Clinical accuracy of Dexcom G6 and G7 continuous glucose monitors in hospitalized pediatric patients with type 1 diabetes: a real-world study
Article information
Abstract
Purpose
Continuous glucose monitoring (CGM) is increasingly being used in inpatient pediatric diabetes care; however, data on its real-world accuracy remain limited.
Methods
We retrospectively assessed the clinical accuracy of factory-calibrated Dexcom G6 and G7 CGM systems in 69 pediatric patients with type 1 diabetes at a Korean tertiary hospital between 2019 and 2025. A total of 1,838 CGM readings were temporally paired with point-of-care (POC) capillary glucose measurements. Accuracy was evaluated using the mean absolute relative difference (MARD), mean absolute difference (MAD), Bland-Altman analysis, and Clarke error grid classification. Subgroup analyses were performed according to wear day, glucose range, care setting, and CGM type.
Results
The overall MARD was 10.6%±10.1% and MAD was 15.4±16.2 mg/dL. Accuracy improved over time, with MARD declining from 13.8% (G6) and 11.0% (G7) on day 1 to <10% on days 6 and 4, respectively. Bland-Altman limits narrowed from ±60 to ±39 mg/dL. During Days 6–11, 85.9% of CGM-POC pairs were within Clarke zone A and 94.1% within zones A and B. The MAD increased with higher glucose levels, whereas the MARD was highest during hypoglycemia (15.2%) and lowest during hyperglycemia (8.5 %). Intensive care unit admission significantly increased the MAD (p<0.001) without affecting the MARD. G7 demonstrated superior accuracy compared with G6 (pooled MARD: 9.8% vs. 11.2%, p=0.003). Mixed-effects modeling confirmed that wear day was an independent predictor of accuracy improvement (-0.34% MARD/day, p=0.001).
Conclusions
The Dexcom G6 and G7 CGM systems met the ISO (International Organization for Standardization) 15197:2013 and U.S. Food and Drug Administration integrated-CGM criteria after a short inpatient stabilization period.
Highlights
· In 69 hospitalized children and adolescents with type 1 diabetes, factory-calibrated Dexcom G6 and G7 sensors reached a mean absolute relative difference below 10% by day 6 and day 4, respectively. Accuracy improved with each wear day and was maintained across ward and intensive care settings. After a brief stabilization period, both systems met ISO (International Organization for Standardization) 15197:2013 and U.S. Food and Drug Administration integrated-continuous glucose monitoring accuracy criteria.
Introduction
Type 1 diabetes mellitus (T1DM) is relatively uncommon in children in Asia. Nationwide registry data from Korea indicate that the annual incidence of T1DM among individuals younger than 30 years rose from 3.02 to 3.75 per 100,000 between 2008 and 2021, while the prevalence more than doubled, from 21.82 to 46.41 per 100 000 over the same period [1]. Global modeling suggests that 8.4 million people were living with T1DM in 2021 and that this figure could reach 13.5–17.4 million by 2040, with the steepest relative increases expected in low-income and lower-middle-income countries [2]. These epidemiological trends, together with rapid improvements in sensor chemistry, algorithm design, wireless data transmission, and expanded national insurance reimbursements, have dramatically increased the use of real-time continuous glucose monitoring (CGM) among children and adolescents in Korea and worldwide [3,4].
T1D requires rigorous glycemic control to prevent acute crises, such as diabetic ketoacidosis (DKA), and mitigate long-term microvascular and macrovascular complications [5-7]. Capillary self-monitoring of blood glucose (SMBG) has long been the standard of care; however, capillary fingerstick testing provides intermittent point-in-time measurements and often misses rapid excursions, notably nocturnal hypoglycemia [8]. CGM closes this gap by delivering real-time interstitial glucose values, trend arrows, and customizable alerts. Randomized outpatient trials in pediatric T1DM cohorts demonstrated that CGM lowers glycated hemoglobin (HbA1c) levels and reduces exposure to both severe hypo- and hyperglycemia [9].
Over the past decade, sensor accuracy has markedly improved. Factory-calibrated systems such as Dexcom G6 and the next-generation G7 now report overall mean absolute relative differences (MARDs) below 9%, satisfying U.S. Food and Drug Administration (FDA) integrated-CGM (iCGM) criteria that require more than 87% of readings within ±20% or ±20 mg/dL of reference plasma glucose [10]. A recent durability study of a 15-day factory -calibrated sensor involving both adult and pediatric participants (age range, 2–17 years) confirmed an aggregate MARD of approximately 8%, with 94% of values within the 20/20 accuracy envelope [11]. A head-to-head trial showed that G7 achieved a lower MARD (8.9%) than G6 (13.6%), and a greater share of readings within ±20 mg/dL of the reference (91.4% vs 78.6%) [12]. These data highlight the point-wise accuracy that is now attainable, which is essential for CGM-guided insulin dosing and automated insulin delivery systems. However, most evidence has been derived from ambulatory adults, and pediatric inpatients present unique physiological and operational challenges that can reduce accuracy.
The coronavirus disease 2019 (COVID-19) pandemic accelerated the adoption of inpatient CGM through regulatory waivers designed to minimize direct patient contact [13,14]. Early implementation studies confirmed feasibility but reported highly variable accuracy: aggregate inpatient MARDs from 9% to 20%, particularly in intensive care units (ICUs) [10]. However, most available data are derived from adult populations, often with steroid-induced hyperglycemia, leaving pediatric-specific evidence limited and highlighting the need to investigate factors influencing CGM performance in hospitalized children.
The accuracy of CGM in pediatric inpatients is influenced by several factors. Wear-day effects remain prominent; Dexcom G6 MARD falls from 13.4% on day 1 to 9.1% on day 3 in hospitalized adolescents [15]. Moreover, clinical setting also matters, with adult data showing the MARD climbing from 12.0% in general wards to 16.1% in critical-care units. Similar stressors in pediatric ICUs (hypoperfusion, vasoactive infusions, edema, and frequent arterial samples) likely have comparable effects [16]. Accuracy further varies with the glucose range; the relative error is modest in euglycemia, spikes in hypoglycemia, and the absolute error widens during marked hyperglycemia [17]. Therefore, any pediatric inpatient accuracy study should stratify the results by sensor day, care setting, and glucose range to capture distinct sources of variability.
Therefore, we conducted a retrospective study at a tertiary children's hospital to evaluate the clinical accuracy of Dexcom G6 and G7 in hospitalized pediatric patients with T1DM. Accuracy metrics (MARD, MAD, Clarke error grid, Bland-Altman analysis) were assessed by glucose range, sensor day, care setting, and device type to identify patterns over time and factors affecting performance. These findings aim to inform safe and effective CGM use in hospitalized pediatric patients with T1DM.
Materials and methods
1. Study design and setting
We performed a single-center, retrospective cohort study at Seoul National University Bundang Hospital (SNUBH), a tertiary-care center, from January 1, 2019, to June 30, 2025. The study protocol was approved by the institutional review board (IRB) of SNUBH (IRB number: B-2506-980-101).
2. Participants
The inclusion criteria included patients who were younger than 19 years of age, had a confirmed diagnosis of T1D, were hospitalized for 24 or more hours, wore a Dexcom G6 or G7 CGM for at least 24 consecutive hours, and had concurrent capillary blood glucose (BG) measurements. The exclusion criteria were incomplete or missing CGM or capillary BG data (e.g., absent CGM records or unclear SMBG timestamps) or infeasible continuous data collection due to sensor malfunction, premature detachment, or signal loss.
3. Devices and data sources
Dexcom G6 and G7 sensors (Dexcom Inc., USA) provided factory-calibrated interstitial glucose readings at 5-minute intervals and stored the values on a receiver or smartphone application [18]. Raw sensor files were exported from Dexcom Clarity (a cloud platform) as comma-separated value files [19]. Reference BG values were obtained using the BAROzen H Expert Plus BG monitoring system (Handok, Korea), which employs a glucose dehydrogenase-flavin adenine dinucleotide enzymatic method and meets the International Organization for Standardization (ISO) 15197:2013 performance criteria [20]. All BG readings were automatically transmitted to the hospital electronic medical record system.
4. Pairing procedure
The CGM data were merged with the POC glucose data using R v4.2.3 (R Foundation for Statistical Computing, Austria) [21]. Each capillary BG measurement was matched with the closest CGM value. When the BG value occurred between two 5-minute CGM readings, the CGM value was linearly interpolated from the flanking points. Duplicate POC values recorded 1 minute apart were consolidated by retaining the initial measurement.
5. Accuracy metrics
Analytical accuracy was evaluated with the following indices:
• MARD: |CGM−POC|/POC × 100% [22]
• Mean absolute difference (MAD): |CGM−POC| in mg/dL
• Clarke error grid: Each CGM-POC pair was categorized into zones A through E. Zones A and B were considered clinically acceptable [23].
• Bland-Altman analysis: Mean bias and 95% limits of agreement (LoAs) were calculated for the entire cohort and key subgroups [24].
6. Subgroup definitions
Sensor-day accuracy was indexed by wear day. Although Dexcom G6 and G7 systems are labeled for 10 days of use, we analyzed data for wear days 1–11. Glycemic zones were defined as hypoglycemia (<70 mg/dL), euglycemia (70–179 mg/dL), and hyperglycemia (≥180 mg/dL) and, for adjusted models, were coded as an ordinal variable. DKA severity was classified as none, mild (venous pH <7.30 or serum bicarbonate <18 mmol/L), moderate (pH <7.20 or bicarbonate <10 mmol/L), and severe (pH <7.10 or bicarbonate <5 mmol/L) [25]. Care setting at the time of pairing was recorded as ICU versus general ward. Device type (G6 vs. G7) was identified from Clarity exports.
7. Statistical analysis
Continuous variables were presented as means±standard deviations or medians (interquartile ranges [IQRs]) as appropriate. Welch t-test or the Mann-Whitney U-test was used to compare MARD and MAD between groups. Clarke grid zone proportions were compared with the Pearson chi-square test.
Repeated measures within individuals were modeled with linear mixed-effects regression (lme4 package, R), with fixed effects for care setting (ICU vs. ward), sensor day, glycemic zone, DKA severity, and device type (G7 vs. G6). We report β estimates with standard error, 95% confidence interval (CI), and 2-sided P-values. The model residuals were examined for normality and homoscedasticity. Statistical significance was set at a 2-sided P-value of less than 0.05. Figures were generated using Python 3.11 (Matplotlib) and refined in Adobe Illustrator [26].
8. Data security and reporting
All data processing and statistical analyses were conducted on a secure institutional server with restricted access.
Results
1. Patient characteristics
Patient baseline characteristics are summarized in Table 1. We analyzed 69 hospitalized children and adolescents with T1DM, of whom 29 (42%) were male. Age at diagnosis was 9.3±4.2 years, the mean age at the index admission was 12.0±3.9 years, and the mean duration of diabetes was 1.0±2.4 years. At admission, the mean HbA1c was 10.4%±3.0% and mean C-peptide was 0.72± 1.44 ng/mL. Seventeen patients (24.6%) were admitted to the ICU. CGM sensors were worn for a mean duration of 6.6±2.5 days. A total of 1,838 CGM readings were paired with contemporaneous POC glucose tests, averaging 26.6±15.0 pairs per participant. Of these, 1,101 pairs (59.9 %) were obtained from Dexcom G6 sensors and 737 (40.1%) from G7 sensors. Across all pairs, the MAD was 15.4±16.2 mg/dL and MARD was 10.6%±10.1%. Hypoglycemic (<70 mg/dL) and hyperglycemic (≥180 mg/dL) readings represented 7.6% and 33.4% of all measurements, respectively.
2. Time-dependent improvement in CGM accuracy after sensor placement
A total of 1,838 CGM-POC glucose pairs were analyzed to assess day-by-day sensor performance (Table 2). The number of daily CGM-POC pairs declined from 349 on day 1 to 5 on day 11. On day 1, the Dexcom G6 showed a MAD of 22.4±22.1 mg/dL and a MARD of 13.8%±12.8%. Accuracy improved over time: the MARD dropped below the clinically acceptable threshold of 10% by day 6 (9.9%±10.4%) and reached a nadir on day 9 (7.8%±6.5%). MAD showed a similar trend, decreasing to 10.7±8.5 mg/dL on day 10. In mixed-effects models adjusting for care setting, glycemic zone, DKA severity, and device type, each additional wear day was associated with lower MAD by 0.83 mg/dL/day (95% CI, -1.24 to -0.42; P<0.001) and lower MARD by 0.44 percentage points/day (95% CI, -0.69 to -0.18; P<0.001), a pattern visually corroborated by the locally estimated scatterplot smoothing (LOESS)-smoothed trajectory in Fig. 1.
Day-by-day mean absolute relative difference (MARD) for Dexcom G6 (light gray bars) and G7 (dark gray bars) sensors in hospitalized pediatric patients with T1D. Dashed and solid LOESS curves trace the nonlinear accuracy trends for G6 and G7, respectively. LOESS, locally estimated scatterplot smoothing.
Dexcom G7 stabilized more rapidly. On day 1, MAD and MARD were 21.2±20.3 mg/dL and 11.0%±9.6%, respectively. MARD dropped below 10% by day 4 (7.8%±6.6 %) and reached 6.3%±3.7% by day 11. MAD followed a parallel pattern, decreasing to 9.6±7.6 mg/dL (Table 2). In the same mixed-effects framework, each additional wear day was associated with lower MAD by 0.56 mg/dL/day (95% CI, -1.09 to -0.03; P=0.038), whereas the change in MARD (-0.17 percentage points/day) was not significant (95% CI, -0.50 to 0.15; P=0.298). Consistently, the LOESS curve in Fig. 1 shows an early decline (crossing <10% by day 4) followed by stabilization.
In the pooled mixed-effects model including both devices, each additional wear day was independently associated with a 0.34% point decrease in MARD (P=0.001).
Both sensors met the ISO 15197:2013 performance standards (MARD <10%) for the majority of the monitoring period, doing so by day 6 for the G6 and by day 4 for the G7 [27]. From day 4 onward, the daily MAD values for both sensors remained at or below 15 mg/dL, supporting consistent absolute accuracy.
Bland-Altman analysis across sensor-day strata (day 1, day 2, days 3–5, and days 6–11) demonstrated minimal systematic bias (mean differences, +1.0 to +2.8 mg/dL), but precision improved markedly over time. The 95% LoA narrowed with increasing wear time, particularly after day 2. Visual inspection revealed mild heteroscedasticity— i.e., greater CGM deviation at higher glucose concentrations—on day 1, which resolved by day 3, suggesting early variability was primarily random rather than systematic (Fig. 2). Clarke error grid analysis further confirmed improving clinical reliability. Zone A proportion increased from 78.3% on day 1 to 85.9% by days 6–11, while zone B decreased from 18.9% to 8.2%. Zones C and E were absent after day 1, and zone D remained consistently low (2.6%–5.9%) (Supplementary Table 1). Across all day strata, ≥94% of readings fell in zones A + B, peaking at 97.2% on day 1 and 96.9% during days 3–5 (Fig. 3).
Bland-Altman agreement between CGM and POC glucose by sensor age. Each gray “×” is a matched CGM-POC pair; the dashed red line shows mean bias and the dashed blue lines mark the 95% limits of agreement (±1.96 SD). CGM, continuous glucose monitor; POC, point-ofcare; SD, standard deviation.
3. CGM accuracy by glycemic range
When CGM glucose readings were compared with POC capillary glucose tests, MAD increased with rising glucose concentrations (Supplementary Fig. 1). MAD reached 20.85±21.20 mg/dL in the hyperglycemic range (≥180 mg/dL; n=612), 12.89±12.03 mg/dL during euglycemia (70–179 mg/dL; n=1,086), and was lowest in hypoglycemia (<70 mg/dL; 9.45±7.46 mg/dL, n=140).
In contrast, MARD was lowest in hyperglycemia (8.51% ±8.11%), increased during euglycemia (11.09%±10.40%), and peaked during hypoglycemia (15.17%±12.25%). The Kruskal-Wallis tests confirmed significant differences across all 3 glucose strata for both MAD and MARD (P<0.001 for each metric).
4. CGM accuracy by device type (G6 vs. G7)
Of the 1,838 paired readings, 1,101 (59.9%) were obtained from Dexcom G6 and 737 (40.1%) from G7 sensors. Pooled across all wear days, the G7 demonstrated higher relative accuracy than the G6, with a lower MARD (9.8%± 9.2% vs. 11.2%±10.5%, P=0.003), whereas the MAD was comparable between devices (14.9±15.0 mg/dL vs. 15.6± 16.6 mg/dL, P=0.37). In the linear mixed-effects model that accounted for repeated measures and adjusted for sensor day, glycemic zone, DKA severity, and care setting, the device-type effect was attenuated and no longer significant (MARD: β=-1.36 percentage points for G7 vs. G6, 95% CI, -3.64 to 0.93; P=0.25; MAD: β=-1.74 mg/dL; 95% CI, -5.04 to 1.57; P=0.30), indicating that the crude G7 advantage was largely explained by its faster postinsertion stabilization.
5. CGM accuracy in ICU and general ward settings
Among 1,838 CGM-POC pairs, 1,739 were obtained on the ward and 99 in the ICU. The 2 care settings were comparable in demographics and diabetes history (age at diagnosis and at admission, body mass index standard deviation score, diabetes duration, C-peptide) and had a similar distribution of device type (Dexcom G7 vs. G6). HbA1c at admission was modestly higher in ICU patients (11.5%±1.8% vs. 10.0%±3.18%). As expected, metabolic derangement was substantially worse in the ICU cohort: lower pH (7.06±0.14 vs. 7.37±0.06), lower bicarbonate (7.73±4.43 mmol/L vs. 23.87±4.67 mmol/L), more negative base excess (-21.26±6.42 mmol/L vs -1.12±3.99 mmol/L), higher lactate (2.04±1.06 mmol/L vs. 1.37±0.48 mmol/L), higher serum ketone (2.58±0.93 mmol/L vs. 1.21±1.28 mmol/L), and higher calculated osmolality (312.7±22.8 mOsm/kg vs. 294.7±9.2 mOsm/kg) (all P≤ 0.022), with DKA severity concentrated in the ICU (mild/moderate/severe: 1/6/10 vs. 6/0/0) (Supplementary Table 2).
MAD was greater in the ICU (24.35±20.14 mg/dL) than on the ward (14.76±15.56 mg/dL, P<0.001), whereas MARD was similar between settings (ICU 11.0%±7.85% vs. ward 10.52%±10.12%, P=0.091). In linear mixed-effects models that accounted for repeated measurements and adjusted for sensor day, glycemic zone, DKA severity, and device type, ICU care remained independently associated with higher MAD (β=+5.69 mg/dL; SE, 1.86; 95% CI, 2.04–9.34; P<0.001), while its association with MARD was not significant (β=+1.67 percentage points; SE, 1.15; 95% CI, -0.59 to 3.93; P=0.150).
Among ICU-time measurements, error metrics varied significantly by DKA severity (Supplementary Table 3). Median (IQR) MAD was 39.44 (22.74–58.22) mg/dL in mild DKA (n=23), 17.59 (6.28–26.98) mg/dL in moderate DKA (n=53), and 14.28 (5.22–23.70) mg/dL in severe DKA (n=23), with an overall P=0.001 and pairwise differences for mild versus moderate P=0.002 and mild versus severe P=0.011. Median (IQR) MARD was 16.14% (8.98–24.02%) in mild DKA, 9.38% (3.43–14.74%) in moderate DKA, and 6.52% (2.86–13.10%) in severe DKA, with an overall P=0.005 and pairwise differences for mild versus moderate P=0.024 and mild versus severe P=0.032.
Discussion
In this retrospective, single-center study of hospitalized pediatric patients with T1DM, factory-calibrated Dexcom G6 and G7 CGM systems demonstrated clinically acceptable accuracy following a brief postinsertion stabilization period. On day 1, the MARD was 13% for G6 and 11% for G7, yet values decreased below the 10% clinical-accuracy threshold by day 4 with G7 and by days 6 with G6, thereafter stabilizing at 6%–8%, with a corresponding MAD of approximately 10–12 mg/dL. These improvements were accompanied by a progressive narrowing of the 95% Bland-Altman LoA, which contracted from ±60 mg/dL on day 1 to ±35–39 mg/dL after day 3. Concomitantly, Clarke error grid zone A readings increased to more than 85% (zones A + B ≈ 97%), confirming performance that met both ISO 15197:2013 and U.S. FDA accuracy benchmarks for therapeutic inpatient CGM use [10,22,23,27].
The time-dependent accuracy gains observed in our cohort mirror those reported in adult outpatient studies, which have demonstrated MARD reductions of 20%–40% within the first 48–72 hours after sensor insertion [28]. While the biological underpinnings of this early variability are not fully understood, transient biofouling and a local foreign-body response are plausible contributors [29]. The G7’s rapid stabilization—achieving MARD <10% by day 4 versus day 6 for G6—is likely attributable to improvements including a shorter introducer needle, updated enzyme formulation, and an enhanced algorithm [11,12], resulting in a clinically meaningful 2-day reduction in sensor "time-to-trust" when sensors are placed on admission.
Although G7 consistently achieved lower absolute MARD levels and crossed <10% earlier (day 4), its adjusted linear wear-day slope for MARD did not reach statistical significance. This likely reflects a floor effect and a nonlinear trajectory characterized by early improvement during days 1–4 followed by rapid stabilization, as visually corroborated by the LOESS-smoothed curves in Fig. 1. In contrast, G6 started from a higher day-1 MARD and showed a larger per-day reduction, yielding a significant linear slope in mixed-effects models.
Accuracy varied by glycemic range: MAD was highest during hyperglycemia (≈21 mg/dL) and lowest during hypoglycemia (≈9 mg/dL). In contrast, MARD was lowest during hyperglycemia (8.5%) and peaked during hypoglycemia (15.2%). This pattern, consistent with prior pediatric inpatient CGM studies, reflects the intrinsic proportional relationship between sensor error and glucose level [17,30]. Despite higher proportional error during hypoglycemia, the associated absolute deviation remained modest (±9 mg/dL at 60 mg/dL), supporting overall safety while still justifying clinical confirmation of low-glucose alarms [19].
Critically ill patients admitted to the ICU exhibited higher MAD (+5.69 mg/dL) but no significant increase in MARD compared to those on the general ward. These results differ from adult data, where vasopressor use and tissue hypoperfusion often impair CGM accuracy [14]. Similar pediatric studies on DKA admissions have reported MARD values of 7%–11% once acidosis resolves [28-31], suggesting that better peripheral perfusion, shorter ICU stays, and careful sensor-site selection mitigate many ICU-specific challenges. Nevertheless, clinicians should be cautious during episodes of hemodynamic instability or severe acidosis when transient inaccuracies may still arise [30].
In the ICU subset, the counterintuitive gradient of error (mild>moderate>severe DKA) likely reflects a convergence of clinical and methodological factors. Mild DKA patients were more often instrumented immediately on admission, yielding a larger share of day-1 to day-2 measurements when early postinsertion stabilization inflates error, whereas severe DKA patients tended to have sensors placed after initial resuscitation or contributed more wear days beyond day 3, when accuracy improves. Faster glycemic rate of change during ambulation and intermittent corrections in milder illness magnifies interstitial-plasma lag and increases proportional error, and greater activity also predisposes to micromotion and compression artifacts that elevate absolute error. By contrast, severe cases typically had stricter site selection and fixation and less movement, creating a more stable sensing environment.
Throughout the observation period, the G7 consistently posted MARD values 1%–2% lower than those of the G6 and crossed the 10% threshold 2 days sooner. Although the study was not powered for formal noninferiority testing, the effect size aligned with outpatient adult trials that showed a 30%–40% MARD reduction with the G7 [12]. Importantly, both devices ultimately satisfied the ISO targets after their respective break-in phases, indicating that either system can be safely deployed in hospitalized youths.
An overall MARD of approximately 10% is comparable to the intrinsic error of many hospital POC glucometers currently used to guide insulin dosing [20,27]. With more than 94% of CGM values falling in Clarke error grid zones A and B at every time point, most readings would lead to identical or clinically benign decisions relative to fingerstick results. Continuous surveillance also addresses the major limitations of intermittent bedside testing by reducing the risk of missing nocturnal hypoglycemia or rapid glycemic excursions. In addition, it decreases the nursing workload and minimizes patient discomfort [13,14,30].
The recent Korean experience during the COVID-19 pandemic further demonstrates CGM’s utility once children have transitioned out of the hospital. A multicenter telehealth program found that approximately 70% of Korean youths with T1DM were already using CGM and could be managed safely at home during mandatory quarantine without episodes of severe glycemic decompensation [32]. In a separate cohort, CGM metrics captured a brief rise in glucose management indicators and time-above-range during mild severe acute respiratory syndrome coronavirus 2 infection that normalized within 2 weeks. Children on insulin pumps showed significantly smaller excursions than those on multiple daily injections, underscoring the synergistic benefit of advanced insulin delivery and real-time sensing [33]. These real-world data highlight CGM's capacity to support continuous, remote decision-making under acute stress and reinforce its role as a cornerstone of comprehensive pediatric diabetes care.
These findings support current U.S. consensus guidance on inpatient CGM use [34]. When a child is admitted with a CGM in place for more than 24 hours, crosschecking the first few CGM values against POC readings is generally sufficient before initiating clinical reliance. For newly applied sensors, clinical care teams should expect higher day-1 variability (MARD>11%) and maintain routine fingerstick frequency or confirm hypoglycemia alarms until days 2–3, particularly in the context of hypoglycemia alarms or when making critical dosing decisions [28]. Safety can be enhanced by setting slightly higher low-glucose alarm thresholds (e.g., 80 mg/dL) and accounting for the physiological lag time (5–15 minutes) between interstitial and plasma glucose. In the ICU setting, CGM use remains feasible once perfusion stabilizes; however, confirmatory POC testing is warranted when CGM values would prompt substantial insulin adjustments or when sensor signal quality is suboptimal [28,29]. Broader adoption of inpatient CGM will require robust integration with electronic medical records, remote monitoring dashboards, standardized calibration and removal protocols, and comprehensive staff education to interpret trend arrows and rate-ofchange alarms [34,35].
This study offers several advantages that enhance its contribution to the field. First, it represents the largest analysis to date that focuses exclusively on hospitalized children and adolescents with T1DM in Asia, an understudied population in CGM accuracy research. Second, each POC glucose value was matched to the nearest CGM reading using linear interpolation, minimizing timing bias and better approximating real-world clinical decision-making compared with prior reports. Third, by encompassing the full spectrum of inpatient care, including ICU admissions, and stratifying results by sensor day and glucose range, we captured CGM performance across the most clinically relevant scenarios. Fourth, using multiple accuracy metrics (MARD, MAD, Clarke error grid, and Bland-Altman) combined with linear mixed-effects modeling allowed a comprehensive assessment that accounted for repeated measures within patients. Finally, the 11-day observation period delineated the complete trajectory of accuracy maturation for both the G6 and G7 devices, thereby defining the practical "time-to-trust" threshold for each. Collectively, these strengths provide a robust and clinically actionable evidence base to inform best-practice guidelines and guide future multicenter investigations on inpatient CGM use in pediatric patients with T1D.
Several limitations warrant consideration. First, the retrospective, single-center design and exclusive inclusion of patients with T1DM limit the generalizability of the findings to other dysglycemic populations. Moreover, the performance of other non-Dexcom devices has yet to be evaluated. Second, because the reference standard was POC capillary meters, with an inherent analytical error of ±15% [20,27], the degree of sensor inaccuracy observed in this study may have been overestimated. Finally, potential selection bias must also be acknowledged, given that the continuation of CGM use was subject to individual clinician judgment.
Future work should involve multicenter prospective trials to confirm accuracy across varied clinical environments and determine whether CGM-guided protocols increase time-in-range, decrease the incidence of hypoglycemia, improve insulin-dose precision, and shorten hospital stay. Workflow investigations should examine the integration of automated data feeds, central alarm monitoring, diabetes nurse educators' training requirements, and cost-effectiveness. Hybrid closed-loop pumps may offer the prospect of semiautomated insulin delivery during routine admissions or recovery from DKA. Incremental sensor improvements, including shorter warm-up periods, enhanced accuracy in the hypoglycemic range, and extended wear times, are expected to further reduce implementation barriers. Our findings indicate that the current sixth- and seventh-generation Dexcom systems already provide reliable, near-real-time glucose data in hospitalized children, provided that clinical teams respect the brief break-in intervals and continue to verify outlier readings judiciously. As professional societies refine clinical guidance and regulatory frameworks, inpatient CGM is poised to become the standard of care for pediatric patients with T1DM, enabling safer, less intrusive, and more precise glycemic management.
In conclusion, Dexcom G6 and G7 sensors rapidly achieved inpatient benchmarks in pediatric inpatients with T1D. The G7 achieved a MARD below 10% by day 4, and the G6 by day 6. By day 4, over 94% of CGM-POC glucose pairs fell within Clarke error grid zones A + B, with Bland-Altman LoAs of ±35 mg/dL. These results satisfied both the ISO 15197:2013 and FDA iCGM accuracy criteria. While modest reductions in accuracy were observed in ICUs and during hypoglycemia, these findings support confirmatory POC testing during the first 48–72 hours of sensor wear and in response to lowglucose alarms. Overall, the results support the routine inpatient use of modern Dexcom CGMs following a brief initial stabilization period.
Supplementary materials
Supplementary Tables 1-3 and Supplementary Fig. 1 are available at https://doi.org/10.6065/apem.2550338.169.
Percent Clarke error grid zone distribution by sensor day for paired CGM-POC readings
Baseline characteristics of hospitalized pediatric T1DM (ICU vs. ward)
MAD and MARD by DKA severity during ICU-time measurements in hospitalized pediatric patients with T1DM
Relative (MARD, solid black; left axis) versus absolute (MAD, dashed gray; right axis) CGM error across glucose zones with 95% confidence intervals (CIs). MAD, mean absolute difference; MARD, mean absolute relative difference; CGM, continuous glucose monitor.
Notes
Conflicts of interest
JK received honoraria from Abbott, Dexcom, and Medtronic, and research fund from i-Sens. Apart from this, no potential conflict of interest relevant to this article was reported.
Funding
This study received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Data availability
The data that support the findings of this study can be provided by the corresponding author upon reasonable request.
Acknowledgments
We thank the nursing staff, diabetes nurse educators, and clinical laboratory personnel of Seoul National University Bundang Hospital for their assistance with continuous glucose monitoring data collection and patient care.
Author contributions
Conceptualization: JK; Data curation: DJH, JH; Formal analysis: DJH, HL, JH, MY; Methodology: DJH, HL, HYK; Project administration: JK; Writing – original draft: DJH; Writing – review and editing: DJH, HL, JH, MY, HYK, JK
