Pediatric Inpatient Continuous Glucose Monitor Accuracy by Sensor Wear Day, Hospital Day, and Glucose Rate of Change

Abstract

Background:

Continuous glucose monitoring (CGM) improves glycemic outcomes but inpatient use in the U.S. is not FDA approved, despite the potential need for frequent glycemic monitoring during hospitalization. We sought to assess CGM accuracy in the pediatric hospital setting and to evaluate the impact of potential modifiers of accuracy including sensor wear day, hospitalization day, and glucose changes.

Methods:

In this retrospective cross-sectional accuracy study among youth and young adults with diabetes hospitalized at two pediatric centers, reference point-of-care (POC) capillary glucose values were compared with Dexcom G7 CGM glucose readings (paired within 5 minutes). Accuracy was assessed using standard metrics including mean absolute relative difference (MARD) and Parkes Error Grids, overall, and stratified by sensor wear day, hospital day, and CGM glucose rate of change.

Results:

We analyzed 544 POC-CGM glucose pairs from 68 participants with diabetes aged 2–22 years (mean 12.4±5.8 years old, 58.8% female, 66.2% white, 8.8% Hispanic) during 80 hospital encounters. The overall MARD was 12.5%. Accuracy was comparable on sensor wear day 1 and days 2–6 (MARD 12.3% vs 13.6%, respectively) and was best on wear days 7–10 (MARD 10.6%). Accuracy improved from hospital day 1 (MARD 13.7%) to hospital days 2–4 and 5–30 (MARD 10.4% and 12.5%, respectively). Overall and on wear day 1, almost all glucose pairs (99%) fell within low-risk Parkes Error Grid zones A and B. Accuracy was best with a flat (−1 to 1 mg/dL/min) rate of change (MARD 11.5–11.9%) and was reduced with rapidly rising or falling (beyond −2 or 2 mg/dL/min) CGM values (MARD 16.6–18.4%).

Conclusions:

In this pediatric hospital CGM accuracy study, MARD was within a favorable range of <14% overall and throughout sensor wear and hospitalization. Reduced accuracy with rapidly changing glucose highlights the need to incorporate trend arrows into hospital CGM protocols.

Graphical Abstract

INTRODUCTION

There is a growing imperative to improve in-hospital glycemic outcomes in people with diabetes by expanding the use of continuous glucose monitoring (CGM) to the pediatric hospital setting. CGM is the standard of care for outpatient diabetes management¹ given proven efficacy at preventing hypoglycemic events² and at improving glycemic outcomes^3,4 and quality of life.⁵ However, more data demonstrating the in-hospital accuracy of CGM, particularly in children, is needed to inform policies and to help call for FDA approval for inpatient CGM use in the United States across all ages.

The management of glycemia can be more complex in the hospital versus home setting. With acute illness, glucose regulation and insulin sensitivity are impacted by complex patient-specific (baseline insulin sensitivity/production), illness-specific (acute infection/injury) and treatment-specific factors (glucocorticoids, nutritional regimen, etc.)⁶ which may contribute to volatile glycemia. CGM provides around-the-clock glucose monitoring with alerts which, in the adult inpatient setting, has demonstrated reduced time hyperglycemic,⁷ improved hypoglycemia detection,⁸ duration^9,10 and recurrence,¹¹ and lower in-hospital complications.¹⁰ Hospitalized pediatric patients would similarly benefit from the inpatient use of CGM, particularly among patients who may not manifest symptoms of low or high blood glucose due to developmental stage, neurocognitive differences or acute illness.

Professional society guidelines advocate for hospital protocols that enable insulin-treated patients to continue CGM in the hospital, but with the use of confirmatory point-of-care (POC) blood glucose measurements to inform treatment decisions.^12,13 Accelerated by the COVID-19 pandemic,¹⁴ guidelines and practice considerations have emerged surrounding the implementation of CGM in the hospital without mandatory POC confirmation.^6,15–17 Hospital CGM policies have demonstrated feasibility and high acceptance by patients and hospital staff.¹⁸ Importantly, multiple studies have demonstrated CGM accuracy in the adult hospital setting.^8,16,19–24 While pediatric hospital studies show similar accuracy,^25–29 there is a need for additional and larger studies of pediatric hospital CGM usage to ultimately drive the implementation of pediatric CGM policies across more institutions.

Understanding factors that may negatively affect CGM accuracy in the hospital setting is of top importance to inform inpatient CGM policies. CGM accuracy has been preserved in studies examining associations with lower pH and other biomarkers of tissue perfusion in pediatric cohorts;^25,27–29 and with blood pressure alterations²¹ and vasopressor use²⁴ in adult cohorts. However, further study is needed to examine in-hospital accuracy on different sensor wear days³⁰ given the commonly cited concern of poor accuracy during the first day of sensor wear¹⁵ and the potential clinical utility to initiating a sensor during a hospital admission.

Assessments of inpatient CGM accuracy throughout phases of hospitalization²⁷ and during stages of glycemic fluctuation are also crucial to inform hospital CGM accuracy and validation protocols. Patients may be transferred to different hospital areas (e.g. emergency department (ED), intensive care unit (ICU), inpatient floor) as their illness evolves. Glucose levels in the hospital may have more rapid fluctuations given evolving illness states and the use of medications that influence insulin production and sensitivity, as well as altered nutritional regimens (IV nutrition, enteral tubes, etc.). In addition, CGM measures interstitial glucose resulting in known lag time³¹ which is more pronounced during times of rapid glucose changes.

The goal of this study was to address the knowledge gap related to understanding the factors that affect hospital CGM accuracy, specifically in the pediatric population. This study pooled data for a single CGM system (Dexcom G7) from two pediatric centers across different levels of care. We assessed CGM accuracy on different sensor wear days, days of hospitalization, and at times of hypoglycemia, hyperglycemia and rapid glucose rates of change.

METHODS

Study Cohort Criteria and Covariates

Study Cohort Eligibility.

We performed a pooled retrospective cross-sectional study of individuals with diabetes who used Dexcom G7 CGM (DexCom, Inc., San Diego, CA, USA) during an inpatient hospitalization at either Boston Children’s Hospital (BCH – Boston, MA, USA) or the Children’s Hospital of Colorado (Aurora, CO, USA) and shared CGM data with the diabetes programs at BCH or the Barbara Davis Center for Diabetes (BDC - Aurora, CO, USA). This study received ethical approval from the BCH institutional review board (IRB-P00050094) on January 7, 2025 and the BDC institutional review board (IRB-22–2157) November 22, 2022. All information was de-identified with the exception of elements of dates, and consent was not required.

Eligible subjects were between ages 2 and 30 years at the time of hospitalization to the respective pediatric hospitals. Young adults were included in this analysis given that they commonly present to pediatric hospitals for care due to their established relationships with the diabetes teams or other specialists at those institutions. The BCH research database encompassed hospitalizations using Dexcom G7 from May 2023 to June 2024. The BDC research database encompassed hospitalizations using Dexcom G7 from January 2024 to December 2025.

Covariates.

Demographic covariates on the day of hospital admission were abstracted from the medical record. Covariates included: sex, age, race, ethnicity, diabetes type, diabetes duration (using recorded date of diabetes diagnosis), pump use, most recent A1c (measured during or prior to the hospital encounter) and encounter diagnoses. Encounter diagnoses were abstracted from the medical record from ICD10 codes or from admission or discharge documentation.

Hospital Day and Level of Care.

The first hospital day was defined as 0 to 23.99 hours following the date and time of hospital presentation, inclusive of ED presentation. Subsequent hospital days were defined as subsequent, sequential 24 hour periods during hospitalization. The final hospital day was defined from the beginning of that 24 hour period to the date and time of hospital discharge. Hospital level of care was defined as the level of care upon initial presentation to the hospital. If patients were seen and discharged from the ED, this was grouped with ‘floor’ level of care in stratified analyses.

POC Glucose (Reference) Data

Point-of-care (POC) capillary glucose measurements were used as the reference glucose values for this analysis and were abstracted from the electronic health record. The Nova Biomedical StatStrip (Nova Biomedical, Waltham, MA, USA) was the POC glucometer used at both pediatric hospitals (BCH and BDC). This glucometer has a mean absolute relative difference (MARD) of 6.3%³² compared to plasma glucose and has been validated in the pediatric hospital setting.³³

CGM Glucose (Test) Data

CGM System.

The test glucose values assessed in this study were interstitial glucose readings measured using the Dexcom G7 10-Day CGM device. The Dexcom G7 system was FDA approved in December 2022 for outpatient use in patients ages 2 years and older.³⁴ In outpatient studies, the Dexcom G7 had an overall MARD for arm-placed sensors of 8.2% in an adult population³⁵ and 8.1% in a pediatric population aged 7–17 years³⁶ compared to plasma glucose.

CGM Data Extraction and Cleaning.

CGM glucose readings were downloaded from the cloud-based Dexcom Clarity Portals at BCH and BDC. CGM data was gathered from 14 days prior to hospitalization (to identify sensor insertion date) through the end of the discharge date. Sensor glucose readings outside of the reportable range (40–400 mg/dL) and reported as “High” or “Low” were excluded from subsequent analysis.

BCH extracted the CGM sensor glucose values from the downloaded BCH Dexcom Clarity dataset using R (version 3.6.2) on the BCH Enkefalos-v3 (E3) high-performance computing cluster (Linux operating system). BDC extracted sensor glucose values from the raw CGM files using R (version 4.5.1).

CGM Transmitter Identification and Duration.

Individual Dexcom G7 sensors were identified using the unique transmitter ID codes recorded in the downloaded CGM data. As the Dexcom G7 sensor and transmitter are one unified device, identification of a new transmitter code was used as the proxy for a new sensor initiation. We calculated transmitter duration (hours) as the elapsed time from the earliest recorded date and time for a given transmitter ID to the time of each recorded sensor glucose measurement by the same transmitter ID. If interstitial glucose values were measured from two transmitters at overlapping times, the data from the newer transmitter was kept. The data from the older transmitter was filtered out following the earliest date and time of data from the newer transmitter.

CGM Sensor Wear Day.

CGM sensor wear day was defined based on calculated transmitter duration. Transmitter duration between 0 and 23.99 hours was considered wear day 1; 24 - <48 hours was considered wear day 2; 48 - <72 hours was considered wear day 3; etc. Wear day 10 included any transmitter durations 0–4 hours past the 10th day.

CGM Accuracy Outcomes

Glucose Pair Definition.

Each POC glucose (reference value) was paired to the single closest CGM glucose (test value) within five minutes before or after the POC glucose timestamp. POC glucose measurements that were not paired to a CGM glucose were not included.

MARD and MAD.

The average MARD was calculated as the mean of the absolute relative differences across all POC-CGM pairs. The mean absolute difference (MAD) was calculated as the average of the absolute differences between the CGM reading and the reference POC value across all pairs, which was utilized for further assessment of accuracy in the hypoglycemic range. The percentage accuracy assessments of test CGM readings within 15, 20, and 30 mg/dL of reference POC values <= 100 mg/dL or 15, 20, and 30% of reference POC values > 100 mg/dL were calculated, (referred to as %15/15, etc, respectively).

Error Grids.

The Parkes Consensus Error Grid was used to assess CGM accuracy relative to the reference POC glucose. The Parkes Error Grid was previously developed to assess clinical accuracy of blood glucose self-monitoring by assigning error in test versus referent glucose measurements to one of five risk categories, based on survey results from 100 endocrinologists.³⁷ Zone A is defined as no effect on clinical action. Zone B is defined as altered clinical action but with little to no effect on clinical outcome. Zone C is defined as altered clinical action and likely to affect clinical outcome. Zone D is defined as altered clinical action which could have significant medical risk. Zone E is defined as altered clinical action which could have dangerous consequences. The boundaries of the final risk zones were created using iterative averaging and final smoothing by hand. The Parkes Error Grid differs from the Clarke Error Grid in that the survey respondents were allowed to assign test and referent glucose pairs to any zone without restricted quantitative definitions (e.g. the Clarke Zone A was defined as test values within 20% of reference values or both values <70 mg/dL³⁸).

Overall CGM accuracy was assessed, for comparison, with alternative error grids - the Clarke Error Grid³⁸ (with five, discontinuous risk zones based on quantitative glucose thresholds) and the Diabetes Technology Society (DTS) Error Grid³⁹ (five-zone representation, with smoothed zones based on clinician survey perceived clinical risk of discrepancies between reference versus test glucose measurement).

Underestimation or Overestimation.

We also assessed the proportion of CGM values that overestimated or underestimated POC glucose values by more than 20%, the accuracy threshold used clinically at the institutions in this analysis. Overestimation was defined as CGM values >20% above POC glucose values. Underestimation was defined as CGM values >20% below POC glucose values.

Subgroup Analyses of POC-CGM Pairs

Glucose Ranges.

Analyses of accuracy in different reference (POC) glucose ranges were performed. Glucose ranges were defined using ADA criteria for hypoglycemia (<70 mg/dL), level 1 hypoglycemia (<70 mg/dL to ≥ 54 mg/dL), level 2 hypoglycemia (< 54 mg/dL), time in range (TIR; between 70 and 180 mg/dL), level 1 hyperglycemia (>180 mg/dL) and level 2 hyperglycemia (>250 mg/dL).⁴⁰

CGM Glucose Rate of Change.

At the time of each of the paired POC-CGM glucose measurements, the preceding CGM glucose rate of change (mg/dL/minute) was calculated in the 15-minute window prior to the paired CGM reading. We fit a regression line (using R function ‘lm’ from package ‘stats’ version 4.3.2 [BCH] or 4.5.1 [BDC]) to all CGM glucose readings within 15.1 minutes of the paired CGM reading, including the paired CGM reading. The slope of the regression line was used to estimate the CGM glucose rate of change. Rate of change was categorized according to clinically meaningful ranges corresponding to the trend arrows⁴¹ displayed to Dexcom G7 CGM users: flat arrow (−1 to 1 mg/dL/min), slowly falling arrow (between −1 and −2 mg/dL/min), slowly rising (between 1 and 2 mg/dL/min), falling arrow (< −2 mg/dL/min), rising arrow (> 2 mg/dL/min). We plotted MARD by rate of change category using box and whisker plots superimposed with the original data points. There were zero to two individual POC-CGM pairs with MARD beyond 50% in each rate of change category. Therefore, the plot of MARD by rate of change category only included MARD between 0 and 50%, with the full original data displayed in the supplemental materials.

Other Covariates.

We evaluated the accuracy within strata of other covariates including institution, age (2 to <18 years, 18–22 years), acuity (ED/Floor, ICU), hospital day (days 1, 2–4, 5–30) and sensor wear day (days 1, 2–6, 7–10).

Data Availability and Reporting Guidelines

The datasets generated and analyzed during the current study are available from the corresponding author on request. We used the STROBE reporting guideline⁴² to draft this manuscript, and the STROBE reporting checklist⁴³ when editing, included in the supplemental materials.

RESULTS

Study Cohort Characteristics

The study evaluated 554 POC-CGM glucose pairs from a cohort of 68 unique participants with 80 unique hospital encounters from two pediatric hospitals (Table 1). The mean (SD) age at admission was 12.4 (5.8) years and there were 63 encounters with participants aged <18 years old. Participants were 58.8% females, 8.8% Hispanic ethnicity and 66.2% White race. Mean A1c was 8.9 (2.3)% and the majority of participants had type 1 diabetes (91.3%) with diabetes duration 3.0 (4.0) years. Data from 91 CGM sensors was analyzed with an average of 1.3 (1.3) sensors per participant. Assessing the distribution of hospital days with POC-CGM pairs, the median [interquartile range] hospital day was 2 [1–6] and the unique number of hospital days with POC-CGM pairs was 1 [1–2]. Primary or contributing reasons for inpatient encounters include hospitalized for DKA (19% of encounters); 44% for hyperglycemia, hypoglycemia or ketosis; 39% gastrointestinal causes; 13% infectious causes; as well as other causes (Table 2). The demographic and clinical characteristics were similar between institutions.

Table 1.

Demographic and clinical characteristics

Characteristics	Overall	Barbara Davis Center (BDC)	Boston Children’s Hospital (BCH)
Demographics
Unique patients, n	68	29	39
Age, yearsa	12.4 (5.8)	12.6 (5.0)	12.3 (6.4)
Female, n (%)b	40 (58.8%)	14 (48.3%)	26 (66.7%)
Race, n (%)b c
Asian	1 (1.5%)	0 (0%)	1 (2.6%)
Black or African American	14 (20.6%)	7 (24.1%)	7 (17.9%)
Other	9 (13.2%)	1 (3.4%)	8 (20.5%)
White	45 (66.2%)	24 (82.8%)	21 (53.8%)
Missing/Declined	2 (2.9%)	0 (0%)	2 (5.1%)
Hispanic Ethnicity, n (%)b	6 (8.8%)	2 (6.9%)	4 (10.3%)
Diabetes History a
Diabetes duration, years	3.0 (4.0)	3.5 (4.6)	2.6 (3.3)
A1c, %	8.9 (2.3)	8.4 (1.7)	9.4 (2.6)
Diabetes Type, n (%)
T1D	73 (91.3%)	35 (97.2%)	38 (86.4%)
T2D	5 (6.3%)	0 (0%)	5 (11.4%)
Diabetes due to medical condition/medication	1 (1.3%)	1 (2.8%)	0 (0%)
Post-transplant diabetes	1 (1.3%)	0 (0%)	1 (2.3%)
Pump User, n (%)	27 (33.8%)	23 (63.9%)	4 (9.1%)
Encounters, Sensors, & POC-CGM Pairs
POC-CGM glucose pairs, n	544	258	286
Glucose pairs per participant	8.0 (11.8)	8.9 (12.1)	7.3 (11.7)
Sensors, n	91	40	51
Sensors per participant	1.3 (1.3)	1.4 (0.9)	1.3 (1.6)
Hospital encounters, n	80	36	44
Hospital encounters per participant	1.2 (0.8)	1.2 (0.7)	1.1 (0.8)
Median hospital encounter day with POC-CGM pair; Median [IQR]d	2 [1 – 6]	1 [1 – 2]	4 [1 – 16.8]
Unique number of hospital encounter days with POC-CGM pair; Median [IQR]d	1 [1 – 2]	1 [1 – 1.3]	1 [1 – 2.3]

Data are presented as count (%) or mean (standard deviation) unless otherwise indicated.

^aReported at the encounter level.

^bReported at the participant level.

^cRespondents were able to select more than one race option at the BDC, so values may not add up to 100%.

^dReported at the participant-encounter level

POC, point-of-care; CGM, continuous glucose monitor; IQR, interquartile range.

Table 2.

Admission diagnoses of hospital encounters

Diagnosis categorya	Number of Admissionsb
Diabetes Mellitus	48
Diabetic ketoacidosis	15
Ketosis	6
New Onset Diabetes	9
Hyperglycemia	26
Hypoglycemia	3
Cardiovascular (Tachycardia, ectopic tachycardia, hypertension)	5
Endocrine (Hyponatremia)	1
Gastrointestinal (Abdominal pain, dehydration, gastroenteritis, vomiting, poor oral intake, hyperbilirubinemia, constipation, abnormal celiac panel)	31
Infectious (Viral illness, pharyngitis, fever, mononucleosis, COVID-19, acute otitis)	10
Neurological (Seizure, PNES, hypothermia, headache, lethargy, agitation, cerebral palsy)	9
Orthopedic (Joint pain/instability, limb pain, hip dysplasia, post-operative pain)	3
Psychiatric/Toxicology (Accidental overdose, depression, hallucination, suicidal ideation, social concern)	7
Renal/Genitourinary (AKI, urosepsis, nephrolithiasis)	5
Respiratory/Otolaryngology (Respiratory distress/failure, apnea, hemoptysis, pneumonia, eustachian tube dysfunction)	6

^aGeneral diagnosis categories listed with examples of specific diagnoses provided in italics

^bPotentially multiple diagnoses per admission and per diagnosis category, total number of diagnoses exceeds number of encounters reported (N=80).

PNES, psychogenic nonepileptic seizure; AKI, acute kidney injury.

Overall CGM Accuracy

The overall MARD was 12.5% (Table 3 and Supplemental Table 1). In a Parkes Error Grid analysis, the vast majority (99.1%) of POC-CGM pairs were in the low-risk A and B zones (Table 3 and Figure 1A). The percent of pairs in low-risk A and B zones was similarly favorable using the Clarke Error Grid (98.6%) and the DTS Error Grid (99.1%), with a side-by-side comparison of Error Grids provided in Figure 1. Overall agreement rates for %15/15, %20/20, and %30/30 accuracy measures were high at 70.2%, 82.5% and 93.4%, respectively. By hospital, the overall MARD was 12.3% at BDC and 12.7% at BCH (Supplemental Table 1). Accuracy analyses by age subgroups showed the same results in the pediatric (2 to <18 years) and young adult (18–22 years) cohorts (MARD of 12.5% and 12.6%, respectively, Supplemental Table 1).

Table 3.

CGM accuracy statistics by day of wear, days of hospitalization, and level of care.

Setting	N	MARD (%)	Within A and B zones (%)	%15/15 (%)	%20/20 (%)	%30/30 (%)
Overall	544	12.5	99	70.2	82.5	93.4
CGM Wear Day
Wear Day 1	113	12.3	99	70.8	83.2	92.9
Wear Days 2–6	282	13.6	99	68.4	80.5	91.1
Wear Days 7–10	149	10.6	99	73.2	85.9	98.0
Hospitalization Day a
Hospital Day 1 (first 24 hours)	250	13.7	99	62.8	79.2	93.2
Hospital Days 2–4	142	10.4	100	81.0	91.5	95.1
Hospital Days 5–30	90	12.5	97	74.4	85.6	90.0
Level of care
ED and Floor	410	12.5	99	70.7	81.5	93.7
ICU	134	12.8	98	68.7	85.8	92.5

^aPOC-CGM pairs from day 31 of hospitalization or later were excluded from this analysis.

MARD, mean absolute relative difference; CGM, continuous glucose monitor.

Figure 1.

Accuracy of overall POC-CGM glucose pairs using the Parkes Consensus Error Grid (A), the Clarke Error Grid (B) and the Diabetes Technology Society Error Grid (C). CGM, continuous glucose monitor; POC, point-of-care.

CGM Accuracy by Sensor Wear Day

CGM accuracy on different sensor wear days was assessed by the approximated duration of G7 sensor wear. Compared to the overall MARD of 12.5%, MARD on wear day 1 was 12.3%, wear days 2–6 was 13.6% and wear days 7–10 10.6% (Table 3). The percentage of values within the low-risk zones (zones A and B) was consistently very high at 99% regardless of wear day (Table 3).

CGM Accuracy by Level of Care and Hospital Day

Accuracy analyses were also stratified by level of care and hospital day. CGM accuracy was similar by level of care, with MARD for POC-CGM pairs from those admitted to the ICU 12.8% versus 12.5% for floor or ED encounters (Table 3). MARD on hospital day 1 was 13.7%, hospital days 2–4 was 10.4% and days 5–30 was 12.5% (Table 3).

Assessing accuracy by the intersection of sensor wear day and hospital day (Table 4), there was no difference in MARD between wear day 1 and wear day 2–6 when examining just hospital day 1 (14.6% vs 14.3%) or just hospital days 2–4 (10.8% vs 10.6%). For POC-CGM pairs from later in the hospitalization (hospital days 5–30), MARD increased from wear day 1 (11.1%) to wear days 2–6 (16.0%), though there were limited pairs on hospital days 5–30 and wear day 1 (N=13). For any given segment of the hospitalization, MARD was lowest on wear day 7–10 (12.3% hospital day 1, 9.0% hospital days 2–4, 7.4% hospital days 5–30), though there were a smaller number of POC-CGM glucose pairs on wear day 7–10 (N= 76, 20, 29 respectively).

Table 4.

CGM accuracy statistics by hospital day and day of wear

Hospital Daya	Wear Day	N	MARD (%)	Within A and B zones (%)	%15/15 (%)	%20/20 (%)	%30/30 (%)
Hospital Day 1 (first 24 hours)	Wear Day 1	40	14.6	100	52.5	75.0	90.0
	Wear Day 2–6	134	14.3	99	64.9	78.4	91.8
	Wear Day 7–10	76	12.3	99	64.5	82.9	97.4
Hospital Days 2–4	Wear Day 1	51	10.8	100	84.3	90.2	94.1
	Wear Day 2–6	71	10.6	100	78.9	91.5	94.4
	Wear Day 7–10	20	9.0	100	80.0	95.0	100.0
Hospital Days 5–30	Wear Day 1	13	11.1	92	84.6	92.3	92.3
	Wear Day 2–6	48	16.0	96	64.6	79.2	83.3
	Wear Day 7–10	29	7.4	100	86.2	93.1	100.0

^aPOC-CGM pairs from day 31 of hospitalization or later were excluded from this analysis.

CGM, continuous glucose monitor; MARD, mean absolute relative difference

CGM Accuracy by Glucose Ranges

When stratifying by different glucose ranges (Table 5), with POC glucose in the hypoglycemia range (<70 mg/dL) the MAD was 8.6% with a MARD of 14.5%. In the target range (70–180 mg/dL), MARD was 13.9%. In the level 1 hyperglycemia range (>180 mg/dL), MARD was 11.1% and in level 2 hyperglycemia range (>250 mg/dL) MARD was 10.2%.

Table 5.

CGM accuracy statistics by POC glucose range

POC Glucose Range	N	MARD (%)	MAD (mg/dL)	Within A and B zones (%)	%15/15 (%)	%20/20 (%)	%30/30 (%)
Hypoglycemia (<70 mg/dL)	38	14.5	8.6	95	86.8	94.7	94.7
Level 1 Hypoglycemia (<70 - ≥54 mg/dL)	32	12.2	7.8	97	90.6	96.9	96.9
Level 2 Hypoglycemia (<54 mg/dL)	6	26.7	13.2	83	66.7	83.3	83.3
In Range (70–180 mg/dL)	240	13.9	NA	99	63.8	77.1	90.0
Hyperglycemia (>180 mg/dL)	266	11.1	NA	99	73.7	85.7	96.2
Hyperglycemia (>250 mg/dL)	110	10.2	NA	100	74.5	88.2	97.3

CGM, continuous glucose monitor; POC, point-of-care; MARD, mean absolute relative difference.

Assessing the proportion of CGM values that overestimated or underestimated POC glucose values (Table 6), 80.5% of paired CGM glucose levels were within 20% of POC glucose, and this was similar across POC glucose ranges (74.6–88.2%). In the hypoglycemia range, CGM values beyond 20% of POC glucose (18.4% of pairs) overestimated POC glucose with no CGM values underestimating POC glucose. In the target range and hyperglycemia ranges, CGM values beyond 20% of POC glucose overestimated POC glucose (8.2–17.9% of pairs) with a minority of CGM values underestimating POC glucose (2.6–7.5% of pairs). This analysis included all POC-CGM pairs without considering glucose rate of change (see rate of change stratified analysis).

Table 6.

Percent of pairs where CGM is beyond 20% higher or lower than POC

POC Glucose Range	N	CGM is within 20% of POC	CGM is 20% higher than POC	CGM is 20% lower than POC
Overall	544	438 (80.5%)	81 (14.9%)	25 (4.6%)
Hypoglycemia (<70 mg/dL)	38	31 (81.6%)	7 (18.4%)	0 (0%)
TIR (70–180 mg/dL)	240	179 (74.6%)	43 (17.9%)	18 (7.5%)
Hyperglycemia (>180 mg/dL)	266	228 (85.7%)	31 (11.7%)	7 (2.6%)
Hyperglycemia (>250 mg/dL)	110	97 (88.2%)	9 (8.2%)	4 (3.6%)

CGM, continuous glucose monitor; POC, point-of-care; TIR, time-in-range.

CGM Accuracy by Glucose Rate of Change

Finally, glucose pairs were stratified by the CGM glucose rate-of-change in the 15-minute window prior to the POC-CGM pair (Figure 2 and Supplemental Table 2). For CGM rates of change that would correspond to ‘flat’ Dexcom trend arrow definitions⁴¹ (less than 1 mg/dL/min in either direction), CGM MARD was between 11.5–11.9%. For a ‘slowly falling’ rate of change (between −1 and −2 mg/dL/min) MARD was 11.7%. For a ‘slowly rising’ rate of change (between 1 and 2 mg/dL/min) MARD was 17.3%. For ‘falling’ rates of change (< −2 mg/dL/min) MARD was 18.4% and for ‘rising’ rates of change (> 2 mg/dL/min) MARD was 16.6%. Even at times of falling and rising CGM glucose, the percent of POC-CGM pairs within the low-risk A and B zones were still very high at 94–95% (Supplemental Table 2).

Figure 2.

Accuracy by CGM glucose rate of change. Blue points represent individual POC-CGM glucose pair absolute relative differences. Black points represent MARD for each rate category and black error bars represent standard error. The MARD range was restricted to 0 – 50% for improved visualization of individual data points, with the full data provided in Supplemental Figure 2. CGM rate of change was calculated in the 15-minute window prior to the paired CGM reading. Arrows above graph represent equivalent Dexcom trend arrows for each rate of change category (left to right): falling (< −2 mg/dL/min), slowly falling (−2 to −1 mg/dL/min), flat (−1 to 1 mg/dL/min), slowly rising (1 to 2 mg/dL/min) and rising (> 2 mg/dL/min). MARD, mean absolute relative difference; CGM, continuous glucose monitor.

DISCUSSION

Our data indicated favorable accuracy of the Dexcom G7 CGM among children and young adults in the real-world hospital setting. The overall MARD of 12.5% was within the common, albeit informal, inpatient accuracy target⁴⁴ of <14% on wear day 1, on hospital day 1 and at any level of care. There is a need for consensus accuracy standards for CGM in the hospital,^6,17 as the only regulatory standards to-date are for ambulatory CGM use.^45,46 Our measured accuracy of 82.5% within %20/20 approached the FDA ambulatory integrated CGM accuracy requirement (overall >87% of measurements within 20% of reference).⁴⁵ Our measured accuracy was in line with published data in adult hospital cohorts using G7 (adult ICU 12%^47,48 and non-ICU 13%⁴⁸ post-op MARD 9.0%⁴⁹). Our overall MARD was higher than a recently published inpatient pediatric cohort (Dexcom G7 MARD 9.8%),⁵⁰ though their cohort had fewer participants in the G7 group with more POC-CGM pairs per participant (26.6 vs 8.0 pairs in our cohort), which may have reduced measurement variability. Nonetheless, our findings provide evidence that inpatient pediatric CGMs can provide accurate glucose measurements which can be clinically useful, even at the beginning of sensor wear and during the first days of hospital encounters.

Glucose rate of change is a crucial factor contributing to CGM accuracy and must be taken into consideration when utilizing CGM data in the hospital setting. MARD increased to 16.6% or higher during rapid of glycemic fluctuation, most notably with slowly rising and rising or falling CGM values. As lag time cannot be distinguished from accuracy, these results show that CGM-measured interstitial glucose is a less effective proxy for blood glucose when there is insufficient time for equilibration with glycemic fluctuation.

While the sample sizes for this sub-analysis were small, this highlights the importance of incorporating rate of change – using readily available CGM trend arrows – into inpatient CGM policies for accuracy assessments and clinical decision making. For inpatient protocols that include CGM validation, accuracy should be verified at times when the CGM trend arrow is ‘flat’ (rate of change between −1 to 1 mg/dL/minute), corresponding to the highest POC-CGM glucose accuracy in our dataset, which is in line with guidelines for CGM sensor calibration.⁵¹ When inpatient providers clinically assess CGM-measured hypoglycemia or hyperglycemia, verifying with blood glucose may be considered in conjunction with the clinical picture, particularly if the CGM trend arrow is not flat at the time of assessment, consistent with expert consensus.¹⁷

In our data, the MARD improved as glucose increased from hypoglycemic to hyperglycemic ranges, concordant with the reference data for Dexcom G7 accuracy in the pediatric population.³⁶ Thus, our findings redemonstrate that Dexcom G7 CGM has the best accuracy in the hyperglycemia range, even in the hospital setting. Even in the hypoglycemia range (POC glucose <70 mg/dL), our calculated MAD of 8.6% in the inpatient setting was comparable to the reported MADs in the reference ambulatory dataset of 11.3% (CGM range 40–60 mg/dL) and 6.4% (CGM range 61–80 mg/dL).³⁶

A common CGM accuracy threshold used in the inpatient setting¹⁸ (including at BDC and BCH) is to verify that CGM glucose is within 20% of POC glucose. Most POC-CGM glucose pairs (80.5% overall) met this accuracy standard, including 81.6% of pairs with POC glucose <70 mg/dL, and 85–88% of pairs with POC glucose >180 mg/dL or >250 mg/dL. Importantly, when POC glucose was in the hypoglycemia range, all CGM values outside of the 20% accuracy threshold overestimated POC glucose values (18.4% of pairs). This indicates a risk of delayed CGM detection of hypoglycemia. Our measured MARD in the hypoglycemia range of 14.5% was just above the inpatient target of <14%, and was more favorable than MARD during hypoglycemia of 15.17% and 28.8% reported in other pediatric studies of Dexcom G6 and G7 sensors.^26,50 Nonetheless, the risk of delayed CGM detection of hypoglycemia – particularly when compounded with higher MARD at times of falling trend arrows – highlights that clinicians cannot rely upon CGM to promptly detect hypoglycemia. Clinicians should instead use POC glucose testing in clinical scenarios when there is increased concern or risk of hypoglycemia (e.g., clinical status change, interruption of intravenous or enteral sources of dextrose, symptoms of hypoglycemia, etc.) or when falling CGM trend arrows raise concern for impending hypoglycemia.

Inpatient CGM accuracy at the beginning of sensor wear was similar to overall accuracy, with MARD on wear day 1 of 12.3% versus 12.5% overall. Likewise, when accounting for hospital day, accuracy on wear day 1 was comparable to accuracy on wear days 2–6. Accuracy at the end of sensor wear was the best, with MARD 10.6% on wear days 7–10. Thus, this real-world data demonstrated that day of sensor wear did not alter sensor accuracy in the hospital setting. This is encouraging data for many hospital scenarios in which a new sensor may be placed (as sensor initiation or replacement) or an older sensor (wear days 7–10) may be already in place and used for glucose monitoring.

Hospital day contributed greater variability to CGM accuracy as compared to sensor wear day. When assessing sensors during wear day 1, MARD was higher on hospital day 1 (14.6%) versus hospital days 2–4 (10.8%). This improvement in accuracy from admission to hospital days 2–4 is aligned with presumed improvements in underlying illness state or stabilization of the clinical management plan. The assessment of accuracy by specific sensor wear days and hospital days is limited by the smaller sample size in these sub-groupings.

In comparing these inpatient findings to the reference data for G7 accuracy in ambulatory pediatric patients,³⁶ it is important to consider key differences in the clinical setting and acquisition of reference glucose values. In the reference data, the primary reported cohort were healthy children aged 7–17 years monitored in an outpatient research clinical setting. Reference glucose values were measured using arterialized venous blood drawn from an IV. In contrast, our real-world, inpatient cohort encompassed multiple additional sources of variability, including a broader age range (ages 2–22), illness requiring hospitalization and reference glucose measured via capillary glucometer which itself has a MARD of 6.3%.³² These factors may explain the differences between our findings, including our overall MARD of 12.5% compared to the ambulatory reference data overall MARD of 8.1% (arm) and 9.0% (abdomen). Surprisingly, our reported MARD of 12.3% on wear day 1 compares favorably to the reference data wear day 1 MARDs of 11.7% (arm) or 11.1% (abdomen), though the G7 algorithm may have been updated to improve accuracy during wear day 1 between publication of the reference data and collection of data for this study.

Strengths of this cross-sectional analysis include the collaboration between two academic medical centers with large pediatric diabetes programs and established hospital CGM policies. Our findings demonstrated effective use of CGM in the inpatient setting that can be generalized to other pediatric hospital settings with the support of CGM policies and resources. Limitations include the retrospective study design, though a large number of matched pairs were available for analysis. A capillary POC glucometer was used as the reference glucose, which itself is a device with inherent measurement error (MARD 6.3%³²). Multiple POC glucose measurements per CGM glucose timepoint (for averaging to improve accuracy of the referent POC glucose estimate) were unavailable in this retrospective analysis. CGM sensor wear time was approximated using the elapsed time from the earliest recorded reading for a given sensor. We were unable to assess for any potential missing data between new sensor insertion and first recorded sensor readings. Body site of CGM placement was not recorded and could not be adjusted for in this retrospective analysis, and thus future work should assess for differences in Dexcom G7 sensor accuracy by site of placement.

We were also unable to account for medications used during hospital encounters, which may differ from common outpatient medications. While we did compare accuracy by level of care, we did not directly account for admission diagnoses or illness acuity, which can vary by level of care and throughout hospitalization. Future studies should examine these additional factors that may affect inpatient CGM accuracy. In this retrospective study, we did not quantify or characterize delayed CGM detection of hypoglycemia, nor did we study the utility of falling CGM trend arrows to alert of impending hypoglycemia, which should be prospectively assessed in future studies. It is critical to note that CGM accuracy directly impacts the safety of insulin delivery via automated insulin delivery (AID) systems, and future work should jointly assess the safety and reliability of AID usage in the hospital setting. Finally, as pediatric hospitals implement CGM policies, future studies should examine any potential adverse events or limitations to the use of CGM for inpatient medical decision making.

CONCLUSIONS

In the pediatric hospital setting, Dexcom G7 CGM demonstrated favorable accuracy with an overall MARD of 12.5%, indicating the reliability of CGM data when used judiciously in the hospital setting. Accuracy was comparable between sensor wear days, hospital days, level of care and across age groups. Importantly, CGM overestimated POC glucose in the hypoglycemia range, and accuracy was lessened at times of rising and falling CGM rate of change, highlighting the importance of integrating trend arrow information and lag time into inpatient CGM policies to guide sensor accuracy validation and clinical decision making.

Data availability:

The datasets generated during and analyzed during the current study are available from the corresponding author on request.