Skip to main content
Diabetes Technology & Therapeutics logoLink to Diabetes Technology & Therapeutics
. 2023 Nov 7;25(11):817–821. doi: 10.1089/dia.2023.0287

Testing the Real-World Accuracy of the Dexcom G6 Pro CGM During the Insulin-Only Bionic Pancreas Pivotal Trial

Martin Chase Marak 1,, Peter Calhoun 1, Edward R Damiano 2, Steven J Russell 2, Katrina J Ruedy 1, Roy W Beck 1
PMCID: PMC10771867  PMID: 37668666

Abstract

Continuous glucose monitors (CGMs) have transformed the way people with type 1 diabetes can self-monitor glucose levels. Past studies have evaluated the accuracy of CGMs in clinic-based studies, but few have analyzed their accuracy in real-world settings. The Insulin-Only Bionic Pancreas Trial provided the opportunity to assess real-world accuracy of the blinded Dexcom G6 Pro sensor over the first 48–60 h of wear using a blood glucose meter (BGM) as a comparator for 1073 CGM–BGM pairs across 53 participants. The mean absolute relative difference (MARD) was 11.0% over a median period of 50 h (range 47–79 h). The MARD was 13.6% in the first 12 h, 10.5% in hours 12–24, and 10.1% after the first 24 h. These results are comparable with accuracy shown previously with laboratory-based measurements and provide real-world evidence of Dexcom G6 Pro accuracy, which improved after the first 12 h and then remained stable thereafter.

Clinical Trial Registry: clinicaltrials.gov; NCT04200313.

Keywords: Continuous glucose monitors, Accuracy, Dexcom, Type 1 diabetes, Blood glucose

Background

Continuous glucose monitors (CGMs) have been used increasingly by individuals with type 1 diabetes (T1D) and type 2 diabetes to achieve good glycemic control. The accuracy of CGMs has generally improved over the years when evaluated in a laboratory setting with a highly accurate glucose-measuring instrument,1 but the accuracy has generally been found to be lower in the real-world setting when point-of-care blood glucose meters (BGMs) were used for reference.

For example, Welsh et al. compared the Dexcom G5, G6, and G7 and found mean absolute relative difference (MARD) of 9.0%, 9.9%, and 9.1%, respectively, when evaluated in the laboratory setting,2 but the MARD of the G5 when measured at home has been reported to be 11.4%3 and 16.3%.4 The higher MARD at home could be partially explained by BGMs being less accurate than the laboratory-grade measuring instrument. However, the CGM accuracy at home could also better reflect the true accuracy when individuals are active, and glucose is rapidly changing. Thus, understanding the accuracy of the Dexcom G6 Pro when used in clinical care is important.

One challenge with evaluating the CGM accuracy at home is that individuals may choose to use the BGM when they believe their CGM is less accurate. For example, G6 users are advised to collect a BGM measurement when CGM alerts and readings do not match their symptoms or expectations. Thus, this could cause a biased accuracy assessment as inaccurate CGM readings may be sampled more often than accurate CGM readings. At the end of the randomized Insulin-Only Bionic Pancreas Pivotal Trial,5 a subset of participants used the iLet® Bionic Pancreas System (Beta Bionics, Inc.) for ∼48–60 h with intermittent BGM input rather than CGM input.6

During this time period, a blinded G6 Pro sensor was worn (i.e., participant masked to their CGM readings). Because the sensor was blinded, the data collected during the substudy provided the opportunity to evaluate the accuracy of the G6 Pro sensor in comparison with frequent BGM measurements without the aforementioned potential sampling bias.

Methods

The institutional review board-approved protocol for the randomized trial (available at https://www.jaeb.org/finaliobp) describes the methods of the substudy, which followed the randomized trial in which the bionic pancreas received BGM input without CGM input for 48–60 h. The methods are briefly described below.

Separate written informed consent, or parental consent and patient assent for children, was obtained for participation in the substudy. Participants were provided with a Contour® Next One Glucometer and instructed to measure their blood glucose approximately every 2 h during waking hours and at least once during each overnight period for 48–60 h. In addition, participants entered these blood glucose values into the iLet bionic pancreas, which set a reminder alarm to trigger 2 h after entry. Participants were also encouraged to measure their blood glucose before meals, ∼2 h after meals, before bedtime, and before exercise.

During this period, a blinded Dexcom G6 Pro sensor (Dexcom, Inc.) was worn that did not communicate with the iLet bionic pancreas. All participants initiated the substudy with a new blinded sensor and wore this sensor throughout the observation period.

For analysis, BGM measurements were paired to the CGM measurement that was closest in time within a range of 5 min. The following metrics were pooled for all pairs: mean bias (difference of CGM and BGM), MARD, and median absolute relative difference (ARD), and proportion of CGM values within 15 and 20 mg/dL of BGM values ≤100 mg/dL or within 15% and 20% of BGM values >100 mg/dL (%15/15 ISO and %20/20 ISO, respectively).

Results were tabulated overall, by glucose rate of change within 15 min (<−2, −2 to <−1, −1 to 1, >1 to 2, and >2 mg/[dL·min]), by BGM values (<70, 70 to <120, 120–180, or >180 mg/dL), and by sensor wear period (<12, 12 to <24, and ≥24 h). In addition, the MARD was also calculated for each individual sensor.

Results

Data from 53 participants with T1D were included in the analyses. Participants' age at the start of the randomized trial ranged from 7 to 70 years, 41 (77%) were ≥18 years old and 12 (23%) were <18 years old, and mean T1D duration was 21 ± 14 years. Mean HbA1c at the start of the substudy was 7.1% ± 0.6%, down from 8.0% ± 1.0% before initiation of iLet therapy 13 weeks earlier.

The substudy lasted a median of 50 h (range 47–79 h, interquartile range [IQR] 48 − 52 h). Participants performed a median of 19 paired BGM measurements (IQR 17–23).

A total of 1073 BGM measurements were paired to a blinded Dexcom G6 Pro CGM measurement within 5 min. The paired BGM and CGM readings generally were closely aligned, with a slight bias in which the CGM values tended to be slightly higher than the BGM values over the whole range of glucose levels (Fig. 1). The overall mean bias was +4 mg/dL, with a bias of +6 mg/dL in the first 12 h, +6 mg/dL in 12–24 h, and +3 mg/dL after 24 h (Table 1). The overall MARD was 11.0% with a median ARD of 8.6%. The MARD was 13.6% during the first 12 h, 10.5% during 12–24 h, and 10.1% after the first 24 h.

FIG. 1.

FIG. 1.

Scatterplot of CGM and BGM measurements. Each dot represents one CGM–BGM pair (N = 1073). The red line represents a mixed linear regression line of CGM values onto BGM values with a spatial power covariance structure to handle correlated paired measurements across time for each sensor. BGM, blood glucose meter; CGM, continuous glucose monitor.

Table 1.

Accuracy of Dexcom G6 Pro CGM When Compared with BGM Measurements

  No. of subjects No. of CGM-BGM pairs Mean bias (mg/dL)a MARD (%)b Median ARD (%)b %15/15c %20/20c
Entire sensor wear
 Overall 53 1073 4 11.0 8.6 78.7 87.2
 BGM reading (mg/dL)
  <70 32 61 6 16.5 12.1 80.3 88.5
  70 to <120 51 217 5 13.7 9.7 73.3 82.9
  120 to 180 52 335 4 10.0 8.0 80.0 88.7
  >180 53 460 4 9.7 8.1 80.0 88.0
 CGM reading rate of change (mg/[dL·min])d  
  <−2 40 60 6 13.4 9.6 76.7 80.0
  −2 to <−1 48 123 4 10.9 8.7 82.9 88.6
  −1 to 1 53 729 4 10.6 8.0 79.3 87.9
  >1 to 2 43 99 6 12.5 10.8 70.7 82.8
  >2 35 59 9 11.0 7.9 78.0 91.5
<12 h
 Overall 53 250 6 13.6 9.8 73.6 80.8
 BGM reading (mg/dL)
  <70 18 18 6 16.5 15.8 83.3 94.4
  70 to <120 25 39 9 24.3 16.5 51.3 64.1
  120 to 180 36 74 –1 13.0 9.6 75.7 82.4
  >180 48 119 9 10.0 7.8 78.2 83.2
12 to <24 h
 Overall 53 251 6 10.5 8.2 79.7 89.2
 BGM reading (mg/dL)
  <70 10 12 7 18.3 13.9 83.3 83.3
  70 to <120 35 49 7 13.3 9.3 71.4 83.7
  120 to 180 43 90 7 9.8 9.1 80.0 90.0
  >180 45 100 5 8.9 7.6 83.0 92.0
≥24 h
 Overall 53 572 3 10.1 7.9 80.4 89.2
 BGM reading (mg/dL)
  <70 21 31 6 15.8 10.4 77.4 87.1
  70 to <120 47 129 3 10.7 8.0 80.6 88.4
  120 to 180 48 171 5 8.9 7.4 81.9 90.6
  >180 53 241 2 9.9 8.7 79.7 88.8
a

Bias = CGM – BGM.

b

ARD = abs(CGM – BGM)/BGM. Mean of the ARD is referred to as MARD.

c

%15/15 is defined as a CGM value within ±15 mg/dL when BGM value ≤100 mg/dL and CGM value within ±15% when BGM value >100 mg/dL. %20/20 is defined as a CGM value within ±20 mg/dL when BGM value ≤100 mg/dL and CGM value within ±20% when BGM value >100 mg/dL.

d

Three CGM measurements have missing rate of change due to no recent prior CGM measurement.

ARD, absolute relative difference; BGM, blood glucose meter; CGM, continuous glucose monitor; MARD, mean absolute relative difference.

The %20/20 was 87.2% overall, 80.8% during the first 12 h, 89.2% during 12–24 h, and 89.2% after the first 24 h. During the first 24 h of sensor wear, the MARD was lowest for BGM values in the 120–180 mg/dL range. After the first 24 h of sensor wear, the MARD was generally lowest for high BGM values, whereas the %20/20 was relatively stable (∼89%) across all BGM values. The MARD was also the lowest when the rate of change of CGM values was between −1 and 1 mg/(dL·min) (i.e., CGM values were changing most slowly) and the MARD was the highest when the rate of change of CGM values was less than −2 mg/(dL·min) (i.e., CGM values were falling rapidly).

Figure 2 provides a histogram of the MARD values for each individual sensor throughout the study period. There were 8 (15%) sensors with a MARD ≥15.0%. For these less accurate sensors, the MARD was 22.7% during the first 12 h, 19.5% during 12–24 h, and 16.2% after the first 24 h, and the overall mean bias was +20 mg/dL, with a bias of +23 mg/dL in the first 12 h, +18 mg/dL in 12–24 h, and +19 mg/dL after 24 h.

FIG. 2.

FIG. 2.

Histogram of MARD for each sensor. Each value represents an individual sensor (N = 53). Mean MARD = 11.0%, median MARD = 9.3%, SD MARD = 5.3%. MARD, mean absolute relative difference; SD, standard deviation.

Discussion

This study provided the opportunity to assess G6 accuracy in a blinded manner and without sampling biases that may occur in accuracy studies of unblinded sensors. CGM accuracy at home was high, with a MARD of 11.0% when compared with BGM measurements over a median period of 50 h. These results show that, in a typical real-world setting, participants can still expect a high degree of CGM accuracy with the Dexcom G6 Pro and correspondingly the Dexcom G6 on the first day and beyond.

The Dexcom G6 Pro and G6 sensors are identical; however, they differ with respect to feature sets and signal processing, in that the former is blinded and does not allow for manual calibrations, whereas the latter is unblinded and can be manually calibrated. Dexcom G6 accuracy is generally the lowest in the first 24 h7 and then improves. The accuracy of the Dexcom G6 was reported to have a MARD of 11.5% on sensor wear day 1 and 10.5% on sensor wear day 2 when compared with a YSI (Yellow Springs Instrument) in a laboratory setting.7 Thus, our findings showed similar accuracy of the G6 Pro when compared with BGM measurements at home.

This was surprising as other Dexcom CGM studies have shown lower accuracy when using BGM as a reference in various settings.3,4 Reasons for why this study found high CGM accuracy may include the lack of a sampling bias and the use of the Contour Next One BGM, which has a significantly lower MARD than other common BGMs.8 Specifically, the Contour Next glucometer had a MARD of 3.1% over all glucose levels when referenced against the YSI glucose analyzer, whereas five other BGMs had MARD values ranging from 4.0% to 9.7%.

The CGM was most accurate when the rate of change of CGM values was low. Moser et al.9 found that intermittently scanned CGMs demonstrated significantly worse accuracy when CGM values were rapidly increasing or decreasing, whereas Wadwa et al.7 found the accuracy was similar for various rates of change. Our results were in between these two findings, in that CGM accuracy was only slightly worse when CGM values were falling rapidly.

The bias in our sample, in which CGM values on average were slightly higher than BGM values, was surprising as the results were consistent across the range of glucose values. Typically, bias suffers from a regression to the mean effect and possibly a floor effect where the CGM often exhibits a positive bias when the reference blood glucose is low, and a negative bias when the reference blood glucose is high. In our study, the CGM values were on average 4–6 mg/dL higher than the BGM values irrespective of the BGM value, with the smallest bias occurring after 24 h of sensor wear. Additional real-world studies are needed to further investigate whether there is a positive bias in the Dexcom G6 Pro over the lifetime of the 10-day sensor wear period.

It is also notable that the bias in the sensors with lower accuracy was positive and was larger than the mean bias for all the sensors collectively. This could result in hypoglycemia that is not detected by the CGM. Any bias would not affect real-time insulin decisions as the Dexcom G6 Pro is blinded, but this could affect insulin dosing and medication recommendations depending on how information from the Dexcom G6 Pro is used. However, only a small percentage of sensors exhibited high MARD.

The main limitations of this study are that we only assessed accuracy during the first 48–60 h of sensor life rather than the full 10 days of sensor life. Although it was not available at the time this study was performed, the Dexcom G7 sensor is now available. However, integration of Dexcom G7 with current pump and AID systems is still in development. Therefore, to the extent that the accuracy of the G6 Pro is representative of the accuracy of G6, these data remain relevant for many CGM users.

This study demonstrates that the Dexcom G6 Pro evaluated in an at-home environment relative to BGM measurements achieved an accuracy that was comparable with the accuracy obtained when evaluated in a laboratory setting relative to YSI measurements. Accuracy improved after the first 12 h and then remained stable thereafter.

Authors' Contributions

M.C.M. performed statistical analysis, interpreted the data, and reviewed/edited the article; R.W.B. researched data and wrote the article; P.C., E.R.D., S.J.R., and K.J.R. researched data, contributed to the discussion, and reviewed/edited the article.

Author Disclosure Statement

M.C.M. has no personal financial disclosures but reports that his employer has received grant support from Beta Bionics, Dexcom, and Tandem Diabetes Care. P.C.'s employer has received consulting payments on his behalf from vTv Therapeutics, Beta Bionics, Dexcom, and Diasome. E.R.D. has issued patents and pending patents on aspects of the bionic pancreas, and is an employee, the executive chair of the board of directors, and shareholder of Beta Bionics.

S.J.R. has issued patents and pending patents on aspects of the bionic pancreas that are assigned to Massachusetts General Hospital and licensed to Beta Bionics; has received honoraria and/or travel expenses for lectures from Novo Nordisk, Roche, and Ascensia; serves on the scientific advisory boards of Unomedical; served on scientific advisory board and had stock in Companion Medical that was bought out by Medtronic; has received consulting fees from Beta Bionics, Novo Nordisk, Senseonics, and Flexion Therapeutics; has received grant support from Zealand Pharma, Novo Nordisk, and Beta Bionics; and has received in-kind support in the form of technical support and/or donation of materials from Zealand Pharma, Ascencia, Senseonics, Adocia, and Tandem Diabetes; and is an employee, the chief medical officer, and a shareholder of Beta Bionics.

K.J.R. has no personal financial disclosures but reports that his employer has received grant support from Beta Bionics, Dexcom, and Tandem Diabetes Care. R.W.B. reports no personal financial disclosures but reports that his institution has received funding on his behalf as follows: grant funding and study supplies from Tandem Diabetes Care, Beta Bionics, and Dexcom; study supplies from Medtronic, Ascencia, and Roche; consulting fees and study supplies from Eli Lilly and Novo Nordisk; and consulting fees from Insulet, Bigfoot Biomedical, vTv Therapeutics, and Diasome.

Funding Information

Study funding provided by the National Institute of Diabetes and Digestive and Kidney Diseases (Grant No. 1UC4DK108612-01), by an Investigator-Initiated Study award from Novo Nordisk, and by Beta Bionics, Inc., which also provided the experimental bionic pancreas devices used in the study. Fast-acting insulin aspart and insulin aspart were provided by Novo Nordisk and insulin lispro by Eli Lilly. Blood glucose meters and test strips (Contour Next One Blood Glucose Monitoring System) were provided by Ascensia Diabetes Care, Basel, CH. Continuous glucose monitor sensors and transmitters were purchased from Dexcom, Inc., at a discounted price.

References

  • 1. Cappon G, Vettoretti M, Sparacino G, et al. Continuous glucose monitoring sensors for diabetes management: A review of technologies and applications. Diabetes Metab J 2019;43(4):383–397. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Welsh JB, Psavko S, Zhang X, et al. Comparisons of fifth-, sixth-, and seventh-generation continuous glucose monitoring systems. J Diabetes Sci Technol 2022:1–5;19322968221099879. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Link M, Kamecke U, Waldenmaier D, et al. Comparative accuracy analysis of a real-time and an intermittent-scanning continuous glucose monitoring system. J Diabetes Sci Technol 2021;15(2):287–293. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Jafri RZ, Balliro CA, El-Khatib F, et al. A three-way accuracy comparison of the Dexcom g5, Abbott Freestyle Libre Pro, and senseonics eversense continuous glucose monitoring devices in a home-use study of subjects with type 1 diabetes. Diabetes Technol Ther 2020;22(11):846–852. [DOI] [PubMed] [Google Scholar]
  • 5. Russell SJ, Beck RW, Damiano ER, et al. Multicenter, randomized trial of a bionic pancreas in type 1 diabetes. N Engl J Med 2022;387(13):1161–1172. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Balliro CA. 96-lb: Safety of glucose regulation by the bionic pancreas without continuous glucose monitoring input. Diabetes 2022;71(Supplement_1):96-LB. [Google Scholar]
  • 7. Wadwa RP, Laffel LM, Shah VN, et al. Accuracy of a factory-calibrated, real-time continuous glucose monitoring system during 10 days of use in youth and adults with diabetes. Diabetes Technol Ther 2018;20(6):395–402. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Klaff LJ, Brazg R, Hughes K, et al. Accuracy evaluation of contour next compared with five blood glucose monitoring systems across a wide range of blood glucose concentrations occurring in a clinical research setting. Diabetes Technol Ther 2015;17(1):8–15. [DOI] [PubMed] [Google Scholar]
  • 9. Moser O, Sternad C, Eckstein ML, et al. Performance of intermittently scanned continuous glucose monitoring systems in people with type 1 diabetes: A pooled analysis. Diabetes Obes Metab 2022;24(3):522–529. [DOI] [PubMed] [Google Scholar]

Articles from Diabetes Technology & Therapeutics are provided here courtesy of Mary Ann Liebert, Inc.

RESOURCES