The present study showed that a subcutaneous CGM system to guide blood glucose regulation was equally effective and safe in glycemic control compared to frequent POC-guided blood glucose regulation. However, CGM significantly reduces nursing workload, blood loss and the daily costs for glucose control.
Comparison with other studies
This is the second but largest randomized controlled trial in which CGM is used to guide glycemic control in critically ill patients. In contrast to our findings, Holzinger and colleagues did find less severe hypoglycemia in the CGM group . This may be caused by the very low incidence of severe hypoglycemia in the present study, which was true for both the intervention and the control group. This may be related to a change of policy after the publication of the NICE-SUGAR trial , which was a reason for our and most other ICUs to increase their blood glucose target range. The increased target range may have reduced the incidence of hypoglycemic events [19,20]. Indeed, the blood glucose target used in the current study (5.0 to 9.0 mmol/L) was higher than in the Holzinger trial  (4.4 to 6.1 mmol/L) and this is reflected in the achieved mean blood glucose levels (8.1 vs. 6.3 mmol/L). Moreover, the use of a fully computerized algorithm for glucose control and the high familiarity of the protocol among our IC nurses may have contributed to the low incidence of severe hypoglycemia. The available studies to date on tight glucose control showed an increase in nursing workload [21-23]. The potential benefits of CGM in the reduction of blood samples, blood loss and nursing workload was assumed in previous studies, but was not systematically assessed before. We now observed that CGM significantly reduced the amount of blood samples and the daily nursing workload for glucose control up to 53%. This finding seems clinically relevant, especially in a busy clinical IC environment. Two studies focused on the cumulative nursing workload accompanied with tight glucose control protocols [21,22]. Gartemann et al. estimated that nurses devoted approximately 42 minutes during a 12-hour shift of their time to administering a tight glycemic control (TGC) protocol, whereas Aragon et al. even reported that up to 2 hours might be required for tight glycemic control for a single patient in a 24-hour period. In our POC control group, the mean nursing workload estimate was less (36 minutes per 24 hours) than the published estimates reported by other groups. This might partly be explained by the use of a fully computerized algorithm for glucose control in our ICU. In addition, the familiarity of the protocol is very high among our ICU nurses.
Effectiveness and costs
The use of CGM did not achieve improved glycemic control in our study. We found similar percentages of time-in-target and below-target range between the study groups. The not-significantly lower percentage of time in the hyperglycemic range in the intervention group could be explained by the fact that CGM measurements were more frequently entered in the glucose protocol than POC measurements in the control group. This probably resulted in more adjustments in the insulin treatment with lower blood glucose levels as a consequence. The significantly increased ICU LOS, which was observed in the intervention group, may be a coincidence or reflect unmeasured case-mix factors but is, in our view, unrelated to the glucose measurement strategy.
In contrast to our expectations, the cost analysis shows that the use of CGM systems for glucose control in an ICU setting is not a priori an expense. However, we should be cautious in interpreting these results due to the rather short time horizon (24 hours) in the analysis of costs determination and the single-centre study design. Also, cost savings cannot immediately be monetized due to the short time horizon used in this cost analysis.
Accuracy of the subcutaneous measurements
The subcutaneous Freestyle Navigator CGM device that we used in the present study showed a median RAD of 13.7%, which is higher than the 10.6 and 11.6% that was found in previous validation studies of this device in critically ill patients, suggesting an accuracy acceptable for clinical use. [11,14]. The lag time that may be needed for the subcutaneous compartment to adapt to the intravenous compartment appeared not to be clinically relevant . However, the accuracy as assessed in the current study seems to indicate a need for improvement, because the accuracy was less than the accuracy of the Accu-Chek and because a substantial number (75% in the CGM group and 33% in the control group) of hypoglycemic events was not detected. Of note, Leelarathna et al.  recently investigated whether there was a difference in accuracy of the Freestyle Navigator in a critical care setting using two methods of calibration: (1) calibration according to the manufacturer’s instructions (1, 2, 10, and 24 h) or (2) calibration at variable intervals of 1 to 6 h using ABG. Using enhanced calibration, at a median (interquartile range) every 169 (122 to 213) minutes, the absolute relative deviation was lower (7.0% (3.5, 13.0) vs. 12.8% (6.3, 21.8), P <0.001). So, further significant improvements in accuracy may be obtained by frequent calibrations with ABG measurements. In the current study forced calibration was not possible, calibration was only performed when the CGM device indicated the need for calibration by itself.
In addition, technical problems with the subcutaneous CGM device were observed during the study and led to a 12% dropout. The most important reason was the temporary loss of sensor signal from several minutes to hours that resulted in a loss of data. Difficulties in the calibration process were also identified as the CGM could only be calibrated if the system indicated a calibration by itself, which occurred for median 1.9 times per 24 hours. Most of the technical difficulties, however, may have been due to lack of experience working with the CGM device despite the training of all ICU nurses. We expect such problems to be easily resolved with additional training and with the improved next generation Freestyle Navigator II, which has recently been introduced and showed good utility and sensor performance in critically ill patients . This study aimed to define safety, efficacy and costs and therefore we neglected the system dropout at this moment. It is true, however, that this device can only become part of routine care when the dropout percentage diminishes.
Strengths and weaknesses
The strengths of our study include the relatively large sample size, the randomised controlled study design and the wide variety in case mix. However, some limitations of the present study merit further consideration. First, the study was performed in a single Dutch intensive care unit, which limits the generalizability of the study. Second, the study was designed to blind the values of the CGM in the control group. However, the CGM needed to be calibrated several times during the study period, which made it impossible to blind it completely. Third, the nursing staff did not verify the severe hypoglycemia that was indicated by CGM in two of the three patients despite specific instructions to do so. One of these two patients had evolved into a ‘withholding care policy’, which was the reason to accept the severe hypoglycemia. We assume that in the other patient priority was given to other important nursing tasks. Thus, the available data are insufficient to define the accuracy of the CGM in the hypoglycemic range. In our previous studies this was not identified as a clinical problem [11,14]. Also, with an adapted algorithm, the CGM should be able to detect a decreasing glucose level before hypoglycemia is present and give a timely alert. Fourth, the computerized algorithm was designed for intermittent POC measurements and not for (semi-) continuous data. As such, the patients did not fully benefit from the frequent glucose measurements by CGM. An algorithm based on 10-minute glucose input might have led to other results. We did identify this issue beforehand but we decided to keep the algorithm for both groups the same to be able to investigate the contribution of CGM per se. It can be expected that an adapted algorithm will further improve the performance of CGM in the guidance of glycemic control.