Development of a glycated hemoglobin prediction model using continuous glucose monitoring metrics in pediatric type 1 diabetes mellitus: insights into average glucose and recent glycemic trends

Article information

Ann Pediatr Endocrinol Metab. 2026;31(3):189-197

Publication date (electronic) : 2025 November 19

doi : https://doi.org/10.6065/apem.2550214.107

Hakyung Lee¹, Mi Yang¹, Hwa Young Kim^1,2, Jaehyun Kim^,1,2

¹Department of Pediatrics, Seoul National University Bundang Hospital, Seongnam, Korea

²Department of Pediatrics, Seoul National University College of Medicine, Seoul, Korea

Address for correspondence: Jaehyun Kim Department of Pediatrics, Seoul National University Bundang Hospital, 82, Gumi-ro 173 Beon-gil, Bundang-gu, Seongnam, Gyeonggi-do, 13620, Korea Email: jaehyun.kim@snu.ac.kr

Received 2025 June 30; Revised 2025 September 16; Accepted 2025 October 3.

Abstract

Highlights

· In Korean children with type 1 diabetes mellitus, continuous glucose monitoring (CGM)-derived metrics—especially average glucose, time-above-range, and time-in-range—were strongly associated with glycated hemoglobin (HbA1c). Glycemic data from the most recent 0- to 4-week predicted HbA1c better than longer intervals, with average glucose showing the highest accuracy. Incorporating recent CGM metrics into routine practice may improve HbA1c interpretation and support individualized diabetes management.

Introduction

Type 1 diabetes mellitus (T1DM) is a chronic autoimmune condition characterized by insulin deficiency and persistent hyperglycemia [1]. Although its prevalence among children and adolescents is relatively low in many Asian countries compared to Western nations, the global incidence of T1DM has been steadily increasing, with similar trends observed in Korea [2-4]. Glycated hemoglobin (HbA1c) has long served as the gold standard for assessing average glycemic control in individuals with T1DM [5]. Tight HbA1c control is crucial, as it reduces the risk of long-term diabetic complications [6]. However, HbA1c is limited in that it reflects only the average blood glucose over the preceding 2–3 months and does not capture glycemic excursions such as hypoglycemia, hyperglycemia, and glycemic variability [7]. Its reliability is further compromised by conditions such as hemoglobinopathies [8] and racial differences in glycation, with ethnic minorities often showing higher HbA1c despite similar glucose levels [9,10].

Continuous glucose monitoring (CGM) systems have emerged as an essential tool in modern diabetes management, offering real-time glucose data and detailed metrics such as time-in-range (TIR), time-above-range (TAR), time-below-range (TBR), and glucose variability [11,12]. Studies have shown that CGM use can improve glycemic control and reduce glycemic variability compared to self-monitoring of blood glucose [13-15]. Recent meta-analyses have lent further support to CGM use in both pediatric and adult populations by demonstrating significant reductions in HbA1c and hypoglycemic episodes [16,17]. Based on these benefits, early initiation of CGM at diagnosis is now widely recommended as a part of standard T1DM management [18-20]. In Korea, CGM adoption in pediatric T1DM patients increased from 1.4% to 39.3% over the past decade, with a corresponding decline in mean HbA1c level from 8.56% to 8.01% [21,22].

The recent 2024 International Society for Pediatric and Adolescent Diabetes (ISPAD) guidelines recommended a lower HbA1c target of ≤6.5%—compared to the conventional target of <7%—for patients using advanced diabetes technologies such as CGM and automated insulin delivery [23]. Along with CGM metrics emphasized as complementary indicators of glycemic control [24], CGM-derived glycemic goals such as TIR >70%, TBR <4%, and TAR <25% are recommended, in addition to HbA1c targets [11,25]. To achieve the stricter HbA1c target of ≤6.5%, tighter CGM goals—including TIR >80% and time-in-tight-range (TITR) >55%—may be necessary, based on the rationale that increased time in optimal glucose ranges reflects reduced glycemic variability and may help prevent long-term diabetes-related complications [23,26].

Despite these advances, most studies evaluating the relationship between CGM metrics and HbA1c have been conducted in adults, with limited data available in pediatric populations, especially in Asia [27]. There remains a clinical need to better understand how CGM metrics relate to HbA1c in children and adolescents and to determine whether recent glycemic data can predict HbA1c more accurately. Therefore, this study aimed to investigate the association between CGM metrics and HbA1c levels and to develop and validate CGM-based HbA1c prediction models in Korean pediatric patients with T1DM.

Materials and methods

Results

1. Patient characteristics

Among 85 pediatric patients with T1DM, mean age was 10.8±4.2 years, and mean diabetes duration was 2.0 (interquartile range, 0.8–3.8) years. At diagnosis, mean HbA1c level was 12.3%±2.1%, which decreased to 6.7%±0.9% at the time of analysis (Table 1). Over the 12 weeks preceding HbA1c measurement, mean CGM-derived metrics were as follows: TIR 65.2%±18.4%; TAR 32.1%±18.8 %; and TBR 2.7%±2.6%. These values remained consistent across all time intervals (0–2, 0–4, 4–8, 8–12, and 0–12 weeks), without significant variations (Supplementary Table 1).

Table 1.

Baseline characteristics of study participants

2. Associations between CGM metrics and HbA1c

Associations between HbA1c and each CGM metric across different time intervals are summarized in Table 2. All predictors demonstrated statistically significant associations with HbA1c (all P<0.05). TIR consistently exhibited the largest negative regression coefficients across all intervals. For the 0- to 4-week interval, the coefficient for TIR was -0.79 (95% confidence interval [CI], -0.90 to -0.69), indicating a strong inverse association with HbA1c. In contrast, TAR and average glucose showed the largest positive coefficients, both at 0.80 (95% CI, 0.70–0.90), reflecting strong positive associations with HbA1c levels (Fig. 1A). TITR showed a smaller negative coefficient of -0.73, while CV and TBR demonstrated weak associations (CV, 0.25; TBR, -0.33), with lower R² and higher AIC values (Table 2). Similar trends were observed across other time intervals (0–2, 4–8, 8–12, and 0–12 weeks). Supplementary Fig. 1 provides an intuitive visualization of the relationships between HbA1c and CGM metrics, presented as unadjusted scatter plots without adjustment for model fit indices such as AIC or R².

Table 2.

Associations between continuous glucose monitoring metrics and glycated hemoglobin by time intervals

Fig. 1.

(A) Regression coefficients with 95% confidence intervals for the associations between glycated hemoglobin (HbA1c) and continuous glucose monitoring metrics across time intervals. Associations between CGM metrics and HbA1c across different time intervals are presented. Estimated regression coefficients and corresponding 95% confidence intervals (CIs) are shown for TIR, TAR, TBR, TITR, CV, and average glucose during each observation period: 0-2, 0–4, 4–8, 8–12, and 0–12 weeks. Each point represents the coefficient derived from a generalized linear model predicting HbA1c from the respective CGM metric, adjusted for age at diagnosis, sex, and diabetes duration, with horizontal error bars indicating the 95% CI. The figure displays effect estimates and corresponding CIs only; differences in model fit statistics (adjusted R² and AIC) are presented separately in Fig. 1B. (B) Comparison of adjusted R² and AIC values of the HbA1c prediction models by metric and interval. Adjusted R² values (bar graph) and AIC values (dot plot) are shown for TIR, TAR, and average glucose across 0–4, 4–8, 8–12, and 0–12 weeks. CGM, continuous glucose monitoring; TIR, time-in-range; TAR, time-above-range; TBR, time-below-range; TITR, time-in-tight-range; CV, coefficient of variation; R², coefficient of determination; AIC, Akaike Information Criterion.

4. Development and validation of HbA1c prediction models

Table 3 shows the performance of HbA1c prediction models incorporating CGM metrics of different time intervals. The equation used was as follows:

Table 3.

Comparison of predictive performance among the continuous glucose monitoring metrics across time intervals

HbA1c=β_pat+β_gly+β₀×metric_0-2wk+β₁×metric_0-4wk···β_n×metric_int + ε

Every CGM metric of each time interval was used as predictor to build separate ridge regression models. Patient-specific intercept β_pat was added to estimate the overall intercept β_gly of the glycemic metric coefficient. In both the training and test datasets, models using average glucose, TAR, and TIR showed strong predictive performance, with consistently low MSE and high R² values. In the 0- to 4-week interval, the model based on average glucose achieved the lowest MSE in the training set (0.14) and a high R² (0.83), which remained stable in the test set (MSE=0.15; R²=0.83), indicating robust generalizability. Similarly, the model using TAR achieved favorable results (MSE=0.16; R²=0.83 in test data), followed closely by TIR (MSE=0.16; R²=0.82). In contrast, models based on CV and TBR demonstrated limited predictive accuracy, with higher MSEs (e.g., TBR: 0.68; CV, 0.75 in test data) and lower R² values (TBR, 0.26; CV, 0.18), suggesting relatively poorer performance. TITR yielded intermediate performance (MSE=0.23; R²=0.74). When comparing time intervals, the 0- to 4-week interval consistently yielded the largest absolute regression coefficients across all CGM metrics, outperforming the 0–2, 4–8, 8–12, and 0–12-week intervals (Table 3).

Discussion

In this study of Korean pediatric patients with T1DM, CGM-derived metrics—particularly average glucose, TAR, and TIR—demonstrated strong and consistent associations with HbA1c across multiple time intervals. Among these, average glucose during the 0- to 4-week period emerged as the most robust predictor of HbA1c, achieving the best model performance in both training and test datasets.

Current ISPAD and American Diabetes Association guidelines recommend achieving not only an HbA1c target of less than 6.5%–7%, but also CGM-derived glycemic goals, such as TIR >70% [23,25]. Numerous studies, mainly conducted in adult populations with T1DM, have reported a strong inverse correlation between TIR and HbA1c [27,34,35]. Vigersky and McMahon [27] reported an excellent correlation between the two (R=-0.84, R2=0.71). Similarly, Díaz-Soto et al. [34] found a comparable relationship between TIR and HbA1c (R=-0.75, R2=0.72), which remained stable when analyzed separately in pediatric and adult patients. Valenzano et al. [35] also reported a strong TIR-HbA1c correlation in a real-world cohort monitored for 1 year. Findings in previous pediatric studies also support the role of TIR as a meaningful marker for glycemic control in children and adolescents [30,36]. Vandenbempt et al. [30] observed strong linear correlations between HbA1c and TIR measured over 2, 4, and 12 weeks before consultation, with R values of -0.57, -0.60, and -0.62, respectively. Consistent with these prior findings, our study in Korean pediatric patients also revealed a significant inverse relationship between TIR and HbA1c across all time intervals from GLM analysis, with the strongest association observed during the most recent 0- to 4-week period.

Furthermore, both TAR and average glucose showed strong positive associations with HbA1c in our study, with average glucose from the most recent 0- to 4-week interval yielding the highest predictive power. Although HbA1c is traditionally considered to reflect glycemia over the preceding 2 to 3 months, previous studies have shown that it is disproportionately influenced by more recent glucose levels [28,29]. Our findings are consistent with this, demonstrating that short-term glycemic trends—particularly those from the past 4 weeks—contribute more substantially to HbA1c than long-term averages. Notably, in this interval, all CGM metrics exhibited uniformly lower AIC values, and average glucose demonstrated no less than an 8- to 10-point reduction compared with other intervals—a difference generally regarded as indicating a statistically superior model fit [37]. This emphasizes the clinical utility of recent average glucose in guiding treatment decisions, especially in pediatric patients with poorly controlled hyperglycemia. In routine clinical practice, 2-week CGM summaries are frequently used to estimate glycemic control [31,32]. However, our data—even in a generally well-controlled cohort—revealed that CGM metrics from the 0- to 2-week interval had lower explanatory power (higher AIC, lower R²) than those from the 0- to 4-week interval. These results suggest that reviewing at least 4 weeks of CGM data may provide a more accurate and clinically meaningful assessment of glycemic status, particularly when interpreting or predicting HbA1c values in pediatric patients.

Our findings are therefore consistent with previous pediatric studies that reported a strong association between HbA1c and CGM metrics such as TIR, TAR, and average glucose, while showing relatively weaker associations with CV and TBR [30]. Although Petersson et al. reported a strong association between TITR and HbA1c [36]—suggesting that TITR, like TIR, may serve as a useful supplement to HbA1c—our data showed that TITR had a weaker association with HbA1c than TIR. CV and TBR demonstrated even more limited associations, indicating that HbA1c is relatively insensitive to variations in glycemic variability and hypoglycemia. This finding reinforces the importance of using these metrics to assess aspects of glycemic control beyond what HbA1c can capture. Taken together, these findings underscore the capacity of CGM metrics not only to reflect average glycemia, but also to capture hyperglycemia, hypoglycemia patterns, and glycemic variability—components that may not be fully identified by HbA1c alone. Consistent with previous evidence linking HbA1c levels to the risk of developing diabetes-related complications [38], several studies have also emphasized the relationship between CGM metrics and microvascular outcomes [39,40]. Beck et al. first demonstrated that TIR was strongly associated with both the progression of diabetic retinopathy and development of microalbuminuria [39]. Similarly, Lu et al. [40] reported that patients with more advanced diabetic retinopathy had significantly lower TIR and higher glycemic variability, as measured by CV or SD. Our results therefore support the complementary use of CGM-derived metrics for a more comprehensive assessment of glycemic control and associated complication risks.

Our study adopted GLMs rather than a linear mixed model, as each CGM metric was repeatedly measured across multiple time intervals, whereas HbA1c was measured only once per participant. This approach allowed us to compare regression coefficients as well as adjusted R² and AIC values across time intervals, enabling a nuanced evaluation of each CGM metric’s contribution to HbA1c variation. In a previous study using fasting and postprandial capillary blood glucose to predict HbA1c in type 2 diabetes mellitus patients, Yuan et al. [33] applied ridge regression instead of linear regression due to collinearity between predictors. Similarly, to address multicollinearity among CGM metrics in our dataset, we employed ridge regression with cross-validation to enhance the stability of our prediction model. Through ridge regression modeling, we demonstrated that models incorporating average glucose and TAR offered the most accurate and robust prediction of HbA1c, followed by TIR. Notably, the 0- to 4-week interval showed the highest prediction accuracy across all metrics. These findings underscore the clinical relevance of recent CGM data—particularly within the 4 weeks preceding HbA1c measurement—as a powerful tool to enhance the interpretation of glycemic control. Integrating such temporally proximal metrics into clinical practice may improve individualized treatment strategies and facilitate timely therapeutic adjustments.

This study has several limitations. First, it was a single-center study with a relatively small sample size. Second, most participants had well-controlled T1DM, which could have contributed to the relatively weak associations observed between CV, TBR, and HbA1c. Third, we analyzed only 2 HbA1c measurements per participant—specifically those closest to the CGM analysis periods—which may not fully capture long-term glycemic trends. Future prospective studies involving larger, more diverse pediatric populations with varying levels of glycemic control are warranted to confirm our findings.

In conclusion, CGM-derived metrics—especially average glucose, TAR, and TIR—are significantly associated with HbA1c in Korean pediatric patients with T1DM. Among these metrics, average glucose over the preceding 4 weeks was the strongest predictor. Our findings support the integration of CGM metrics, particularly recent glycemic data, into routine clinical evaluations to improve HbA1c interpretation and enable more personalized diabetes management.

Supplementary materials

Article information Continued