Study design and population
This investigator-initiated prospective study was conducted in the mixed medical-surgical ICU of the Academic Medical Center, Amsterdam, The Netherlands. The Institutional Review Board of the Academic Medical Center approved the study protocol, and informed consent was obtained from all patients or their legal representatives before start of the study. Maquet Critical Care AB provided the CGM device and its disposables free of charge. Maquet Critical Care AB had influence neither on the design of this study nor on reporting of the results. The study was registered at the Netherlands Trial Register (NTR4527).
Patients were eligible for participation if they were at least 18 years old, were expected to stay in the ICU for ≥48 h, had an arterial catheter in place and were in need of a (new) CVC. Patients were excluded if they participated in another investigational drug or device study or were known to be pregnant.
Blood glucose control
ICU nurses followed a local guideline aiming at a blood glucose level between 90 and 144 mg/dL (5–8 mmol/L) as part of standard care. This guideline mandated nurses to measure blood glucose every 4 h, or more frequently when glucose levels were out of range or when rapid changes were expected. Infusion of insulin was started when glucose levels were over 144 mg/dL and stopped when glucose was lower than 61 mg/dL. Adjustments of insulin titration were based on sliding scales. More details can be found in Additional file 1. In addition, details on how nurses were trained can also be found in Additional file 1.
During the study, ICU nurses were not allowed to change insulin infusion rate based on the readings by the device. The ICU nurses, however, had access to device readings and additional arterial blood glucose measurements were allowed if the device suggested rapid changes in the glucose level or when there was a trend toward hypoglycemia. In addition, the ICU nurses could also adjust insulin infusion rates based on reference blood glucose values obtained during study observation periods (see below).
The study device
For intravenous microdialysis-based glucose monitoring, a special CVC with a semipermeable membrane (Maquet Critical Care AB, Solna, Sweden) is needed. This CVC has five lumens, three ‘normal’ ports for intravenous administration of fluids or medication and two ‘special’ ports for transport of normal saline alongside the semipermeable membrane, which should not be flushed and cannot be used for intravenous administration of fluids or medication. The ‘afferent’ port is connected to a saline-filled syringe placed in the syringe pump of the device. The ‘efferent’ port is connected to the disposable sensor. Small metabolites such as glucose pass through the semipermeable membrane creating equilibrium between blood and the dialysate. The dialysate is pumped over the sensor in a continuous fashion, where the glucose oxidase method is used to measure the glucose level [4, 5]. The device can be used for a maximum of 96 h per sensor. Reference measurements are needed for calibration of the device, which is performed at start-up and every 8 h thereafter.
Of note, because the dialysate needs to be transported to the sensor outside the patient, where measurements are performed, there is a delay in time of 5 min between dialysate formation and the actual measurements.
Study procedures
In three blocks of 8 h per day, and for a maximum of 3 days, every 15 min an arterial blood sample of 200 µL was drawn through an existing arterial line. Blood glucose levels were measured using a blood gas analyzer (RAPIDLab 1265, Siemens Healthcare Diagnostics, The Hague, The Netherlands).
Definitions of the metrics used to assess device reliability, including those suggested by recent consensus recommendations [8], are described in detail in Additional file 1 and included the percentage of real-time data, skips in data acquisition, failures to calibrate, sensor failures and CVC failures.
Power calculation
Based on previous studies [9, 10], we chose to collect approximately 1000 paired measurements or to connect the device to a minimum number of 11 patients. Inclusion of patients was restricted by the time the device was available for this study and the number of disposable CVCs and sensors provided by the manufacturer.
Analysis plan
Patient characteristics were reported as means, medians or percentages, where appropriate. Because of the delay between dialysate formation and the actual measurements of the blood glucose level, we subtracted 5 min from the time stamp of the values of the CGM device; as such reference blood glucose values matched with the moment dialysate was formed. Subsequently, device and reference measurements were merged. Paired measurements were used for determining point and trend accuracy of the device. To be considered for the statistical analysis, each patient needed to have at least multiple samples with at most 30 min in between. However, patients excluded for statistical analysis remained included in the reliability analysis. While each paired sample was included in the point accuracy analysis, only the samples with a gap of at most 30 min to the next sample were included in the trend accuracy analysis. In addition, when the device was calibrated within the daily 8-h block of intense sampling, the calibration sample and the subsequent sample were not considered for trend accuracy analysis. This way, large changes in trend due to the calibration were excluded from the analysis.
Point accuracy was expressed using a Clarke error grid, a Bland–Altman plot, the glucose prediction error analysis and the mean absolute relative difference (MARD). To be considered point accurate, at least 95 % of values must be in zone A, a maximum of 5 % can be in zone B, and no values are allowed in zones C to E of the Clarke error grid [11]. Also, the MARD should be below 14 %; a value above 18 % represent poor accuracy [3].
Trend accuracy was expressed using rate error grid analysis (R-EGA) [12]. Values outside zones A and B of the R-EGA corresponding to values in zones A and B of the Clarke error grid were considered benign errors. On the other hand, values outside zones A and B of the R-EGA corresponding to values outside zones A and B of the Clarke error grid were considered erroneous readings [12].
Post hoc analysis
Point accuracy was also expressed using the recently published surveillance error grid [13].
Two of the CVCs were malfunctioning. In one case, it was immediately clear that the CVC was defect, and no additional measurements were performed. In another case, this was not immediately clear, and only after reviewing the readings it became clear that the CVC started to malfunction from a certain time point. We chose to perform a post hoc analysis excluding the data from that patient.