Use of Continuous Glucose Monitor in Critically Ill COVID-19 Patients Requiring Insulin Infusion

An Observational Study

Eileen R. Faulds; Andrew Boutsicaris; Lyndsey Sumner; Laureen Jones; Molly McNett; Keaton S. Smetana; Casey C. May; Elizabeth Buschur; Matthew C Exline; Matthew D. Ringel; Kathleen Dungan


J Clin Endocrinol Metab. 2021;106(10):e4007-e4016. 

In This Article


This was a retrospective observational study among COVID-19 patients requiring IV insulin in a medical ICU at a large Midwest academic medical center. The study was approved by the health system's institutional review board. Participants included all patients who received a CGM for managing IV insulin using a hybrid POC plus CGM protocol between May 1, 2020, and September 8, 2020.

Management of Hyperglycemia at Study Institution

A multidisciplinary team created hospital guidelines incorporating IV insulin in conjunction with a Dexcom G6 (San Diego, CA) CGM in critically ill COVID-19 patients with diabetic ketoacidosis, hyperosmolar nonketotic state, or refractory hyperglycemia. Education and training regarding sensor use and insertion was provided by the inpatient diabetes consult service to ICU nurse leaders. A detailed discussion of CGM training and implementation has been previously reported.[23] Patients requiring insulin infusion initiated by the inpatient teams for standard clinical practice, in conjunction with the diabetes consult service, were identified as potential patients for CGM use. The institution's guidelines require hourly POC glucose testing (24 POC glucose tests/day) for patients receiving insulin infusion. The hybrid POC plus CGM protocol was initially reserved for use in patients for whom IV infusion pumps could be externalized outside of the patient room.

Device Setup and use

Details of device setup and use were reported previously.[23] In brief, the CGM was placed on the left upper arm because frequent prone positioning limited abdominal placement. Although the upper arm is not a specifically FDA-approved site, there is evidence to support the accuracy of its use.[24,25] Alert thresholds were set at 100 mg/dL and 300 mg/dL, and the "Urgent Low Soon" alert was activated. Initial sensor validation was conducted using hourly paired sensor meter (POC, Novo Stat Strip) readings obtained within 5 minutes. The POC glucose values were obtained from both fingerstick and arterial line samples. The CGM was used nonadjunctively (standalone) to inform insulin rate adjustments once initial validation was obtained using 2 consecutive sensor-meter pairs approximately 1 hour apart meeting either of the following criteria: (1) sensor glucose <20% of the meter glucose (positive or negative direction) when the POC glucose is ≥100 mg/dL or (2) sensor glucose <20 mg/dL of the meter glucose (positive or negative direction) when the POC glucose is <100 mg/dL. POC glucose measures were required at a minimum of every 6 hours for continued validation. Additionally, a 1-time POC glucose was required if no CGM glucose value or trend arrow was displayed, for predicted low or low threshold alerts, if clinical signs or symptoms did not match the CGM value, for a change in clinical status (ie, hemodynamic compromise, intubation, change in nutrition), and/or initiation of a new sensor. Otherwise, nurses were encouraged to use the CGM data for adjusting the insulin infusion.

Upon transitioning to subcutaneous insulin, patients could continue to wear the CGM for up to 10 days. The frequency of POC glucose measures continued every 6 hours or before meals and at bedtime while the patient remained in the ICU. Once the patient transferred out of the ICU, CGM use was fully nonadjunctive with exceptions as noted previously.


Demographics, medication use, medical history, and laboratory values were collected from the patient's electronic health record (EHR). CGM devices were downloaded using Dexcom Clarity software, and discrete CGM glucose values were extracted for days 1 through 7 of wear, with each day representing a rolling 24 hours from the time of sensor placement. Sequential organ failure assessment (SOFA) score was calculated, though noted to have limited discriminant function for predicting mortality.[26] The SOFA score has been validated in critically ill patients[27] and is currently the scoring system recommended by the Sepsis 3 definition.[28] The SOFA score is not specific to COVID-19, but has been proposed as reasonable clinical trial endpoint in a variety of settings.[29]


The primary outcome, number of POC glucose measures, was collected from the EHR. Secondary outcomes included workflow, safety, and accuracy measures reported by day of CGM wear. Workflow measures included number of sensors placed, CGM values used for IV insulin titration, defined as any CGM value recorded in the EHR without a POC value documented within 1 hour, and time to initial sensor validation. Safety measures including frequency of hypoglycemia (POC glucose <55 mg/dL and <70 mg/dL) and daily percent time in range (TIR) (70–180 mg/dL), time above range (>180 mg/dL and >250 mg/dL), and time below range (<54 mg/dL and <70 mg/dL) were aggregated from discrete CGM values.[30] Accuracy measures included mean absolute relative difference (MARD) and mean absolute difference (MAD). POC glucose was used as the standard reference rather than laboratory measurement because of the real-world retrospective observational design. MARD and MAD were also calculated by glucose strata (<70, <100, 70–180, 180–250, and >250 mg/dL, respectively) and for sensor meter pairs obtained during hypoxia (oxygen saturation <90%) or hypotension (mean arterial pressure <65 mm Hg).


Continuous variables were reported as mean (SD) or median (interquartile range [IQR]) for normal or nonnormal distributions, respectively. Univariate associations between variables were measured using Wilcoxon rank-sum and Spearman rank correlation. It was determined that longer time to initial sensor validation on day 1 of sensor wear was associated with sensor/POC pairs (r s = 0.84, P < 0.001) resulting in greater representation of these participant data. Therefore, MARD and MAD were aggregated by day for each participant to control for subject clustering and variability in frequency of measurement. For daily MARD, linear mixed effect modeling for repeated measures was used to model daily MARD as a function of patient characteristics (Table 1 and Table 2), adjusting within subject clustering from daily measures. A 2-sided significance level of 0.05 was used for all the statistical tests.