Accuracy of a Real-Time Continuous Glucose Monitor in Pediatric Diabetic Ketoacidosis Admissions

Abstract

Objective:

Continuous glucose monitoring (CGM) devices are integral in the outpatient care of people with type 1 diabetes, although they lack inpatient labeling. Food and Drug Administration began allowing inpatient use during the coronavirus disease 2019 (COVID-19) pandemic, with some accuracy data now available, primarily from adult hospitals. Pediatric inpatient data remain limited, particularly during diabetic ketoacidosis (DKA) admissions and for patients receiving intravenous (IV) insulin.

Design and Methods:

This retrospective chart review compared point-of-care glucose values to personal Dexcom G6 sensor data during pediatric hospitalizations. Accuracy was assessed using mean absolute relative difference (MARD), Clarke Error Grids, and the percentage of values within 15/20/30% if glucose value >100 mg/dL and 15/20/30 mg/dL if glucose value ≤100 mg/dL.

Results:

Matched paired glucose values (N = 612) from 36 patients (median age 14 years, 58.3% non-Hispanic White, 47.2% male) and 42 inpatient encounters were included in this subanalysis of DKA admissions. The MARDs for DKA and non-DKA admissions (N = 503) were 11.8% and 11.7%, with 97.6% and 98.6% of pairs falling within A and B zones of the Clarke Error Grid, respectively. Severe DKA admissions (pH <7.15 and/or bicarbonate <5 mmol/L) had a MARD of 8.9% compared to 14.3% for nonsevere DKA admissions. The MARD during administration of IV insulin (N = 266) was 13.4%.

Conclusions:

CGM accuracy is similar between DKA and non-DKA admissions and is maintained in severe DKA and during IV insulin administration, suggesting potential usability in pediatric hospitalizations. Further study on the feasibility of implementation of CGM in the hospital is needed.

Introduction

Type 1 diabetes (T1D) is characterized by autoimmune destruction of the insulin-producing beta cells of the pancreas, requiring exogenous insulin administration and frequent glucose monitoring. Continuous glucose monitors (CGMs) have led to significant advancements in diabetes care by decreasing the need for fingerstick glucose testing, improving the quality of life for youth with T1D and their caregivers,^1,2 increasing the glucose data available for medication management, and significantly improving hemoglobin A1c.³ Despite the proven benefits of CGM use in the outpatient management of T1D, these devices currently lack Food and Drug Administration (FDA) labeling for inpatient use, despite American Diabetes Association recommendations,⁴ and many hospitals continue to require point-of-care (POC) fingerstick glucose values for the purposes of insulin dosing.

As part of the response to coronavirus disease 2019 (COVID-19), the FDA issued an enforcement discretion for the use of CGM in the hospital. Many adult hospitals permitted inpatient CGM use,^5–8 resulting in some evidence regarding inpatient CGM accuracy. While the accuracy of CGMs for inpatient has previously been shown to be lower compared to outpatient studies,⁹ inpatient adult results have shown promise for CGMs being a reasonable alternative to standard of care.^6–8 However, many pediatric hospitals in the United States have continued to rely solely on fingerstick blood glucose monitoring. This has led to limited data in the inpatient pediatric population, particularly for those admitted for diabetic ketoacidosis (DKA).

DKA is a frequent cause of hospitalization for youth with T1D and often occurs at presentation of previously undiagnosed T1D.¹⁰ In people with previously diagnosed diabetes, DKA occurs due to insulin deficiency, resulting from insulin pump or infusion set malfunction or missed or insufficient insulin administration through injections.^11,12 This life-threatening condition is characterized by metabolic acidosis from ketone production and moderate-to-severe dehydration from hyperglycemia-induced osmotic diuresis. Without prompt treatment, the risk of morbidity and mortality is high, particularly from cerebral injury secondary to cerebral edema.¹² The mainstays of DKA management are intravenous fluid and insulin administration.¹³ However, hyperglycemia generally improves more rapidly than the ketosis, necessitating dextrose intravenous (IV) fluids to maintain safe glucose levels until the acidosis resolves. Due to the complexity of the pathophysiology and treatment for this condition, hourly glucose checks are required for frequent adjustments in insulin infusion rate, IV fluid rate, and dextrose concentration.

There have been minimal data available regarding CGM accuracy in pediatrics during DKA, severe acidosis, or IV insulin administration where physiology is complex and glucose variability is high, however, frequent glucose monitoring is especially critical. As CGMs measure glucose levels from the interstitial fluid rather than plasma, these devices are known to have some lag time. This lag time can be increased during periods of rapid glucose fluctuations^14,15 and potentially during times of impaired perfusion¹⁶; however, the impact of this on sensor accuracy during critical illness in youth is not clear.

This study included a large retrospective chart review, examining accuracy of the Dexcom G6 CGM in pediatric patients with T1D, who were admitted to Children's Hospital Colorado (CHCO). Previous accuracy analysis for all hospital encounters based on the mean absolute relative difference (MARD) was comparable to or better than published adult hospital data.¹⁷ In this report focused on accuracy during hospital admissions for DKA, we hypothesized that accuracy will be maintained, even during periods of acidosis, while receiving IV insulin.

Research Design and Methods

Study design

This study was a retrospective chart review involving children and adolescents (<18 years of age) hospitalized at one of CHCO centers. This study was approved by the Colorado Multiple Institutional Review Board. Inclusion criteria comprised a clinical diagnosis of T1D, previously receiving diabetes-related care at the Barbara Davis Center for Diabetes at the University of Colorado, and a hospitalization or emergency department visit through the CHCO system between July 2020 and October 2022. Exclusion criteria included a lack of real-time data during hospitalization from the patient's G6 rtCGM system (Dexcom). Only home sensors were used, and no new sensor was placed by staff during the study period.

The electronic medical record was used to obtain demographic information, location of admission (intensive care unit [ICU] vs. medical floor/emergency department), primary admission diagnosis, and POC glucose values. Initial laboratory evaluations, including pH, bicarbonate, and beta-hydroxybutyrate, were obtained when available. For patients who received IV insulin, the initiation and discontinuation times were recorded.

DKA management at CHCO

Patients were diagnosed with DKA based on standard clinical criteria, including the presence of a pH <7.3 and/or bicarbonate <15 mmol/L. Severe DKA was defined as a pH <7.15 and/or bicarbonate <5 mmol/L based on CHCO protocol indicating the need for ICU admission.¹⁸ Per CHCO protocol, all patients meeting criteria for DKA must remain on IV insulin until acidosis resolves, as defined by a bicarbonate ≥18 mmol/L and/or beta-hydroxybutyrate <1 mmol/L. At this time, patients are transitioned to either insulin injections or their home insulin pump.

Glucose monitoring

POC glucose data

POC glucose measures were obtained using Nova Biomedical StatStrip (MARD 6%),¹⁹ which was specifically developed for use in critically ill patients and used by CHCO in the ICUs and medical floors. The frequency of POC glucose data collection was based on standard of care protocols throughout the hospitalization.²⁰ POC glucose data, including time obtained, were collected from the electronic medical record.

Continuous glucose monitor

The factory-calibrated Dexcom G6 CGM was used for this analysis as it is our center's most worn sensor in the outpatient setting, accounting for 95.2% of outpatient CGM use. Of patients with known T1D admitted to the hospital during the study period, Dexcom G6 was used in over 97% of cases. Due to the lack of sufficient patients on other sensors admitted to the hospital in DKA, analysis was limited to patients on the Dexcom G6 system. This device was approved by the FDA in 2018 for outpatient, nonadjunctive use. The G6 has a MARD of 7%–10% in outpatient studies.^9,21,22 Comma-separated values (CSV) data were obtained from the Dexcom Clarity cloud system, which the patients had previously granted access to the Barbara Davis Center as part of routine clinical care.

POC: CGM matched glucose pairs

Any POC value over 400 mg/dL was excluded from the analysis due to limitations of the G6 reporting capabilities. The POC glucose was included for analysis if there was a recorded CGM value occurring within 5 min of the POC glucose measurement. Matched pairs were additionally classified as occurring while the patient was receiving IV insulin or while receiving insulin injections or home pump use.

Statistical analysis

MARD was used as the primary accuracy measure. MARD is a standard method of comparing reference glucose and CGMs.²³ Participants with DKA were compared to those without DKA, and participants with severe DKA were compared to those with nonsevere DKA, using two-sample t-tests.

Matched glucose pairs were also plotted on Clarke Error Grids.²⁴ Zone A includes values that fall within 20% of the reference glucose level. Zone B includes values outside of 20%, but do not result in inappropriate treatment (benign errors). Zone C includes overcorrection errors with treatment leading to hypoglycemia or hyperglycemia. Glucose in Zone D indicates potentially dangerous failure to detect hypoglycemia or hyperglycemia. Zone E includes glucose points that would confuse hypoglycemia for hyperglycemia or vice versa. Calculation of the percentage of values that fall within Zones A and B, which are deemed clinically acceptable, is commonly reported.

%15/15, %20/20, and %30/30: These measures calculate the percentage of values within 15% (20% or 30%) of POC glucose if >100 mg/dL and within 15 mg/dL (20 or 30 mg/dL) if POC glucose is ≤100 mg/dL. This is a commonly used measure for CGM accuracy studies.^25,26

Each analysis was performed for the overall sample, for those matched glucose pairs when the patient was on IV insulin, and for those obtained while receiving insulin injections or utilizing their home pump. In addition, to account for the period of increased glucose variability and changing degree of acidosis following initiation of IV insulin, further analysis of the matched pairs within the first 3 h, 3 to <6 h, and ≥6 h following initiation of IV insulin was performed.

Results

Data for the larger study were available from 83 unique patients with T1D from 100 unique hospital encounters. The median age for the entire sample was 12.0 years, 54% were male, and 68.7% identified as non-Hispanic White. Data for the DKA analysis were obtained from 36 patients from 42 unique admissions due to DKA. For those not admitted in DKA in the larger study, primary diagnoses were neurologic, infectious, gastrointestinal, mental health, ENT, and allergic. The mean age for those admitted with DKA was 14.0 years, 47.2% were male, and 58.3% identified as non-Hispanic White. For patients admitted with DKA, the mean initial beta-hydroxybutyrate level was 5.47 mmol/L (standard deviation [SD] 2.63 mmol/L) with a mean pH of 7.21 (SD 0.12). Full demographic information is included in Table 1.

Table 1.

Demographics

	Overall sample	DKA admissions
Patients	83	36
Age, years, median [IQR]	12.0 [8.0, 15.0]	14.0 [10.8, 16.0]
Gender, male, n, %	45 (54.2)	17 (47.2)
Race, non-Hispanic White, n (%)	57 (68.7)	21 (58.3)
Type 1 diabetes duration, years, median [IQR]	7.6 [4.7, 10.9]	7.0 [4.7, 9.7]
HbA1c, %, median [IQR]	8.5 [7.6, 10.5]	10.2 [8.3, 10.8]
Pump user, n (%)	51 (61.4)	23 (63.9)

DKA, diabetic ketoacidosis; HbA1c, hemoglobin A1c.

CGM accuracy

DKA admissions

In total, 1115 matched pairs were available from all hospital encounters, with DKA admissions accounting for 612 pairs. The MARD from the DKA admissions was 11.8% (N = 612) compared to 11.7% from non-DKA admissions (N = 503). When examining severe DKA [pH <7.15 and/or bicarbonate <5 mmol/L] (N = 288) compared to nonsevere DKA (N = 324), the MARD was significantly better (8.9% vs. 14.3%, respectively, P = 0.004).

For patients in DKA, 53.0% of values were within 15/15, 67% were within 20/20, and 83.6% were within 30/30, compared to 60.4% within 15/15, 73.6% within 20/20, and 88.3% within 30/30 for patients not in DKA. See Table 2 for full summary.

Table 2.

Accuracy Statistics for Diabetic Ketoacidosis Versus Nondiabetic Ketoacidosis Admissions

	MARD	% within A and B Zones	% 15/15	% 20/20	% 30/30
DKA (N = 612)	11.8%	97.6%	53.0%	67.0%	83.6%
Non-DKA (N = 503)	11.7%*	98.6%	60.4%	73.6%	88.3%
Severe DKA (N = 288)	8.9%	98.3%	50.4%	66.5%	85.3%
Nonsevere DKA (N = 324)	14.3%**	96.9%	55.3%	67.3%	82.0%
IV insulin (N = 266)	13.4%	98.1%	50.6%	66.4%	85.0%
Subcutaneous insulin (N = 346)	10.5%	97.1%	54.9%	67.4%	82.4%

^* P-value 0.95 (DKA vs. non-DKA).

^** P-value 0.004 (severe DKA vs. nonsevere DKA).

IV, intravenous; MARD, mean absolute relative difference.

We performed Clarke Error Grid analysis with the matched pairs. For those admitted with DKA, 61.2% fell within zone A (clinically accurate), 36.4% in zone B (benign errors), 1.1% in zone C (potential for overcorrection), 1.1% in zone D (dangerous failure to detect hypoglycemia or hyperglycemia), and 0.2% in zone E (mistaking hypoglycemia for hyperglycemia or vice versa).

For those admitted for conditions other than DKA, 67.8% fell in zone A, 30.8% in zone B, 0.2% in zone C, 1.2% in zone D, and none in zone E. For those admitted with severe DKA, 60.4% fell in zone A, 37.8% in zone B, 0.4% in zone C, 1.4% in zone D, and none in zone E. For those admitted with nonsevere DKA, 61.7% fell in zone A, 35.2% in zone B, 1.9% in zone C, 0.9% in zone D, and 0.3% in zone E. Therefore, under all circumstances, 97% or more matched pairs were within zones A and B (Fig. 1).

FIG. 1.

Clarke error grids for DKA versus non-DKA admissions. DKA, diabetic ketoacidosis.

IV insulin administration

Median IV insulin duration was 14.9 h [IQR: 11.1, 20.7 h]. The MARD when receiving IV insulin was 13.4% (N = 266) compared to 10.5% for the same patients following transition to subcutaneous insulin at the resolution of acidosis (N = 346). For the Clarke Error Grid analysis for data points during IV insulin infusion, 61.3% of matched pairs fell within zone A, 36.8% within zone B, 1.9% in zone C, and none in zones D and E. When not receiving IV insulin, 61.0% of matched pairs fell within zone A, 36.1% within zone B, 0.6% in zone C, 2.0% in zone D, and 0.3% in zone E (Fig. 2).

FIG. 2.

Clarke error grids for pediatric patients admitted for DKA, while on and off IV insulin. IV, intravenous.

When receiving IV insulin, 50.6% of values were within 15/15, 66.4% were within 20/20, and 85.0% were within 30/30. When not receiving IV insulin, 54.9% of values were within 15/15, 67.4% were within 20/20, and 82.4% were within 30/30.

Additional analysis was performed for the matched pairs obtained during IV insulin administration based on the time on the drip. The MARD for 0 up to 3 h on the insulin drip was 13% (N = 68), compared to 11.8% (N = 64) between 3 and up to 6 h, and 14.3% for ≥6 h after initiation of IV insulin (N = 134).

For the Clarke Error Gride analysis, 60.3% of values between 0 and up to 3 h on the insulin drip fell within zone A, 35.3% in zone B, 4.4% in zone C, and none falling within zones D or E. For pairs obtained between hours 3 and up to 6 on the insulin drip, 64.1% fell within zone A, 35.9% within zone B, and none in zones C, D, or E. For pairs obtained ≥6 h on the insulin drip, 60.4% of values fell within zone A, 38.1% in zone B, 1.5% in zone C, with none falling in zones D or E (Fig. 3).

FIG. 3.

Clarke error grids for 0 to <3 h, 3 to <6 h, and ≥6 h on IV insulin.

In the first 3 h on IV insulin, 50% of values were within 15/15, 66.1% were within 20/20, and 77.4% were within 30/30. From 3 up to 6 h on IV insulin, 62.1% of values were within 15/15, 70.7% were within 20/20, and 93.1% were within 30/30. After 6 or more hours on IV insulin, 45.7% of values were within 15/15, 64.6% were within 20/20, and 85% were within 30/30. See Table 3 for full summary.

Table 3.

Accuracy Statistics While on Intravenous Insulin

	MARD	% within A and B zones	% 15/15	% 20/20	% 30/30
Overall IV insulin (N = 266)	13.4%	98.1%	50.6%	66.4%	85.0%
IV insulin 0–<3 h (N = 68)	13.0%	95.6%	50.0%	66.1%	77.4%
IV insulin 3–<6 h (N = 64)	11.8%	100%	62.1%	70.7%	93.1%
IV insulin ≥6 h (N = 134)	14.3%	98.5%	45.7%	64.6%	85.0%

Conclusions

Our data show that the accuracy for a real-time CGM worn during pediatric T1D DKA admissions was consistent with previously published adult inpatient data.^6–8 While accuracy was lower during IV insulin infusions, it is still similar to prior published studies, even in the first few hours of treatment when acidosis and glucose variability are theoretically the highest. This suggests that CGMs could be a potential reasonable alternative to the current standard of care but requires further prospective evaluation. Our subanalysis of CGM accuracy during DKA admissions and IV insulin infusion showed similar accuracy statistics to our overall study cohort.¹⁷

The MARD for DKA admissions (11.8%) was similar to the overall MARD, while the MARD for those admitted in severe DKA indicated better accuracy at 8.9%, although the reason for the improved accuracy requires further evaluation. Our findings of the sensor accuracy in the setting of severe acidosis is consistent with a previous study using Dexcom G6 in the PICU, although glucose data from that study were only available after the patients had received initial IV fluids and been started on IV insulin.²⁷

Overall, the data from pediatric admissions remain limited, and accuracy during IV insulin infusion has not previously been published. As our sample used the patient's own sensor, data were able to be captured immediately upon presentation to the hospital and were not limited by the time to enroll in the study or with the warmup period for new sensor placement. This allowed the data during IV insulin initiation and administration to be analyzed.

The MARD during IV insulin infusion was higher (13.4%) compared to when the same patients transitioned to subcutaneous insulin following resolution of acidosis (10.5%). When examining the MARD during IV insulin infusions across all time points, accuracy was generally maintained between 11.8% (hours 3 to <6 h) and 14.3% (≥6 h). Even during the period of highest acidosis and increased glucose variability at initiation of treatment of DKA, the MARD of 13% is often considered clinically acceptable for inpatient management and comparable to previously published adult data.^6–8

Our study does have several limitations. While we had large sample sizes for DKA and non-DKA admissions (N = 612 and 503, respectively), the sample size for those receiving IV insulin was much smaller, particularly for the subanalysis of IV insulin duration. Smaller sample sizes can lead to either falsely increased or decreased MARDs, making it potentially more difficult to interpret the data. This was alleviated to a degree by utilizing Clarke Error Grids and 15/20/30 analyses, which provide a visual and representative relationship of the comparator glucose value to the reference. POC glucose values are also not the ideal reference for CGM accuracy, as the blood glucose meters themselves have a degree of inaccuracy, as indicated by an elevated MARD when compared to the gold standard laboratory or YSI reference.

In this analysis, the POC reference was the Nova Biomedical StatStrip blood glucose meter, which is one of a limited number of meters approved for use in critical care units and has a relatively good MARD of around 6%. In addition, the retrospective nature of this study limited the ability to collect laboratory reference glucose values or more frequent beta hydroxybutyrate levels following initial presentation to the emergency department, which are not often obtained as frequently during standard of care hospital management. Average sensor wear duration for the patients' home CGM upon admission was not obtainable from the CSV files for analysis and this should be considered in future CGM assessments.

The results from this analysis are promising regarding CGM accuracy during pediatric hospital admissions for DKA. While the close monitoring required during IV insulin and dextrose infusions can be accomplished using frequent POC glucose testing, this process can be cumbersome, painful, and disruptive to patients and staff. The use of CGM may relieve some of this burden, even for those that are newly diagnosed and not previously on a sensor. However, additional studies are needed before adoption of CGM as part of routine care in the hospital setting, including evaluation of sensor accuracy during the use of certain medications (i.e., acetaminophen, vasoactive medications).

In addition, while the MARDs were similar across all time periods when receiving IV insulin, even at the start of treatment when acidosis is theoretically the highest, further evaluation of the impact of acidosis on accuracy through more frequent evaluation of beta-hydroxybutyrate levels may also be beneficial. Further evaluation of the CGM accuracy compared to laboratory glucose values may also be beneficial to provide a more reliable reference comparison. A prospective study examining accuracy in a larger pediatric population (T1D, type 2 diabetes, cystic fibrosis related diabetes, medication induced hyperglycemia, etc.) is currently under development. If CGM accuracy can be demonstrated under a variety of hospital scenarios, safety and efficacy protocols will need to be developed before widespread implementation.

Authors' Contributions

L.A.W. collected data, developed the analysis plan, and wrote and edited the article. E.C. developed the research idea, analysis plan, collected data, and edited the article. L.P., G.P.F., and R.P.W. developed the research idea and analysis plan and reviewed and edited the article. L.T., A.J.K., E.J., and C.B. collected data, and reviewed and edited the article. All authors approved the final version of the article.

Disclosure Statement

L.A.W., L.P., A.J.K., L.T., and E.J. have no conflicts of interest to report. G.P.F. conducts research sponsored by Medtronic, Dexcom, Abbott, Tandem, Insulet, Lilly, and Beta Bionics and has been a consultant, speaker, or advisory board member for Medtronic, Dexcom, Abbott, Tandem, Insulet, Lilly, and Beta Bionics. C.B. has received speaking honoraria for Insulet, Dexcom, and Embecta and has been a consultant for Insulet. R.P.W. conducts research sponsored by Medtronic, Dexcom, Abbott, Tandem, Insulet, Lilly, and Beta Bionics and has been a consultant, speaker, or advisory board member for Medtronic, Dexcom, Abbott, Tandem, Insulet, Lilly, and Beta Bionics. E.C. has been a speaker and on an advisory board for Dexcom.

Funding Information

5T32DK063687, NIDDK

5-ECR-2022-1179-A-N, JDRF