Dexcom G7 Accuracy and Reproducibility in the Intensive Care Unit

Abstract

Objective:

To evaluate the accuracy of Dexcom G7 continuous glucose monitor (CGM) in the intensive care unit (ICU) setting.

Methods:

We performed a prospective, single-center study in patients with known diagnosis of diabetes or stress hyperglycemia and treated with insulin. Two Dexcom G7 sensors were placed on the abdomen and/or upper arm. Blood glucose (BG) measurements obtained according to usual ICU care were paired with sensor glucose values, and accuracy metrics were analyzed. For further comparison, non-ICU patients were also studied.

Results:

The analyses included 30 participants with mean ± standard deviation age of 55 ± 12 years, with preexisting diabetes in 40% and stress hyperglycemia in 60%. A total of 1515 sensor-BG pairs were analyzed. The mean difference (bias) was −12 mg/dL (median: −6), and the mean relative absolute difference (RAD) was 16% (median: 12%). Mean RAD was 13% (median: 9%) using plasma glucose as the reference and 17% (median: 13%) using capillary glucose. For comparison, in 35 adults with type 2 diabetes in a non-ICU setting, the mean RAD was 15% (median: 13%). No meaningful differences were observed across the duration of time since sensor insertion. No correlation was found between mean RAD and severity of illness.

Conclusions:

Mean RAD of the Dexcom G7 sensor in the ICU setting was slightly higher than the outpatient use labeling, but was similar to a non-ICU hospital setting. Further studies are needed to determine whether CGM can be used nonadjunctively in an ICU setting for insulin management, including use of glucose trends and alarms for hypoglycemia or hyperglycemia.

Introduction

Hyperglycemia, hypoglycemia, and glucose variability are associated with increased risk of morbidity and mortality in intensive care unit (ICU) patients, ^{1 –4} and achievement of glycemic targets is critical to preventing poor clinical outcomes. Frequent blood glucose (BG) measurements are necessary to adjust insulin doses and maintain glucose levels within goal, which lead to high workload demands for clinical staff.

Continuous glucose monitoring (CGM) systems may offer an adjunctive option to periodic BG testing in the ICU setting with some potential advantages. ^5,6 However, the required accuracy of CGM to titrate insulin in the ICU non-adjunctively is not known by rigorously controlled trials and may not be the same as the outpatient setting.

Several recent pilot studies evaluating CGM technology in the ICU have been conducted and have shown variable levels of accuracy. ^{7 –11} Multiple factors related to the patient’s medical condition or treatment in the ICU setting may impact CGM accuracy.

Thus, additional data are needed to better understand the degree of accuracy that can be achieved with CGM in the ICU. In recent years, new CGM systems have been launched on the market with improvements in CGM technology and accuracy. The aim of this study was to assess the accuracy of Dexcom G7 CGM in the ICU setting. These data are required to evaluate the utility and limitations for using CGM as an adjunct to care in future ICU trials of glycemic control.

Methods

The study was conducted at one hospital. The protocol and informed consent were approved by an institutional review board. Study participants were adults age ≥18 years who were hospitalized in an ICU, with an expected length of stay in the ICU of at least 2 days. Participants had either a known diagnosis of diabetes or stress hyperglycemia with at least two BG measurements >180 mg/dL and were being treated with insulin.

After informed consent was obtained and eligibility determined, two Dexcom G7 sensors were placed in different locations on either the abdomen or upper arm and blinded by covering the receiver screen. All alarms were silenced except for the Urgent Low alert. Blinded CGM data were collected, with the sensor replaced as needed, for up to 10 days or until discharge from the ICU, or death if before 10 days. CGM sensors were planned to be removed before magnetic resonance imaging, diathermy, or other incompatible procedures (however, this did not occur during the study). The sensor glucose data were not used as part of glucose management. To blind the receiver, all possible alerts were turned off. Alerts that were not possible to turn off (Urgent Low and Technical Alerts) were set to Vibrate Only using Alert Sounds in the receiver.

BG measurements were made according to the ICU’s usual routine and could be capillary, venous, or arterial samples for laboratory or BG meter (Accu-Chek Inform II) ¹² measurements. Medical information was obtained from the electronic health record. Severity of illness was assessed by calculating Acute Physiology and Chronic Health Evaluation (APACHE) II scores for each participant. ¹³

For analyses, the BG measurements were paired with available sensor glucose values obtained within 5 min. Standard accuracy metrics were computed overall and in ranges based on the BG value as follows: <70, 70–99, 100–139, 140–179, 180–249, and ≥250 mg/dL. The effect of lag was assessed by computing accuracy metrics in which BG values were paired with CGM values that were shifted back by various amounts of time (2.5, 5, 7.5, 10, 12.5, and 15 min). To assess reproducibility (precision), values from two sensors worn simultaneously by participants were compared. A Spearman correlation coefficient was computed to assess the association of the APACHE II score and the mean relative absolute difference (RAD) of the sensors worn by a participant.

Accuracy metrics were also computed for G7 (blinded) data obtained in a non-ICU setting from a randomized trial (“TIGHT,�? NCT05135676) involving adults with type 2 diabetes that was conducted at the same hospital plus five additional academic hospitals. ¹⁴

Results

The analyses included 30 participants ranging in age from 25 to 77 years (mean 55 ± 12 years); 37% were female, 80% were White, and 7% of Hispanic ethnicity. Diabetes was present in 40% and stress hyperglycemia in 60%. The primary medical condition was cardiac in 40%, respiratory failure in 27%, lung transplant in 13%, and other conditions in 20% (Table 1). APACHE II scores ranged from 10 to 25 (median 14, interquartile range 12–16).

Table 1.

Participant Characteristics

	N = 30
Age (years)
Mean (SD)	55 (12)
Range	25–77
Sex—n (%)
Female	11 (37%)
Male	19 (63%)
Race—n (%) a
White	24 (80%)
Black	1 (3%)
Asian	1 (3%)
American Indian/Alaskan Native	1 (3%)
Native Hawaiian/Other Pacific Islander	2 (7%)
Ethnicity—n (%)a
Hispanic/Latino	2 (7%)
Diabetes/Stress hyperglycemia diagnosis
Diabetes	12 (40%)
Stress hyperglycemia	18 (60%)
IV insulin received—n (%)	26 (87%)
Vasopressor drug received—n (%)	30 (100%)
Medical condition—n (%)
Cardiac	12 (40%)
Respiratory failure	8 (27%)
Lung transplant	4 (13%)
Vascular	2 (7%)
Malignancy	2 (7%)
Shock	1 (3%)
Liver failure	1 (3%)
ICU type—n (%)
Cardiac/Cardiothoracic	23 (77%)
Medical	7 (23%)
APACHE II Score
Median (IQR)	14 (12–16)
Range	10–25

^aMissing for one participant.

APACHE, Acute Physiology and Chronic Health Evaluation; ICU, intensive care unit; IQR, interquartile range; SD, standard deviation.

Accuracy analyses

Among the 30 participants, there were 60 sensors worn for which there were corresponding BG values for analysis. The sensor location was the abdomen in 15% and arm in 85%. There were 1515 sensor-BG pairs included in the analyses, with 1191 (79%) of the BG measurements from capillary blood and 324 (21%) from plasma.

Overall, the mean difference (bias) was −12 mg/dL (median: −6) (CGM minus BG), and the mean RAD was 16% (which was 1.3 times larger than the median value of 12%) (Table 2). The mean and median differences between CGM and BG values were small for BG values <180 mg/dL. For higher BG values, CGM values tended to be lower than BG values. The mean and median RADs were slightly higher for BG values <140 mg/dL or ≥250 mg/dL than for 140–249 mg/dL.

Table 2.

Accuracy Metrics for Blinded G7 Sensors Compared with Blood Glucose Measurements in Intensive Care Unit Setting

	N (pairs)	Mean RAD	Median RAD	Mean difference (mg/dL)	Median difference (mg/dL)	Mean absolute difference (mg/dL)	Median absolute difference (mg/dL)	15%/15 rule met	20%/20 rule met	30%/30 rule met
Overall	1515	16%	12%	−12	−6	25	18	59%	72%	86%
BG (mg/dL)
<70	30	16%	9%	3	−1	10	5	92%	95%	95%
70–99	120	17%	14%	1	2	15	12	58%	71%	88%
100–139	590	16%	13%	−5	−3	20	15	57%	69%	83%
140–179	382	15%	11%	−11	−7	23	18	62%	74%	88%
180–249	268	14%	11%	−19	−16	30	23	62%	74%	90%
≥250	125	18%	15%	−47	−40	56	44	49%	66%	84%

The relative absolute difference (RAD) is computed as the absolute value of (CGM−BG)/BG. Difference is computed as CGM minus BG. The 15%/15 rule is met if the CGM is within either ±15% or ±15 mg/dL of the BG reference (i.e., RAD ≤15% or absolute difference ≤15 mg/dL). The 20%/20 rule and 30%/30 rule have similar definitions. ¹⁵ Sensor values that were <40 or >400 mg/dL were analyzed as 40 and 400 mg/dL, respectively.

BG, blood glucose; CGM, continuous glucose monitoring.

RAD values were similar with abdomen sensor insertions compared with arm insertions (Supplementary Table S1). Accuracy appeared similar across the duration of time since sensor insertion, without an indication of lower accuracy on day 1 (Supplementary Table S2). Adjusting for a lag effect for up to 15 min had no impact on the results (Supplementary Table S3). Similarly, results were similar when stratifying by rate of change (Supplementary Table S4). Sensor accuracy appeared to be better compared with plasma glucose concentrations than capillary glucose concentrations (mean 13% versus 17%; median 9% versus 13%; Supplementary Table S5). The correlation between the APACHE II score and mean RAD for sensors worn by the participant was −0.10 (Supplementary Fig. S1).

In a non-ICU setting, a blinded G7 sensor was worn by 35 adults with type 2 diabetes. The mean RAD was 15% (median: 13%) and the mean absolute difference was 24 mg/dL (median: 19) for 846 sensor-BG pairs (Table 3).

Table 3.

Blinded G7 Accuracy Metrics in Non-Intensive Care Unit Setting

	N (pairs)	Mean RAD	Median RAD	Mean difference (mg/dL)	Median difference (mg/dL)	Mean absolute difference (mg/dL)	Median absolute difference (mg/dL)	15%/15 rule met	20%/20 rule met	30%/30 rule met
Overall	846	15%	13%	6	10	24	19	57%	73%	91%
BG (mg/dL)
<70	11	20%	17%	11	9	12	9	76%	84%	92%
70–99	57	23%	17%	6	5	20	15	50%	64%	80%
100–139	252	17%	16%	8	11	20	19	49%	69%	90%
140–179	260	14%	12%	10	11	22	19	60%	74%	92%
180–249	203	12%	11%	6	9	26	22	63%	79%	95%
≥250	63	15%	11%	−12	6	45	29	66%	74%	89%

The 846 pairs from 44 sensors were worn by 35 participants. The relative absolute difference (RAD) is computed as the absolute value of (CGM-BG)/BG. Difference is computed as CGM minus BG. The 15%/15 rule is met if the CGM is within either ±15% or ±15 mg/dL of the BG reference (i.e., RAD ≤15% or absolute difference ≤15 mg/dL). The 20%/20 rule and 30%/30 rule have similar definitions. ¹⁵ Sensor values that were <40 or >400 mg/dL were analyzed as 40 and 400 mg/dL, respectively.

Precision analyses

Two blinded G7 sensors were worn simultaneously by all 30 participants, which generated 28,926 glucose pairs. The mean absolute difference was 22 mg/dL (median: 16), and the mean RAD was 16% (median: 11%) (Table 4). Mean RAD ranged from 33% (median: 25%) for glucose values <70 mg/dL to 8% (median: 7%) for glucose values ≥250 mg/dL.

Table 4.

Precision Analysis for Concurrently Worn G7 Sensors ^a

	N (pairs)	Mean RAD	Median RAD	Mean absolute difference (mg/dL)	Median absolute difference (mg/dL)	15%/15 rule met	20%/20 rule met	30%/30 rule met
Overall	28,926	16%	11%	22	16	63%	74%	88%
Average of Concurrent CGM (mg/dL)
<70	448	33%	25%	20	15	51%	61%	74%
70–99	3200	27%	20%	23	18	45%	55%	73%
100–139	10,110	17%	12%	20	15	58%	70%	85%
140–179	7835	13%	10%	21	16	66%	78%	91%
180–249	5465	12%	9%	24	18	70%	80%	94%
≥250	1868	8%	7%	24	20	87%	94%	99%

^aThis analysis compares CGM values from two sensors worn simultaneously by the participant (does not use BG data). This analysis is limited to the N = 30 participants who wore simultaneous CGM sensors. The relative absolute difference (RAD) is computed as the absolute value of the difference between the two CGM readings divided by their average. The 15%/15 rule is met if the CGM is within either ±15% or ±15 mg/dL of the CGM average (i.e., RAD ≤15% or absolute difference ≤15 mg/dL). The 20%/20 rule and 30%/30 rule have similar definitions. ¹⁵ Sensor values that were <40 or >400 mg/dL were analyzed as 40 and 400 mg/dL, respectively.

Discussion

The goal of TIGHT-ICU was to observe the accuracy and precision of the Dexcom G7 sensor in the ICU. We learned that the degree of accuracy was similar in the ICU as the non-ICU setting with the mean RAD of 16% (median: 12%) and 15% (median: 13%), respectively. Not surprisingly, accuracy appeared better when sensor glucose levels were compared with plasma glucose concentrations than with capillary glucose. These findings compare to a mean absolute relative difference of 8.8% and median absolute relative difference of 7.8% in an outpatient registration study. ¹⁶ Precision analysis with almost 29,000 paired glucose measurements showed a mean RAD of 16% (median: 11%), with the absolute difference being similar across the glucose range, but the RAD being higher at low glucose levels.

Like in the outpatient setting, there were no meaningful differences between abdomen and arm sensor insertion in the ICU. Interestingly, while CGMs are reported to perform less well during the first 24 h after insertion, we observed no meaningful difference in median RAD on day 1 compared with subsequent days. It is difficult to draw conclusions on accuracy after day 7 as the sample sizes were small. Furthermore, accounting for a possible lag of up to 15 min between interstitial fluid and BG measurements did not affect the results.

Our results with Dexcom G7 show intermediate accuracy to sensors from prior generations in the ICU. ^{7 –10,17} For example, the Dexcom G6 CGM in the ICU has shown mean RADs of 10.4%–12.7%, ⁹ 9.4%, ¹⁷ and 14.8%. ¹⁸ A pediatric ICU study assessing the accuracy of the Medtronic Guardian Sensor 3, the Dexcom G6, and the FreeStyle Libre 1 reported mean RADs of 13.4%, 11.1%, and 11.3%, respectively. ¹⁹ When considering the entire literature about CGM in critically ill patients, 97 reports were reviewed and for the newer factory-calibrated CGMs (performed before the introduction of Dexcom G7), mean absolute relative differences ranged from 7.9% to 20.6%. ^20,21 The other challenge with the literature is that the comparators are not consistent. A recently published outpatient study found considerably different CGM performance results when three different comparison measurements were used, despite being performed on the same venous blood samples. ²² In our study, we measured plasma and capillary blood (mean/median RAD of 13%/9% and 17%/13%, respectively), whereas the accuracy data noted above could be arterial, venous plasma, or capillary. Ideally, the same comparator should be used when comparisons of RAD are made.

Accuracy of blood or interstitial glucose monitoring in the ICU is complicated by factors that can alter sensor accuracy. Changes in blood pressure, vasopressors, ²³ dialysis, and interfering substances are several of the many factors thought to impact accuracy. However, there was virtually no correlation between accuracy and disease severity as measured with APACHE II scores, the standard method of assessing severity of critical illness. Putting our population in perspective, the mean APACHE II score for the surgical critically ill patients studied in the 2001 study from Van den Berghe et al. was 9, ²⁴ whereas the mean score from the Normoglycemia in Intensive Care Evaluation–Survival Using Glucose Algorithm Regulation (NICE-SUGAR) study (37% surgical) was 21. ² Our median APACHE II score of 14 was between these two well-known studies.

Currently, the U.S. Food and Drug Association (FDA) has only one approved hospital BG meter, the Nova StatStrip®. ²⁵ In the 2018 meeting of the FDA Clinical Chemistry and Clinical Toxicology Devices Panel for capillary BG meters in the hospital (including the ICU), there were 5 proposals for final accuracy, some more permissive than others. ²⁶ It is unclear what accuracy requirements would be required for CGM in the hospital for those requiring insulin, especially for critically ill patients, where intravenous insulin is frequently used.

Limitations of this study include a single center evaluation with a relatively small number of mostly White subjects, relatively few glucose levels in the hypoglycemic range, and lack of a standard method for measuring BG with use of a BG meter for many of the measurements. Still, we used one of the most advanced CGM sensors currently available, and our results showed accuracy that was similar to the non-ICU setting. Whether this sensor is accurate enough to use in an ICU where the majority of administered insulin is provided by the intravenous route is unclear.

In conclusion, we showed that the mean RAD of the Dexcom G7 in an ICU setting is slightly higher than the outpatient use labeling. Further studies are needed to determine whether CGM can be used nonadjunctively in an ICU setting for insulin management, including use of glucose trends and alarms for hypoglycemia or hyperglycemia.

Authors’ Contributions

All authors contributed to the interpretation of study results and provided critical review of the content of the article. I.B.H. and R.W.B. are the guarantors of the work and, as such, had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Author Disclosure Statement

F.J.P. has received research support through his institution from Dexcom, Tandem, Insulet, Novo Nordisk, and Ideal Medical Technologies, personal consulting fees from Dexcom, and has provided consulting services for Insulet (services paid to his institution). J.U. reports research grants from the Cystic Fibrosis Foundation and the Jaeb Center for Health Research, payment, or honoraria for lectures from the American Diabetes Association and leadership or fiduciary roll for the Greater Pittsburgh Diabetes Club. C.C.L.W. has received research support from Dexcom, Inc. and Virta Health. J.B.B. reports research support from Bayer, Boehringer-Ingelheim, Carmot, Corcept, Dexcom, Eli Lilly, Insulet, MannKind, Novo Nordisk, and vTv Therapeutics; consulting fees from Alkahest, Altimmune, Anji, Aqua Medical Inc, AstraZeneca, Boehringer-Ingelheim, CeQur, Corcept Therapeutics, Eli Lilly, Embecta, GentiBio, Glyscend, Insulet, Mediflix, Medscape, Medtronic MiniMed, Mellitus Health, Metsera, Moderna, Novo Nordisk, Pendulum Therapeutics, Praetego, ReachMD, Stability Health, Tandem, Terns Inc, and Vertex; and stock options from Glyscend, Mellitus Health, Pendulum Therapeutics, Praetego, and Stability Health. UNC efforts in the conduct of this study were provided by grants from NIDDK (P30DK124723) and NCATS (UM1TR004406). R.W.B. reports no personal financial disclosures, but reports that his institution has received funding on his behalf as follows: grant funding, study supplies, and consulting fees from Insulet, Tandem Diabetes Care, and Beta Bionics; grant funding and study supplies from Dexcom; grant funding from Bigfoot Biomedical, Embecta, Sequel Med Tech, and MannKind; consulting fees and study supplies from Novo Nordisk; consulting fees from Vertex, Hagar, Ypsomed, Sanofi, and Zucara; and study supplies from Medtronic, Ascencia, Roche, and Eli Lilly. I.B.H. reports research support from Dexcom, Tandem, and Mannkind and consulting fees from Abbott Diabetes Care, Roche, Embecta, and Vertex. J.D.B., C.S., R.B., H.-H.C., J.H.C., D.K., P.P., M.S.J., J.S., and C.K. report no financial disclosures.

Funding Information

This study was supported by funding and donation of CGM supplies from Dexcom Inc to the JAEB Center for Health Research, which then provided funding to the clinical sites.