Glucose Color Index: Development and Validation of a Novel Measure of the Shape of Glycemic Variability

Abstract

Background: Standard continuous glucose monitoring (CGM) metrics: mean glucose, standard deviation, coefficient of variation, and time in range, fail to capture the shape of variability in the CGM time series. This information could facilitate improved diabetes management.

Methods: We analyzed CGM data from 141 adults with type 2 diabetes in the Hyperglycemic Profiles in Obstructive Sleep Apnea (HYPNOS) trial. Participants in HYPNOS wore CGM sensors for up to two weeks at two time points, three months apart. We calculated the log-periodogram for each time period, summarizing using disjoint linear models. These summaries were combined into a single value, termed the Glucose Color Index (GCI), using canonical correlation analysis. We compared the between-wear correlation of GCI with those of standard CGM metrics and assessed associations between GCI and diabetes comorbidities in 398 older adults with type 2 diabetes from the Atherosclerosis Risk in Communities (ARIC) study.

Results: The GCI achieved a test-retest correlation of R = .75. Adjusting for standard CGM metrics, the GCI test-retest correlation was R = .55. Glucose Color Index was significantly associated (p < .05) with impaired physical functioning, frailty/pre-frailty, cardiovascular disease, chronic kidney disease, and dementia/mild cognitive impairment after adjustment for confounders.

Conclusion: We developed and validated the GCI, a novel CGM metric that captures the shape of glucose variability using the periodogram signal decomposition. Glucose Color Index was reliable within participants over a three-month period and associated with diabetes comorbidities. The GCI suggests a promising avenue toward the development of CGM metrics which more fully incorporate time series information.

Keywords: CGM metrics; diabetes comorbidities; glycemic variability; type 2 diabetes.

Figure 1.

Diagram of steps in the calculation of GCI. We collect and clean the raw CGM data, calculate the log-periodogram for each CGM wear, summarize the result using disjoint linear model fits, and then combine them via linear combination with weights determined by canonical correlation analysis to produce a “raw” GCI. The weights and corresponding log-periodogram summary values are referred to using acronyms representing the frequency bands and summary values of interest. The initial characters of L, I, and S refer to long, intermediate, and short periodicity, respectively, while the secondary acronym characters MP and S refer to midpoint and slope. The optional standardization procedure is not pictured here but will be performed after Step 4. Finally, the resulting metric values are validated.

Figure 2.

Illustration of GCI and normalized CGM data (subset days 3-5 of wear, chosen at random) in 10 ARIC participants with diabetes. The left column corresponds to higher GCI, while the right column corresponds to lower GCI. Continuous glucose monitoring time series data are standardized to glucose Z-scores (mean value 0 and standard deviation of 1). The mean at 0 is indicated in each time series by the blue line. This standardization ensures that each of the 10 individuals has effectively equivalent features according to the standard CGM metrics.

Figure 3.

Test-retest correlations of GCI and standard CGM metrics in HYPNOS. The faceted figure presents raw (left column) and corresponding mutually adjusted (right column) scatter plots of the follow-up versus the baseline value for each metric. The blue solid line in each panel is the linear regression line, while the blue dotted line is the line of identity, y = x. Test-retest correlation R is marked.

Figure 4.

Associations between CGM metrics, including GCI, and comorbidities in ARIC participants with diabetes. Model 1 adjusts for the demographic factors of age, sex, and race. Model 2 adjusts for these demographic features as well as the duration of diabetes and the use of common diabetes medications (insulin, sulfonylurea, and metformin). Model 3 incorporates all adjustment from Model 2 as well as mutual adjustment for the other CGM metrics. Standard deviations were 17.0 for GCI, 15.4% for TIR, 26.4 mg/dL for mean glucose, and 6.7% for CV. Note that the standard deviation of glucose is not considered here due to it being a secondary variability metric to CV, to which it is also highly colinear.