Real-world Accuracy of CGM in Inpatient Critical and Noncritical Care Settings at a Safety-Net Hospital

Management of hyperglycemia in the hospital is essential, and poor control leads to adverse outcomes (1,2). Continuous glucose monitor (CGM) technology has gained increased use, study, and interest in inpatient settings (3–14) but has yet to receive formal U.S. Food and Drug Administration approval for inpatient use.

Previous studies have documented the accuracy of real-time CGM in noncritical care and critical care settings, but most have been in controlled experimental situations and with small patient cohorts. Additionally, almost all previous studies have compared CGM with capillary point of care (POC) testing, while data on the accuracy of CGM compared with laboratory glucose values remains limited. We present here a multiyear analysis of CGM accuracy using both POC and laboratory (Lab) glucose reference values, in a real-world setting, across critical and noncritical care units in a safety-net hospital.

We conducted a retrospective analysis of adult patients at a large, 525-bed community academic hospital who used >72 h of CGM during their inpatient admission between 15 May 2020 and April 2022. Eligible patients for CGM use included those requiring contact/respiratory isolation, a planned admission of >48 h, and dysglycemia, including hyperglycemia, severe hypoglycemia, high insulin requirements (>2 units/kg/day), or use of an insulin drip. Patients who were pregnant during admission were excluded from this analysis. We obtained data from electronic medical records (EMR) and the Dexcom G6 Clarity website. The study protocol was exempt from Institutional Review Board review by the Colorado Multiple Institutional Review Board (COMIRB; Protocol 21-4681), Aurora, CO.

Dexcom G6 real-time CGMs (Dexcom, San Diego, CA) were used in the medical intensive care unit (ICU), surgical ICU, and designated medical-surgical units per a hospital CGM protocol. Bedside nurses were trained to place, troubleshoot, and remove CGMs. A provider order was required for CGM placement. CGM sensors were placed on the upper arm rather than the abdomen to facilitate patient repositioning, including prone positioning for patients with respiratory failure. Following placement, POC monitoring continued until two consecutive glucose results were within 30 mg/dL of a concurrent CGM reading. After two consecutive CGM-reference glucose results within 30 mg/dL, CGMs were used for glucose monitoring and insulin dosing, based on provider orders.

Per CGM protocol, additional POC checks were performed for signs and symptoms discordant with CGM readings, lack of displayed CGM glucose (e.g., sensor error message), CGM glucose <80 mg/dL, “low” or “high” alert, or if “urgent low soon” alert was displayed. Calibration was performed if the discrepancy with the reference glucose was >30 mg/dL for two consecutive values. The CGM sensor was replaced if discrepancies persisted after calibration. CGM use was continued until patients no longer met inclusion criteria and the sensor expired.

CGM data were transmitted to smartphones outside patient rooms for clinical decision making and transmission into the Dexcom Clarity database. Nurses documented the sensor glucose value displayed on the smartphone, and corresponding action (insulin administration, hypoglycemia treatment) if any, into a glucose flow sheet in the patient’s EMR. Frequency of monitoring and insulin dosing was based on provider orders.

Demographic, clinical, and laboratory data were extracted from the EMR. Laboratory analyzed (Lab) glucose was measured in serum (Atellica, Siemens) or whole blood (ABL 825 Flex Analyzer, Radiometer). POC glucose was measured with a StatStrip glucose meter (Nova Biomedical). CGM data were extracted from Dexcom Clarity. Glucose data from calibration events were excluded from the analyses. Patient encounters were designated as critically ill (CI) or not critically ill (NCI) based on presence of critical care billing code or interventions restricted to ICUs, including use of ventilation or pressor medications. Patients who spent part of their hospitalization in the ICU were defined as CI.

CGM glucose was matched to reference (POC or Lab) glucose closest by time. For multiple glucose values at the same time, values were averaged. Pairs with CGM or reference glucose values outside the reference range (i.e., “high” >400 mg/dL, “low” <40 mg/dL on CGM, or >600 mg/dL on reference glucose) were excluded from subsequent analyses.

Clarke Error Grids (CEG) (15) were created with the R ega package (16). Average differences between reference and CGM values were calculated by the aggregated mean absolute relative difference (MARD) and average MARD. Aggregated MARD was calculated as the average relative difference within all the matched pairs for each patient. Median and interquartile ranges were reported across patients. The average MARD was calculated as the mean of the absolute relative differences across all matched pairs, without accounting for correlation within patients. Additionally, the percentage accuracy assessments of CGM readings within 15, 20, and 30 mg/dL of reference glucose values ≤100 mg/dL or 15, 20, and 30% of POC >100 mg/dL (%15/15, %20/20 and %30/30) were calculated.

CEG errors in regions E and D were assessed for frequency, and medical records of patients with more than four errors were manually reviewed to identify diagnoses and interventions predisposing to inaccuracies.

CGM data from 239 encounters in 233 individual patients were analyzed. Six encounters were readmissions. Patient characteristics are shown in Table 1. The population analyzed was of high acuity, with 153 CI patients (66%). A diabetes diagnostic code was present for 85% of patients, including 95% of NCI patients and 79.7% of CI patients. The diagnosis was type 2 diabetes in 81% of patients, type 1 diabetes in 10.7%, and “other diabetes” in 14.6%. Diagnostic codes for both type 1 diabetes and type 2 diabetes were found for 17 patients (8.6%). Most patients (79.9%) received acetaminophen, but none received >4 g/day, and none received hydroxyurea. CGM calibration was required for 170 of 233 patients (73%) (Supplementary Table 4).

There were 6,598 CGM results to reference glucose matches, consisting of 2,875 CGM-POC and 3,723 CGM-Lab. Included were 6,449 matched pairs, including 2,744 CGM-POC and 3,705 CGM-Lab, after exclusion of 131 CGM-POC (4.55%) and 18 CGM-Lab (0.5%) with values beyond the measuring instrument detection limit.

Aggregated MARD, average MARD, and percentage accuracy assessments are noted in Table 2, average MARD by day of wear, time of day, and BMI in Supplementary Tables 2 and 3, and Supplementary Fig. 1, and ARD by reference glucose ranges in Supplementary Table 1. Aggregated MARD was 13.8% (CGM-POC) and 9.8% (CGM-Lab) for NCI and 16.3% and 12.1%, respectively, for CI. Overall mean absolute difference for CGM ≤80 mg/dL was 25.1 (SD 33.0) mg/dL for POC and 20.3 (SD 30.0) mg/dL for Lab. All accuracy measures were better in NCI than in CI and for CGM-Lab than for CGM-POC. MARD was higher on day 1 of CGM use than on days 2–7 and more variable by time of day for NCI than for CI. The average MARD was higher for patients with BMI <18 kg/m² than those with a higher BMI.

For CEG (Fig. 1) of CGM-POC pairs, 96.5% of values were in regions A and B, while 3.3% were in the critical error D and E regions. For CEG of CGM-Lab pairs, 99.1% were in regions A and B and 0.9% in region D. Accuracy per CEG was better for patients assigned NCI compared with those assigned as CI. A lower absolute relative difference was more often observed for referent glucose of ≥180 mg/dL versus ≥70 and <180 mg/dL or <70 mg/dL, and for Lab versus POC referent glucose (Supplementary Table 1).

In assessment of type D/E errors, 43% occurred within 24 h of sensor start. Type D/E errors were experienced by 53 individuals, although 50% were attributed to 9 patients who had more than four type D/E errors. Among these nine patients, three had anorexia nervosa with hypoglycemia, one had a hypoglycemia disorder, one underwent cooling after cardiac arrest, one was on extracorporeal membrane oxygenation (ECMO) and continuous renal replacement therapy, and three were intubated with hypoxia but had no other unique conditions.

We measured the real-world accuracy of inpatient CGM in a large, diverse, safety-net hospital, using both POC and Lab glucose as referent values. Without restricting data to a controlled setting, we demonstrated acceptable MARD in both CI and NCI patients. As expected, accuracy of CGM assessed by MARD and CEG was better for Lab compared with POC reference values and for NCI compared with CI patients. Most critical errors by CEG were in a subset of patients with unique conditions, including anorexia nervosa, severe hypoglycemia, and decreased skin perfusion while undergoing therapeutic cooling and ECMO.

In NCI, our study found MARD of 13.8% for POC-matched pairs, which is within the range previously reported for real-time CGM (7,10,11,14). A pooled analysis of 218 patients/4,067 matched pairs from three inpatient CGM studies (two interventional and one observational) reported MARD of 12.8% (14). Smaller studies of 9 to 49 patients/105–596 matched pairs have reported MARD of 9.4–14.8% (11). Compared with POC-matched pairs, we found higher accuracy in Lab-matched pairs, with MARD of 9.8% in NCI. This is lower than the previously reported MARD of 11.3% in 117 CGM-Lab matched pairs from 18 patients (10).

Few studies have evaluated CGM accuracy during critical illness, and almost all have used POC glucose as the referent measure. We analyzed 2,121 POC-matched pairs and 3,171 Lab-matched pairs from 153 CI patients and found MARD of 16.3% and 12.1%, respectively. Recent small studies of 5–19 patients/199–493 matched pairs have demonstrated MARD for POC-matched pairs of 10.9–13.9% (8–11,13,17). CGM accuracy using Lab referent measures was reported for one small study of 84 matched pairs from 10 patients showing MARD of 10.4% (10), and a large, randomized trial evaluating older CGM technology during intensive insulin management in the ICU reported MARD of 13.3% (12). A study of CGM accuracy using Lab referent measures in patients undergoing open cardiac surgery (18) reported very high MARD during surgery (23.8%), extracorporeal circulation (29.1%), and induced hypothermia (41.6%), with improved MARD (15.0%) after surgery.

In our analysis of data from 153 CI patients in a real-world setting, MARD values for POC-matched pairs were higher than, and MARD values for Lab-matched pairs were similar to, those of prior studies. Higher MARD in our POC-matched pairs compared with prior reports could reflect higher variability in real-world application of the device or inclusion of patients with conditions associated with lower CGM accuracy (therapeutic cooling, ECMO). Alternatively, this discrepancy may have resulted from our hospital protocol, which required POC glucose checks in situations where higher CGM discordance is expected (e.g., hypoglycemia, hyperglycemia, symptom discrepancies, etc.) or when discordance with POC was observed. In contrast, Lab glucose was obtained at regularly scheduled times (typically with daily laboratory assessments or every 4 h), regardless of glucose pattern. Future inpatient studies of three-way matched POC, Lab, and CGM are needed to further evaluate accuracy in real-world settings. While recent recommendations (3) are focused on noncritical care use of CGM, the potential utility and impact of CGM use in the critical care setting could be greater given more intensive monitoring and higher risk management. Continued assessment and greater understanding of CGM benefits and limitations will be necessary for wide-scale implementation of CGM in ICUs.

Strengths of our study include a large cohort, large number of measures from both CI and NCI patients, use of both POC and Lab glucose as reference measures, and the real-world setting.

Limitations include the retrospective-observational design and encounter-level analyses that precluded the exploration of effects of interventions/medications on CGM accuracy, and use of real-world data, including documentation of timing of Lab and POC measures, which may be prone to error. Reliance on encounter-level diagnostic codes to determine diabetes type is an additional limitation.

Our analysis of a large, diverse hospital population from both critical care and noncritical care units as patients undergo numerous procedures and interventions provides us an increased understanding of CGM’s real-world accuracy and shortcomings. As this technology continues to progress and eventually becomes integrated into hospitals, continued study and discussion of its benefits and limitations will be crucial.

Funding. L.G. was supported by a collaboration between the Colorado School of Public Health Center for Innovative Design & Analysis (CIDA) and the University of Colorado School of Medicine Division of Endocrinology, Metabolism and Diabetes. I.D. was supported by National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases H1RO1DK130351-01. R.P. is a Clinical Scholar supported by the Robert Wood Johnson Foundation (77887).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. E.F., L.S., I.S.D., and R.I.P. acquired the data. E.F. and R.I.P. conceptualized the study and wrote the manuscript. L.G. analyzed the data. All authors contributed to study design and reviewed and edited the manuscript. R.I.P. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.