From data to insights: a tool for comprehensive Quantification of Continuous Glucose Monitoring (QoCGM)

Data preprocessing and removal of duplicates

The preprocessing of the CGM signal begins with the removal of duplicate entries in the dataset. The process removes duplicate values that share identical timestamps, as these can arise during data export from CGM platforms or due to synchronization issues. This step is critical to ensure that each time point in the data is unique, preventing potential errors in subsequent analyses.

Identification of missing periods in the CGM signal

CGM systems are configured to sample glucose levels at specific intervals. For example, a common sampling frequency involves recording data every 5 min. This interval is considered the expected time between consecutive measurements.

Small deviations from the expected sampling interval can occur due to natural variability in device performance or other factors. To account for this, a tolerance range is established around the expected interval, defining the maximum allowable intervals between consecutive readings. The time differences between consecutive glucose measurements are calculated to identify intervals that fall outside the permissible % range. Any interval that is longer than the defined tolerance indicates a potential gap or missing period in the data. The specific points in time where data is missing are identified by examining the calculated time differences. These points mark the beginning of periods when the expected glucose measurements are absent.

Interpolation of missing data

To reconstruct the missing portions of the CGM signal, interpolation is performed over the identified gaps using Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) (Rabbath & Corriveau, 2019).

The final PCHIP interpolating function H(x) is defined piecewise over each interval (x_i, x_i+1) as: ${\displaystyle H\left(x\right)=H_i\left(x\right)for{x}_i\le x\le x_{i+1}.}$

This method was chosen for its ability to maintain the shape and smoothness of the original data, making it suitable for CGM signals, which typically exhibit smooth and continuous changes in glucose levels. This reconstructed signal filled the gaps identified earlier, providing a continuous and complete signal for further analysis. Figure 1 illustrates a CGM signal with missing periods, the raw signal, and the PCHIP interpolation. It is also possible to handle missing periods without interpolation by setting input argument to removing these periods.

Basic descriptive statistics

We included a range of basic descriptive statistics to characterize the CGM data for each participant. These metrics provide foundational insights into the central tendency, dispersion, and distribution of glucose values over the monitored period.

• Mean glucose: The average glucose level over the monitoring period.

• Median glucose: The midpoint of the glucose distribution, less influenced by outliers.

• Standard deviation (SD): A measure of glucose variability around the mean.

• Coefficient of variation (CV): The standard deviation normalized by the mean, expressed as a percentage.

• Interquartile range (IQR): The range between the 25th and 75th percentiles, representing the spread of the middle 50% of glucose values.

• 75th and 25th percentiles: Indicators of the upper and lower bounds of the middle half of the glucose distribution.

These metrics were calculated for the entire period, as well as separately for daytime and nighttime, to capture diurnal variations in glucose levels.

Time-in-range (TIR) metrics

TIR metrics were employed to assess the proportion of time that glucose levels remained within clinically relevant thresholds, providing a detailed view of glycemic control. The following TIR metrics were included in the analysis:

• TIR (70–180 mg/dL) (Danne et al., 2017): Percentage of time within the target range, indicating overall glycemic control.

• TITR (70–140 mg/dL) (Beck et al., 2024): Percentage of time within a narrower target range, reflecting tighter glucose control.

• TBR1 (54–70 mg/dL) (Danne et al., 2017): Percentage of time in level 1 hypoglycemia.

• TBR2 (<54 mg/dL) (Danne et al., 2017): Percentage of time in level 2 hypoglycemia.

• Total TBR (Hill et al., 2007): Combined percentage of time in TBR1 and TBR2.

• TAR1 (180–250 mg/dL) (Danne et al., 2017): Percentage of time in level 1 hyperglycemia.

• TAR2 (>250 mg/dL) (Danne et al., 2017): Percentage of time in level 2 hyperglycemia.

• Total TAR (Danne et al., 2017): Combined percentage of time in TAR1 and TAR2.

These metrics were calculated for the whole day and separately for daytime and nighttime.

Glycemic risk indicators

Glycemic risk indicators were calculated to quantify the risk associated with hyperglycemia and hypoglycemia, providing a nuanced assessment of the overall glycemic profile. The following metrics were included from the Glycemic Risk Index (GRI), Low/High Blood Glucose Index (LBGI/HBGI), and The Glycemic Risk Assessment in Diabetes (GRADE):

• Glycemic Risk Index for Hypoglycemia (GRI Hypo) (Klonoff et al., 2023): This metric quantifies the risk associated with hypoglycemic episodes, with higher values indicating a greater risk of low glucose levels.

• Glycemic Risk Index for Hyperglycemia (GRI Hyper) (Klonoff et al., 2023): This metric assesses the risk of hyperglycemia, with higher values indicating a greater risk of elevated glucose levels.

• Overall Glycemic Risk Index (GRI) (Klonoff et al., 2023): A composite metric combining the risks of both hypoglycemia and hyperglycemia, calculated as a weighted sum of GRI Hypo and GRI Hyper.

• Low Blood Glucose Index (LBGI) (Kovatchev et al., 1998): This score measures the risk of experiencing low blood glucose levels, with higher values indicating increased hypoglycemia risk.

• High Blood Glucose Index (HBGI) (Kovatchev et al., 2002): This score quantifies the risk of high blood glucose levels, with higher values indicating an increased risk of hyperglycemia.

• GRADE (Overall) (Hill et al., 2007): This metric provides a summary measure of glycemic risk across all glucose ranges, combining the risks of hypoglycemia, euglycemia, and hyperglycemia.

• GRADE for hypoglycemia (Hill et al., 2007): The specific component of the GRADE metric that assesses the risk of hypoglycemia.

• GRADE for euglycemia (Hill et al., 2007): The component of the GRADE metric that evaluates the time spent in the euglycemic range (normal glucose levels).

• GRADE for hyperglycemia (Hill et al., 2007): The component of the GRADE metric that quantifies the risk associated with hyperglycemia.

These glycemic risk indicators were calculated based on the entire CGM signal, offering a detailed view of the participant’s glycemic risk profile. A detailed mathematical definition of the metrics is provided in Supplementary Material S1.

Glycemic variability metrics

Glycemic Variability Metrics were utilized to assess fluctuations in glucose levels over time, providing insights into the stability and predictability of glycemic control. The following metrics were included:

• Continuous overall net glycemic action (CONGA) (Mcdonnell et al., 2005): This metric measures glucose variability over different time intervals. We calculated CONGA for 1-hour, 2-hour, 6-hour, and 24-hour periods (CONGA 1H, CONGA 2H, CONGA 6H, and CONGA 24H), reflecting short-term to daily fluctuations in glucose levels.

• Mean amplitude of glycemic excursions (MAGE) (Service et al., 1970): MAGE captures the average magnitude of significant glucose swings, both increases and decreases, by focusing on excursions that exceed one standard deviation from the mean. It is a widely used indicator of glycemic variability and the likelihood of large glucose fluctuations.

• Mobility: This metric assesses the signal mobility by measuring the variance of glucose changes (differences between consecutive glucose readings) relative to the overall variance of the glucose signal. Higher mobility indicates greater variability in glucose levels. The metrics have especially been shown to separate healthy form individuals with dysglycemia (Cichosz et al., 2025)

• Distance traveled per minute (DTpM) (Peyser et al., 2018): DTpM quantifies the total amount of glucose fluctuation over time by summing the absolute changes in glucose levels and normalizing by the total duration of monitoring. This metric provides a rate of glucose change, highlighting periods of rapid fluctuations. The metric is closely related to the established metric mean absolute glucose (MAG) (Hermanides et al., 2010; Kohnert et al., 2013)

• Standard deviation day-to-day (SD d2d): This metric measures the day-to-day variability in glucose levels, specifically assessing the mean glucose and TIR variability across days. It captures the consistency of glycemic control from one day to the next.

These glycemic variability metrics provide a comprehensive understanding of the dynamics of glucose levels, identifying periods of instability that may increase the risk of hypo- or hyperglycemic events and contribute to long-term complications in diabetes management. A detailed mathematical definition of the metrics is provided in Supplementary Material S1. For example, in Fig. 2, the peaks and nadirs above σ (1 SD) is highlighted in the CGM trace. The absolute excursions $\left|\lambda \right|$, both positive and negatives, are used to calculate the MAGE metric as show in the equation below: ${\displaystyle MAGE=\frac{1}{n}\sum _{i=1}^n\left|{\lambda }_i\right|if\left(\left|{\lambda }_i\right|\ge \sigma \right).}$

Glycemic control indicators

Glycemic control indicators were assessed to evaluate the effectiveness of diabetes management strategies by examining specific aspects of glucose control. The following metrics were included in the analysis:

• Fasting glucose proxy (FGxP) (Millard et al., 2020): This metric estimates the fasting glucose levels based on the CGM data. It serves as a proxy for assessing baseline glycemic control and helps to approximate the glucose levels typically observed during fasting periods.

• Glucose management indicator (GMI) (Bergenstal et al., 2018): The GMI provides an estimate of average glucose levels by applying a formula that correlates with hemoglobin A1C values. This metric offers a standardized measure of overall glycemic control and helps in evaluating the long-term efficacy of diabetes management.

• Hypoglycemia events: This metric quantify the numbers of hypoglycemic episodes below 70 mg/dL for a minimum duration of 15 min.

Several methodologies have been proposed for approximating fasting glucose levels from CGM data. We employed a robust approach developed by Millard et al. (2020) which does not require information about mealtimes. This method estimates fasting glucose by calculating the mean of the 30 lowest consecutive minutes of glucose readings during nighttime. Figure 3 illustrates the CGM readings used to estimate fasting glucose from an individual.

GMI is calculated based on Bergenstal et al. (2018): ${\displaystyle GMI\left(\text{\%}\right)=3.31+0.02392\times MeanGlucose\left[mg/dL\right].}$

Entropy and complexity measures

The entropy and complexity measure multiscale entropy (MSE) were utilized to evaluate the irregularity and unpredictability of glucose fluctuations, providing insights into the dynamic nature of glycemic control. MSE quantifies the complexity of the glucose time series by analyzing its entropy across multiple time scales. This metric captures the degree of randomness and structure within glucose data, providing a comprehensive measure of variability and the underlying complexity of glycemic fluctuations. The calculation involves embedding the data into various time scales and assessing the entropy at each scale. The sum of these entropy measures gives the Multiscale Entropy index, reflecting the overall complexity of the glucose signal. We adopted the implementation described by Kohnert et al. (2017) where Multiscale Complexity Index (MCI) is defined as the sum over the range of scales, from 1 to 7, at the window length m = 2, the sensitivity criterion r = 0.15 times the standard deviation. This metric has previously been as an early marker for changes in glucose control (Colás et al., 2019).

Tip: To reduce computational time significantly, the calculation of this metric can be omitted by removing the MSE code part.

Computation time

The average computation time for a 30-day CGM signal sampled at 288 points per day is approximately 24.3 s for the full set of metrics, and 4.6 s when excluding MSE—which is notably the most computationally intensive metric. These benchmarks were obtained on a PC with an 11th Gen Intel(R) Core(TM) i7-11850H @ 2.50 GHz processor and 32 GB of RAM.

Study sample

We used CGM data (Dexcom, G6) from the Diabetes teleMonitoring of patients in insulin Therapy (DiaMonT) trial (NCT04981808), comprising 331 people with insulin-treated T2DM (Hangaard et al., 2022; Hangaard et al., 2025). We encompassed participants from both the intervention and the control arm. Written and oral informed consent was obtained from each participant prior to their enrollment in the study. For consistency, we limited the analysis to the initial seven days of CGM monitoring post the inclusion date for all participants. We included all participants with any measurements in the first seven days. The baseline characteristics of the participants were as follows: mean age 61.3 years (SD 10.6), mean duration of diabetes 17.5 years (SD 11.5), mean BMI 33.1 kg/m² (SD 6.4), mean HbA1c 64.0 mmol/mol (SD 14.4), and 61.6% of the participants were male. QoCGM was used to calculate CGM metric based on the seven-day profile from each participant. Sampling frequency was set to 5 min, with a sampling variation allowed threshold of 0.05, morning start was set to 8AM, and PCHIP interpolation was applied on periods without CGM coverage.

Analyses

We conducted a coefficient of determination analysis (R²) to explore the linear relationships between the derived CGM metrics. R² represents the proportion of variance in one metric that is explained by another metric. R² ranges from 0 to 1, where 1 indicates that the variability in one variable can be completely explained by the other, and 0 indicates no explanatory power. This analysis provides insights into how different CGM-derived metrics are interrelated, offering guidance on their potential differences.