2) a slowly changing estimate of an insulin rate to maintain normoglycaemia. The individualised model forecasted plasma glucose excursions over a 1.5-hour prediction horizon when calculating the insulin rate and a 30- to 40-minute horizon when calculating the dextrose rate. Information about enteral or parenteral nutrition was not provided to the algorithm. A reference blood glucose value was used every 6 hours for calibration of the glucose sensor. If sensor readings were not available because of sensor failure or loss of data capture, then hourly blood glucose levels were used to inform the algorithm for up to 4 hours. At this time, the algorithm continued to provide advice every 15 minutes.

Assessments and data collection

Demographic and clinical characteristics were collected with the following antenatal variables based on clinical diagnoses recorded in the maternal medical file: antenatal steroids (having received at least one dose prior to delivery), prolonged rupture of membranes (a rupture > 24 hours prior to delivery), hypertension and chorioamnionitis. 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 randomisation to the end of CGM.

Statistical analysis

The outcome measures and the statistical analysis plan were agreed in advance. The primary outcome was a comparison of the time that sensor glucose was in the glucose target range of 4–8 mmol/l between the study arms. Secondary outcomes were time spent with sensor glucose levels between 2.6 and 10 mmol/l, prevalence of hyperglycaemia (percentage of time sensor glucose is > 10 mmol/l) and mean (and SD) sensor glucose level. Safety end points included frequency of hypoglycaemia (any blood glucose levels < 2.6 mmol/l) and other AEs. Sample size was selected to reflect this study’s status as a pilot to ensure that we had comparable control data, while minimising patient exposure. Formal power calculations were not performed and all analyses were based on intention to treat. An independent t-test was used to compare normally distributed variables. Non-normally distributed variables were compared using a Mann–Whitney U-test. Calculations were carried out using IBM SPSS Statistics version 23 (IBM Corporation, Armonk, NY, USA). Values are given as mean (SD) or median (IQR). p-values < 0.05 were considered statistically significant.

Glucose control and insulin and dextrose administration between 48 and 72 hours

There was no difference between the study groups in the baseline (48 hours) mean sensor glucose level (Table 22). During the intervention period, the median time spent in the target range of sensor glucose level (i.e. 4–8 mmol/l) was significantly higher in the closed-loop arm (91%, IQR 78–99%) than in the control arm (26%, IQR 6–64%) (p < 0.001). This was equivalent to 22 (IQR 19–24) hours in the closed-loop group, compared with 6 (IQR 1–15) hours in the control group. Similarly, the time spent in the wider target range of 2.6–10 mmol/l was higher in the closed-loop group (median 100% vs. median 84%). This was because the median percentage of time with sensor glucose values > 10 mmol/l was lower in the control arm than in the closed-loop arm (16% compared with 0%). There was no difference in the time spent with sensor glucose levels < 2.6 mmol/l. The median sensor glucose level was lower in the closed-loop arm (6.2 mmol/l, IQR 6.1–7.1 mmol/l) than in the control arm (8.6 mmol/l, IQR 7.4–11.1 mmol/l) (p = 0.002). Glucose variability, as measured by the SD of sensor glucose, was similar (p = 0.6).

Summative glucose profiles and insulin delivery are provided in Figure 20. Nine out of the 10 infants in the closed-loop group had received insulin prior to the 24-hour intervention, with the one remaining infant being started on insulin during the 24-hour closed-loop period. For comparison, four infants in the control arm received insulin prior to the intervention period and eight infants received insulin during the intervention period. Four infants in the closed-loop arm received additional 20% dextrose for short periods during the intervention period (up to 3.5 hours). The mean infusion rate in these infants ranged from 0.13 to 0.53 ml/kg/hour, with the highest rate being infused in an infant who was hypoglycaemic 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 the study groups during the 24-hour intervention period.

FIGURE 20.

Closed-loop study: glucose control and insulin delivery. 3 Median (IQR) sensor glucose level and insulin infusion rate in infants randomised to closed-loop management or CGM with paper guideline (control). The closed-loop intervention period is denoted by the vertical lines and the glucose target range of 4–8 mmol/l is denoted by horizontal lines. The shaded area represents the IQR of sensor glucose values and insulin infusion rates, respectively. Reproduced with permission from Beardsall et al. 3 This is an open access article distributed in accordance with the Creative Commons Attribution (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Beardsall et al. 3 This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Glucose control and 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 (i.e. 4–8 mmol/l and 2.6–10 mmol/l) in the closed-loop arm compared with the control arm (Table 23), but these differences did not reach statistical significance.

Nutrition and clinical care

All infants received parenteral and enteral nutrition in accordance with the standard local neonatal unit protocol. During the closed-loop intervention period, between 48 and 72 hours, four infants in each study group were receiving minimal amounts of trophic feeds. In the post-intervention period, after 72 hours, there was no difference between the two study groups in the volume of milk received (see Table 23). None of the infants received hydrocortisone, but seven infants in the control arm and two infants in the closed-loop arm received inotropes during the first week. During the intervention period, the median number of blood glucose values taken per day was 5.5 (IQR 4.8–6.6) in the control arm and 6.3 (IQR 5.1–6.5) in the closed-loop arm. Over the whole study period, the median number of blood glucose values taken per day was 5.1 (IQR 4.3–5.4) in the control arm and 5.3 (IQR 4.6–5.8) in the closed-loop arm.

Safety

No reported concerns about the sensor site, in terms of inflammation or infection, were reported either during the study or after removal of the sensor. Two infants in the closed-loop arm experienced documented episodes of hypoglycaemia, with a blood glucose level < 2.6 mmol/l. One episode occurred when checking the baseline blood glucose prior to the start of the closed-loop intervention. At that time, maintenance fluids were being changed, but no insulin was being infused. The second episode was on day 6 and, again, this was associated with a change of maintenance fluids. A further two infants experienced periods (after the 72-hour closed loop) when sensor glucose levels fell below 2.6 mmol/l, but the blood glucose level at that time was documented to be above 2.6 mmol/l. In the control arm, no infants had a documented blood glucose value of < 2.6 mmol/l. One infant in the control arm experienced an episode lasting 205 minutes when the sensor glucose fell to < 2.6 mmol/l (blood glucose level was not checked at this time). None of the infants was on insulin and there was no clinical evidence of hypoglycaemia in these infants during any of these episodes.

Continuous glucose monitoring versus point-of-care blood glucose

Considerable controversy remains as to the best way to manage neonatal glucose control, primarily because of a lack of good evidence on which to base clinical practice. 81,82 Measurement of glucose levels has evolved greatly over the last 40 years, with CGM being introduced in the last 20 years. The increasingly rapid innovations in technology could open up the opportunity for better understanding and management of the many infants who present with disordered glucose levels.

Current clinical care depends on a spectrum of medical devices and a range of clinical practices regarding the frequency of glucose sampling. Standard care includes capillary, venous or arterial blood sampling, and POC devices that have the advantage of rapid results and small samples compared with the ‘gold standard’ of laboratory analyses, which can result in a significant delay before results are available. 83 CGM could provide consistency in terms of interstitial levels measured continuously in real time. Over the course of our studies, the regulatory landscape for CGM models has changed during the time of these studies from an adjunct to care (requiring confirmatory blood glucose levels) to one where some CGM models are approved as being accurate enough to be a standalone guide to insulin dosing in patients with diabetes. 84

The fact that CGM measures not blood but interstitial glucose levels has given rise to some concern about its ‘validity’. Even when CGM devices are calibrated to blood glucose levels, there can be differences, particularly in certain circumstances. 85 However, it can be argued that interstitial glucose levels provide much more useful physiological and, therefore, clinical information, as cellular glucose levels may be particularly relevant to neuroglycopenia and the developing brain. 85

The acknowledged lag of up to 15 minutes in sensor glucose level compared with blood glucose levels needs to be placed in a clinical context. Although there may be a short lag of 15 minutes using CGM, the system still provides a continuous trend in real time. In contrast, the current standard of care using intermittent blood sampling can result in a delay of many hours between levels being checked. CGM can therefore reveal significant periods of clinically silent hypoglycaemia. 86,87 This was highlighted in our studies, in which falling sensor values highlighted the need for blood glucose levels to be taken. 1 However, there were no SAEs related to a ‘falsely’ reassuring sensor glucose level during an episode of hypoglycaemia. The detection of clinically silent hypoglycaemic episodes may be the most useful impact of CGM, highlighted by the association of such episodes with deficits in attention and visual acuity at 4.5 years. 79 Although, theoretically, this could lead to overmedicalisation, the potential to prevent harm should not be missed.

The unbiased nature of the CGM data also has the potential to help us better understand the physiology and pathology of changes in glucose levels. However, CGM devices are not currently licensed for use in the neonate and are not designed for ease of use with these infants. The next steps are to engage with the industry to address issues specific to this population.

The use of paper and computerised algorithms

This programme of studies involved the use of CGM technology combined with a variety of algorithms, both paper and computer based. The interventions should therefore be considered ‘combined interventions’.

We purposefully designed the paper algorithm as a guideline and not as a protocol. Although a ‘protocolised’ approach to ensure compliance may be considered optimal within a clinical trial, it was clear that the complexity needed for such a paper protocol would have increased risk. This was based on our previous experience, which highlighted wide variation in drug infusions, clinical complications and interventions across multiple sites. During the feasibility study, the algorithm evolved, as early iterations that aimed to avoid hypoglycaemia resulted in infusions being stopped and restarted more often than was deemed optimal. The advice was simplified for use by nursing staff, without requiring formal changes in prescriptions (similar to following a sliding scale). This was important as it provided a safe framework for the nursing staff to be able to titrate insulin infusion rates based on sensor glucose trends over short periods of time and nursing staff took ownership of decision-making, with a more comprehensive approach to care.

The ‘closed-loop’ study was entirely exploratory, but dramatically demonstrated the potential impact of combing the wealth of data from CGM with a computer-based algorithm. It provided a more personalised approach to care. A fully automated system has the potential to optimise clinical care by targeting glucose control and reducing workload. However, there are significant challenges to realisation, including CGM technology designed for the preterm infant, manufacturing engagement with open access to sensor and pump data, and regulatory hurdles. However, this route clearly could have significant impact for infants and staff.

Device trials and future regulations

Large multicentre device trials are uncommon compared with drug trials, and very limited clinical data are required for devices to obtain CE marking. This is particularly relevant for the neonatal population, in whom the unique risks and benefits may not have been fully assessed.

This trial highlighted some of the challenges of undertaking device trials, highlighting both practical clinical complexities and the importance of working closely with the company, with complete transparency about developments in the technology during the study.

Prior to starting the feasibility study, the sensor technology had been redesigned from the Medtronic plc Sof-Sensor^® to the Medtronic plc Enlite, with a change in the angle of insertion. This was an advance for the adult market, but made sensor insertion more challenging in our preterm infants, who have limited subcutaneous tissue. It was necessary to modify procedures and study guidance. Further modifications were made to the Enlite sensor (from Enlite to Enlite plus) during the studies to improve the accuracy of the sensor technology.

On completion of the feasibility study, a new monitor was due to be released to supersede the Paradigm Veo. The advantages included better accuracy at low glucose levels and a clearer user interface. Therefore, after consideration and discussion with the TSC about risk and benefits, a decision was made to use the latest model. However, it became apparent that changes to the algorithm were having an impact on the calibration frequency and lockout of sensors which then needed to be replaced. Discussions with the company revealed that changes made to the algorithm to provide better accuracy and safety led to increased error readings within our population. There were further modifications to the algorithm to address some of these issues during the study. These potentially subtle changes from a manufacturing perspective had a significant impact on our sensor failure rate at the beginning of the study, but could not have been anticipated. This shows that the rapid advances in the technology present clinical challenges and it is imperative that manufacturers and clinicians work together with complete transparency.

Clinical aspects

Training the large number of staff within a busy intensive care unit to use devices as part of a clinical trial requires significant resources, both financial and time. In our studies, the short recruitment window increased this challenge for recruiting sites, and therefore the provision of on-site training and an online training video, as well as a 24-hour helpline, was important. It was clear that the level of support required for an interventional device trial was greater than that for an investigational medicinal product study.

Addenbrooke’s Hospital

A Ogilvy-Stuart and L Thomson.

Hospital Sant Joan de Déu (Barcelona)

I Iglesias.

John Radcliffe Hospital

Z Molnar, S Barlow and S Baugh.

Leeds General Infirmary

K Johnson, L Uryn and C Spencer.

Leicester Royal Infirmary

M Hubbard, J Hamilton, TC Kwok and S Mittal.

Luton and Dunstable University Hospital

S Somisetty, O Adesiyan, J Kapadia and Y Millar.

New Cross Hospital

K Gurusamy, L Bibb and K Jones.

Norfolk and Norwich University Hospital

P Muthukumar, A Nichols and all of the nurses and the medical team at the Norfolk and Norwich University Hospital NICU.

Princess Anne Hospital

M Johnson, J Pond, P Crowley, C Mellish and H Dewar.

The Royal London Hospital

D Shah, P Vincent, A Iringan, M Abraham, S Anilkumar, B Aninakwa and RA Dalangin-Chalmers.

St Michael’s Hospital

A Collins.

St Peter’s Hospital

P Reynolds, N Holland and K Wells.

Vrije Universiteit University Medical Center: Amsterdam University Medical Center

M van Weissenbruch and A de Lange.

National Institute for Health Research Evaluation, Trials and Studies Coordinating Centre

R Bloom, C Dukes and S Hammerson.

UK Clinical Research Network