Sampling duration matters: Intraindividual variation of glycaemic metrics in type 1 diabetes using automated insulin delivery

Abstract

Introduction

Understanding intraindividual variability in glycaemic metrics among users of automated insulin delivery (AID) is essential for interpreting data from continuous glucose monitoring (CGM).

Methods

CGM data over 56 days were collected from 142 individuals with type 1 diabetes using AID systems. Glycaemic metrics were derived from four 14‐day periods and two 28‐day periods.

Results

The 95% limit of agreement for the difference between two measurements of time in range (TIR) from the same individual over 14‐day periods was ±13.4%. The risk that two 14‐day TIR values from the same individual differed by more than 5% points was 46.5%. If TIR was based on two 28‐day periods, the 95% limit of agreement decreased to ±9.0%. The risk that two 28‐day TIR values from the same individual differed by more than 5% points was 27.6%.

Conclusion

A 28‐day CGM sampling period is recommended to minimise random variation and obtain more reliable glycaemic metrics.

What's new?

What is already known?

• Consensus statements recommend 14‐day continuous glucose monitoring to evaluate glycaemic metrics. A difference of 5% time in range is considered clinically meaningful.

What this study has found?

• The risk that two time in range values from the same individual differed by more than 5% points was 47% if based on 14 days of glucose sampling and 28% if based on 28‐day glucose sampling.

What are the implications of the study?

• A 28‐day glucose sampling period is recommended to minimise random variation and obtain more reliable glycaemic metrics.

1. INTRODUCTION

Glycaemic metrics derived from continuous glucose monitoring (CGM) provide valuable insights for healthcare professionals and individuals with diabetes. The ambulatory glucose profile (AGP), typically generated from 14 days of CGM data as recommended, ¹ serves as a basis for clinical discussions and therapeutic advice. In this context, understanding the random variation of core metrics—including time in range (TIR), time below range (TBR) and mean sensor glucose—is essential. We have previously reported intraindividual variability in glycaemic metrics in individuals with either intermittently scanned CGM (isCGM) using multiple daily insulin injections (MDI) or continuous subcutaneous insulin infusion (CSII) without integrated CGM. ²

This study has two aims: to assess the difference in within‐subject variability of different glucose metrics across two different monitoring periods of 14 and 28 days in individuals with type 1 diabetes using automated insulin delivery (AID) pumps, and to provide information for power calculations needed to calculate the number of study participants.

2. MATERIALS AND METHODS

We have previously downloaded CGM raw data for 160 persons with type 1 diabetes who have used the AID systems Tandem Control IQ (CoIQ) or Minimed 780G (MM780G) for at least 3 months. ³ CGM sampling was truncated to obtain data for 56 days (8 weeks) before download. Glycaemic metrics were analysed across four 14‐day periods and two 28‐day periods, the most commonly used CGM download periods.

Glucose data were sampled every 5 min using the Dexcom G6 sensor (CoIQ) or Guardian 3 or 4 sensor (MM780G), and retrieved from the Glooko or Carelink software platforms, respectively. Active CGM time (%) was calculated as the proportion of 5‐min periods with a valid glucose reading, divided by the total number of intervals in 14 days (4032 = 12 × 24 × 14) or 28 days (8064), and multiplied by 100. Individuals with an active CGM time < 70% in one or more 14‐day periods were excluded (n = 12). The final study population is 148 persons with AID (CoIQ, 83 persons, MM780G 65 persons).

TIR (%) and TBR (%) over 14 and 28 days were calculated from the raw data as the proportion of 5‐min periods with glucose values between 3.9 and 10.0 mmol/L, and below 3.9 mmol/L, respectively, divided by the total number of available glucose readings and multiplied by 100.

The study was approved by the Central Denmark Region Review Board and registered in the Central Denmark Region research database as no. 1‐16‐02‐388‐22. Informed consent was obtained from the participants to use the data. No ethics approval was needed for these observational studies.

2.1. Statistical analysis

Data that were normally distributed, as assessed by visual inspection of Q‐Q plots, are presented as mean ± standard deviation (SD), and paired data were compared using Student's paired t‐test. Data that were not normally distributed are reported as median and interquartile range (IQR). TBR has a strongly skewed distribution and was log‐transformed to obtain a normal distribution after omitting TBR values <0.5%. TBR values were presented as geometric mean, ×/÷ tolerance factor.

Data from the four 14‐day periods were analysed using a linear mixed effects model with a random subject effect. This is identical to a variance component model with one within‐subject and one between‐subject variance component. Data from the two 28‐day periods were analysed similarly. We explored the possibility of a systematic period effect by extending the variance component model to include a systematic period effect. We also investigated the possibility of the within‐subject variance component being different for the two AID systems. This was done using the same linear mixed effects model but allowing for different within‐subject variation depending on the AID system. A test for equal within‐subject variation was performed by comparing the two linear mixed effects models using a likelihood ratio test.

For the normally distributed data, the probability that the difference between two measurements from the same subject exceeds a value × is given by two times the standard normal cumulative distribution function evaluated at the point ‐×/SD, where SD is the standard deviation for the difference. In this context, SD is equal to the square root of two times the standard deviation corresponding to the within‐subject variation. Similar calculations can be done for TBR on a logarithmic scale, giving probabilities that the ratio of two TBR measurements exceeds a value.

Theoretically, SD within‐subjects for 14‐day periods is √2 = 1.41 multiplied by SD within‐subjects for 28‐day periods, given the same number of values in each 14‐day period and making the simplified assumption that daily TIR and TBR are independent within each period. We compared our results with the expected ratio SD 14‐day/SD 28‐day = 1.41.

The 95% limits of agreement, representing the prediction interval for the difference between two measurements in the same individual, were calculated as 1.96 × $\surd 2$ × the within‐subject SD.

The SD of the difference between two measurements (SDD), which is necessary to perform power calculations, was calculated as $\surd 2$ x the within‐subject SD. The statistical packages SPSS ver. 20.0 was used.

3. RESULTS

Of the 148 individuals, six had one of the four 14‐day periods with active CGM time below 70%. Among the remaining 142 individuals, active CGM time was consistently high (ranging from 96.7 to 97.8). Mean age was 45.9 ± 15.0 years, and diabetes duration averaged 27.0 ± 13.6 years. The female‐to‐male ratio was 83/59; of the participants, 80 used CoIQ and 62 used MM780G. The number of glucose values in the four 14‐day periods varied between 552.243 and 553.948.

3.1. Within‐subject variation

Mean TIR for 14‐day periods remained around 71% and a within‐subjects SD of 4.84% (Table 1). The SD within‐subjects for TIR for 14‐day periods was 4.66% for the CoIQ group and 5.06% for MM780G (p = 0.24). For the combined AID group, the 95% limit of agreement of TIR was ±13.42% for repeated 14‐day measurements from the same individual. For two 28‐day periods, the 95% limit of agreement was reduced to ±9.00%. Bland–Altman plots of the difference between two TIR assessments (14 and 28 days) are depicted in Figure 1.

TABLE 1.

Glycaemic metrics for 142 persons with type 1 diabetes and pumps with automated insulin delivery. CGM data are downloaded for four periods of 2 weeks and two periods of 4 weeks.

	Period 4 before download (Weeks 7, 8)	Period 3 before download (Weeks 5, 6)	Period 2 before download (Weeks 3, 4)	Period 1 before download (Weeks 1, 2)	Between‐subjects ±SD	Within‐subjects ±SD	95% limits of agreement for the difference between two measurements
Total number of glucose values	552,243	554,222	553,938	553,498
TIR (%)	70.99 ± 13.02	70.81 ± 12.07	70.77 ± 13.24	70.89 ± 11.91	±11.60	±4.84	±13.42
TBR (%)	1.01 (0.45–2.19)	1.00 (0.47–2.46)	1.03 (0.44–2.19)	1.14 (0.35–2.09)
ln(TBR) % (TBR < 0.5% omitted)	0.48 ± 0.68	0.47 ± 0.69	0.52 ± 0.68	0.53 ± 0.67	±0.47	±0.42	±1.18
TBR (%) Geometric mean%, ×/÷ tolerance factor (TBR <0.5 omitted)	1.62 ×/÷ 1.98 (n = 103)	1.60 ×/÷ 1.99 (n = 104)	1.68 ×/÷ 1.97 (n = 103)	1.70 ×/÷ 1.96 (n = 99)	×/÷ 1.61	×/÷ 1.53	×/÷ 3.24
Mean glucose (mmol/L)	8.73 ± 1.17	8.74 ± 1.10	8.73 ± 1.19	8.71 ± 1.06	±1.05	±0.43	±1.19
Glucose CV (%)	32.73 ± 5.49	32.82 ± 5.25	32.72 ± 5.53	32.80 ± 5.39	±4.76	±2.58	±7.15
Active CGM time (%)	97.9 (96.1–98.6)	98.1 (96.7–98.8)	98.1 (96.2–98.7)	97.9 (95.6–98.6)

	Period 3 + 4 before download	Period 1 + 2 before download
	Weeks (5–8)	Weeks (1–4)
TIR (%)	70.89 ± 12.11	70.83 ± 12.05	±11.63	±3.25	±9.00
TBR (%)	1.16 (0.49–2.14) (n = 142)	1.18 (0.42–2.12) (n = 142)
ln(TBR) % (TBR < 0.5% omitted)	0.47 ± 0.66	0.53 ± 0.62	±0.53	±0.33	±0.91
TBR (%) Geometric mean%, ×/÷ tolerance factor (TBR < 0.5 omitted)	1.59 ×/÷ 1.94 (n = 105)	1.71 ×/÷1.86 (n = 103)	×/÷1.70	×/÷1.39	×/÷2.48
Mean glucose (mmol/L)	8.74 ± 1.10	8.72 ± 1.07	±1.05	±0.29	±0.80
Glucose CV (%)	33.02 ± 5.08	33.07 ± 5.13	±4.77	±1.81	±5.03

FIGURE 1.

Bland–Altman plots of the difference between two time in range (TIR) measurements for 142 individuals with type diabetes and automated insulin delivery (AID) pumps plotted against their average values. To the left is the TIR assessment based on 14‐day continuous glucose monitoring (CGM) sampling. The mean difference is −0.1% (SD 6.9%). The dashed lines denote the 95% prediction interval from −13.6% to 13.4%. To the right is the TIR assessment based on 28‐day sampling. The mean difference is −0.1% (SD 4.6%), and the 95% prediction interval is from −9.1% to 9.0%.

When TBR values <0.5% were excluded, the 95% limit of agreement for the ratio of two TBR values from 14‐day periods ranged from 3.24 to 0.31 (= 1/3.24). For 28‐day periods, the 95% limits for the ratio between two TBR values were between 2.48 and 0.40, again excluding values <0.5%. The theoretical risk that the ratio between two TBR values ≥0.5% exceeds 2 or falls below 0.5 by chance is 24.6% for 14‐day samples and 13.4% for 28‐day samples. We counted how many persons with TBR ≥0.5% for both 4‐week periods (n = 94) who at the second 4‐week period had a TBR value more than double (7 persons (7.4%)) or less than half (6 persons (6.7%)) of the TBR in the first period, showing that the empirical value (14.1% = 7.4% + 6.7%) is close to the theoretical (13.4%).

The mean sensor glucose across the four 14‐day periods was 8.7 mmol/L, with a 95% limit of agreement of ±1.19 mmol/L between the two measurements. This was reduced to ±0.80 mmol/L when based on two 28‐day periods. The glucose coefficient of variation (CV) remained stable over the four periods, with a 95% limit of agreement of ±7.15%.

The mean of glucose CV for four 14‐day periods and two 28‐day periods was 32.77 ± 4.93% and 33.05 ± 4.94%, respectively (difference 0.28 (95% CI: 0.22%–0.33%)), p < 0.001.

We found that the calculated empirical ratio between SD within‐subject for 14 days/SD within‐subjects for 28 days was 1.49 for TIR, 1.48 for mean glucose and 1.43 for glucose CV.

3.2. Risk calculation for the difference between two measurements

The probability that the difference between two 14‐day TIR values exceeds 5 or 10 percentage points by natural variability—and the corresponding probability for 28‐day periods—is presented in Figure 2. As an example, the risk that the difference between two TIR values by chance is higher than 5 percentage points is 46.5% if calculated from 14 days and 27.6% if calculated from 28 days of CGM download.

FIGURE 2.

Risk that the difference between two measurements of time in range (TIR) (Y‐axis) is larger than the limit (X‐axis) calculated for two periods of 14 days and two periods of 28 days.

3.3. Power calculations

The standard deviation of the difference (SDD) between two 14‐day TIR values was 6.85% (=$\sqrt{2}$ × 4.84). This value was reduced to 4.59% (=$\sqrt{2}$ × 3.25) for 28‐day periods (Table 1). These findings inform the design and sample size of studies comparing AID systems. Assuming a parallel‐group design comparing two AID systems using the same sensor, with a minimal relevant difference in TIR of 5%, a type 1 error risk of 5% and a type 2 error risk of 10%, a minimum of 40 persons per arm is required if based on a 14‐day CGM sampling and 18 persons if based on 28‐day CGM sampling. To detect a minimum TIR difference of 3% with the same power and significance level, the number of patients in each study arm should be 110 (14‐day CGM sampling) or 50 (28‐day sampling). The SDD for the mean glucose difference was 0.61 mmol/L for the 14‐day periods and decreased to 0.41 mmol/L for the 28‐day periods.

4. DISCUSSION

In this study of the within‐subject variation of glycaemic metrics in 142 persons with type 1 diabetes using AID, we report a substantial intraindividual variability as demonstrated in the Bland–Altman plots. Although the 95% limit of agreement for the difference between two TIR values for 14‐day periods is lower (±13.4%) than observed in users of MDI or CSII with unintegrated CGM (±17.6%), ² it should be considered when healthcare providers and users interpret CGM reports from AID. Caution is warranted when interpreting single changes in TIR of less than ±10 percentage points from 14‐day data or less than 5 percentage points from 28‐day data.

Several approaches have been proposed to determine the minimal CGM sampling period in clinical practice. Several studies have calculated the correlation (R ²) between CGM data from 7, 14 or 28 days and more extended periods, often 90 days. ⁴ , ⁵ , ⁶ , ⁷ The goal has been to identify the duration at which R ² values plateau or reach approximately 0.88. However, interpreting R ² values in a clinically meaningful context remains challenging. Recently, the correlation‐based approach has been questioned in favour of an equation‐based strategy tailored to clinical trial design. ⁸ , ⁹ , ¹⁰ Presenting 95% limits of agreement provides a straightforward and clinically intuitive way to communicate variability to AID users through standard AGP reports for 14‐ or 28‐day periods. Analysis of TBR variations is complicated by zero values and values <0.5%. In many AGP reports, TBR is unfortunately displayed without decimals, resulting in all values <0.5% being shown as zero. ¹¹ , ¹² Therefore, we calculated the 95% limits of agreement for the ratio between two TBR values after excluding those <0.5%. For individuals with a TBR value >0.5%, the 95% limit of agreement for the ratio between two TBR values was ×/÷ 3.34 for 14‐day CGM sampling and ×/÷ 2.48 for 28‐day sampling. It is reassuring that only 6.7% (=13.4/2) of persons with TBR > 0.5% (based on 28‐day periods) are expected to have a subsequent TBR more than twice the initial value.

The international consensus panel ¹ recommends a minimum CGM sampling duration of 14 days, based on two correlation‐based studies. ⁴ , ⁶ Recent evidence supports the adoption of a 28‐day standard CGM sampling period in both research and clinical practice to provide a more robust assessment of hypoglycaemic events. ¹³ , ¹⁴ Our findings support the use of 28‐day CGM sampling as the default approach. In this study, we present SD within‐subjects for the two commonly used CGM sampling periods in clinical diabetes care (14 and 28 days). Still, the theoretical SD can be calculated for any period.

Glucose CV reflects a combination of intra‐ and interday variability. The contribution of interday variation increases with longer sampling durations. ¹⁵ , ¹⁶ This may explain why the glucose CV calculated over 28 days was marginally higher than that for 14‐day periods.

The computed empirical ratio of SD within‐subjects from 14 and 28 days was close to but not precisely the theoretical reference value of 1.41 due to a slight variation in the number of glucose values between the four 14‐day periods.

The SDD for two TIR measurements and two mean glucose measurements from 14‐day periods were 6.85% and 0.61 mmol/L, respectively—approximately half the values reported previously for unintegrated CGM. ² This profoundly influences the sample size for parallel studies comparing AID systems with different sensors with available baseline and follow‐up values. Due to differing sensor performances, participants in one study arm must wear both their own AID‐connected system and a blinded version of the comparator group's sensor. Wearing two sensors for more than 14 days is generally considered unrealistic for most individuals. However, our results show that a sample size of 40 participants per arm is sufficient to detect a minimum 5 percentage point change in TIR, assuming the actual mean value of TIR is comparable to those in the present study. If no baseline is available (as with newly diagnosed type 1 diabetes), sample size depends on SD of TIR at follow‐up, which was nearly identical for 14 and 28 days of CGM sampling.

It is a strength of the study that we have downloaded raw glucose data and do not depend on software‐specific calculations attributed to the different platforms. ¹² We did not find any statistically significant difference in the within‐subject variation of TIR between the two AID systems, but it is a limitation that the results cannot be extrapolated to users of a pump with unintegrated CGM, and we have not performed a subgroup analysis regarding age. It is also a limitation that the four 14‐day periods are only separated by a maximum of 28 days, and the second 28‐day period is an extension of the first. In real life, CGM metrics are usually compared with a time distance of 3 to 6 months, which will increase the limit of agreement between two assessments of TIR. Our calculation of sample size is only valid for studies with relatively short intervention periods; for more extended periods, the sample size will increase.

5. CONCLUSIONS

Within‐subject variation in TIR and mean sensor glucose is substantial among AID users. These findings should inform clinical interpretation and guidance when comparing two CGM downloads. This is preferably based on 28‐day rather than 14‐day sampling periods to obtain a reasonable compromise between access to the newest data and reducing natural variation.

FUNDING INFORMATION

The authors have nothing to report.

CONFLICT OF INTEREST STATEMENT

KWH has received research grants for an investigator‐initiated study from Abbott Diabetes Care and Novo Nordisk.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.