Glycemic Matrix and Segmentation: A New Metabolic Visualization and Analysis Tool

New technologies arise the opportunity to understand the complex systemic background of multidimensional diseases and allow a personalized approach. Continuous Glucose Monitoring (CGM) sensors and their broad use have been key in the discovery of the metabolic heterogeneity surrounding many disorders such as diabetes type II, and has placed the scientific community a step closer to determining which factors are contributing to their complications and evolution. However, gathering data extending beyond glucose levels linked to lifestyle factors, such as nutrition, physical activity, sleep quality, and stress, poses a significant challenge in terms of representation, considering the substantial amount of data involved. To comprehend the relationship between these variables in a practical manner that empowers individuals to make choices enhancing their quality of life, there is a need for new graphics. These graphics would enable the observation of the overall framework in a contextualized manner and assist in establishing clear visual goals. This article introduces glycemic matrix and metabolic segmentation, a new method for representing and evaluating functional profiles by combining glucose and lifestyle data. In this early-phase trial, the potential of this approach to represent the complete glycemic spectrum within its context and adapt to a diverse range of objectives is demonstrated. Moreover, it is a promising tool to finally be able to cluster metabolic types through artificial intelligence (AI) and adapt clinical interventions to the metabolic heterogeneity. This research is a private research under Glucovibes companies R&D initiatives.

In recent years, the increasing availability of glucose data, thanks to the wider accessibility of CGM devices [1], has significantly expanded our understanding of glucose metabolism, not only in pathological scenarios like diabetes but also in populations traditionally considered normoglycemic [2,3]. However, the vast amounts of data generated require the development of visualization and analysis techniques that can help interpret this information in an actionable way, both for clinicians and patients [2,4].

Currently, glycemic control is often evaluated using HbA1c, an indirect measure reflecting average glucose levels over the past 2-3 months. However, meaningful differences have been found between HbA1c and mean plasma glucose [5]. The advent of CGM has provided access to continuous glucose monitoring, enabling the calculation of metrics such as Glucose Variability (GV) and allowing for the direct observation of daily glucose profiles, including the number, frequency, severity, and timing of hypoglycemia and hyperglycemia episodes [6]. These factors are crucial for the precise evaluation of glucose metabolism [7]. As the use of CGM continues to expand in clinical practice, Time In Range (TIR) has emerged as a key metric, defined as the percentage of time glucose levels remain within specified ranges [2,8].

Glycemic variability has been linked to various adverse outcomes in patients with diabetes, from mild to life-threatening effects [9-11]. TIR has already been associated with cardiovascular disease mortality in patients with type II diabetes [12], as well as the risk of developing or progressing retinopathy and microalbuminuria [4]. GV also contributes significantly to oxidative stress [13] and is an important risk factor in the pathogenesis of vascular complications in diabetes [14]. This highlights the importance of maintaining stable glucose levels.

However, TIR lacks the ability to fully represent glucose curve variability, as it relies on predefined glucose target ranges that may not be applicable to all individuals. Metabolic diseases such as diabetes have been shown to be highly heterogeneous [15], necessitating the development of metrics and visualizations that allow for the study and adjustment of target glucose values in an individualized manner [2,12].

To address this need, Matabuena M, et al. [16] proposed Glucodensities, a distributional measure similar to TIR but with narrower glucose ranges, representing the distribution of glucose concentrations over time. This provides a more precise illustration of glucose metabolism. However, this model is limited by its focus on glucose distribution alone, without considering the context or order of events.

Metabolic disorders are closely linked to lifestyle factors [17], and there is growing recognition of the need to move beyond a glucocentric approach to advance precision medicine [18]. From a clinical perspective, it is critical that both clinicians and patients understand the unique metabolic profile of each individual. This requires tools that integrate and represent the context of glucose data, including meals, physical activity, mood, and rest. Additionally, CGM technology allows access to real-time metabolic information, enabling immediate intervention adjustments [19]. Zeevi D, et al. [20] demonstrated that variables such as nutrition provoke different interpersonal responses in blood glucose. Integrating glucose monitoring with these aspects can shed light on the unique metabolic functioning of each individual, facilitating a personalised approach.

Metrics and visualizations that integrate CGM data with lifestyle factors, while considering the frequency, severity, and timing of hypoglycemia and hyperglycemia, are essential for capturing the complete metabolic picture. In this paper, we apply glucose densities to Glucovibes' extensive CGM dataset and explore a methodological approach that incorporates context, thereby enriching glucose density representations. We also examine how this representation can be transformed into a metric that enables AI to classify metabolic types and assign target glucose density curves for different profiles [21].

This research represents an innovative approach being used by Glucovibes to understand the glycemic evolution of users across different segments of the day and under varying lifestyle conditions. As a technology-based nutritional advisory company, Glucovibes is committed to helping users understand how their metabolism responds to daily stimuli, empowering them to make informed decisions about their health and well-being. Data representation and the development of informative metrics are key to achieving this goal.

Blood glucose level representation

Blood glucose levels (mg/dL) are typically represented over time, providing a local visualisation that is useful for understanding specific events. The variables that can correspond to a timeline include:

• Peak: Maximum value (mg/dL)
• Amplitude: Difference between peak value and initial value (mg/dL)
• Stabilization: Difference between initial and last value (mg/dL)
• AUC: Area Under the Curve (mg/dL*minute)

Global visualization through glucose density

To profile metabolic behavior, a global visualization is more suitable. By representing glucose density, the distribution of glucose values within the chosen time frame can be observed in relation to the mean. If the mean and the distribution peak position are similar, the distribution is centered. If not, there is bias. The width of the density curve reflects glucose variability, with a higher peak indicating greater glucose stability.

Data collection

This research is conducted as part of Glucovibes' R&D initiatives. Glucovibes, a biotechnology company, focuses on understanding metabolism through lifestyle and biosensor data. The company analyses these parameters with different objectives, depending on the user: performance and recovery optimization for athletes, lifestyle and well-being for the general population, and clinical management for patients with metabolic disorders.

Data analysis

Glucose point frequency collected from CGM sensors varies, with one value point recorded every minute from the last 15 minutes, and one value point every 15 minutes from the last 8 hours. In order to standardize data frequency, glucose values were filtered for gaps of 15 minutes in length. Windows with larger gaps were excluded from the analysis. Monitoring periods that covered at least 70% of the time for glucose values were included in the analysis.

Density calculation

All analyses were performed using R statistical suite version 3.6.3. Kernel density estimation was applied, a non-parametric method to estimate the probability density function of a random variable based on kernels, in order to smooth the curve (Figure 1) [22]. The bandwidth used was 5 and 512 density points were calculated between the glucose values of 55 mg/dL and 250 mg/dL.

Figure 1: Comparing time frames for estimation of glucose values.
This graphic shows the estimated glucose value distribution curve for different fixed time measures (1, 5 and 15 minutes). The narrower the time window, the more fluctuation we can observe in the distribution curve and change is more defined and the higher is the maximum density value. Anyway, density is very similar in all of them.

Clustering and validation

The data was clustered using k-means with a maximum of 10 iterations. A training and testing set (80% of the data in the training set) were defined to analyse the representativeness of the results.

Analysis

First of all, density distribution for each included period was calculated from the estimated glucose data. Estimated glucose values were used to define column limits of the Glycemic Matrix through k-means clustering. Density distribution AUC of the resulting columns were used to define row limits of the Glycemic Matrix through k-means clustering. These limits were validated through the comparison of a training and a testing set. After defining the Glycemic Matrix, estimated glucose values are classified in order to the context they belong to and 3 density curves are calculated by period, one for each segment: Day with non-physical activity (this includes events such as stress, meals…), physical activity and night. This is called the Metabolic Segmentation. As a result, each density curve gets an AUC evaluation (row) for each column.

The 14 values (7 columns for each of the 3 density curves) for each period will be the input for a new clustering in order to classify metabolic types. These metabolic clusters will then be represented and compared.

Metabolic segmentation for profiling: A new approach

In order to understand, classify and tag all the glycemic information of an individual, the overall glucose levels or densities are not enough, especially if it is intended to generate a dynamic database that stores information that could be interpreted by AI. AI should be a potential partner for massive and real time analysis once a clear way to classify this information is adopted, and this is the aim of the segments. These segments are explained next.

To avoid losing information about the context, and based on the understanding of the human physiology, different segments that allow the study of metabolic variations on glucose response to different parameters and circumstances were defined:

• All: All the estimated glucose points. It represents the whole general framework. Up to date, the most advanced representations (Glucodensity or TIR) use this approach and it has been tested as an interesting but limited indicator for precise analysis.
• Night: Estimated glucose values between midnight and 6 in the morning as a fixed time window. Despite the awareness that not every user has the same sleep habits, most of the analyzed subjects sleep between 12pm and 6am, so this time window was selected as a first test for this method. With this segmentation the aim was to analyse and classify the glycemic behavior under recovery and sleep time, where a different behavior from the day was expected. At night and under recovery the human body should have more stable glycemic curves. In some cases, we could use this segment to analyse the recovery and energy needs of subjects under study. Potential improvement of this segment involves information and habits from questionnaires or sleep tracking devices.
• Day non-physical activity: All the estimated glucose points from 6am to midnight but the values within activity during the day. It is where all the intakes and other events happen. It is known that stress, work, meals, and other situations could generate a direct impact on glucose levels and under this segmentation all these criteria but physical activity should be found, which is explored in a new segment.
• Physical activity: Only values within physical activity of any intensity, from walking, pilates or dancing, to a cycling race or a marathon.  It represents metabolic response to physical activity and allows a better understanding of the impact of intensity during physical activity in glycemia. The app reports heart rate values during sport integrated directly from activity tracking devices such as smart watches, bands, rings, and tracking apps, and even the Rate of Perceived Exertion (RPE) from 1 to 10. Thanks to this segment high glucose values can be identified to be due to meals or activity.

These segments allow different scopes of analysis, and the addition of each one of them brings new insights to understand the metabolism of each subject in the study, as the glycemic behavior can vary depending on the segment.

For example, some questions that can be answered are:

With the night segment curve:

• Is glycemia stable at night?
• Does the body have enough energy to keep glycemia stable at night, showing a good glucose recovery?
• Are hypoglycemic events present during sleep time?

With the physical activity segment curve:

• How does intensity affect glycemia for each individual?
• Are glucose levels stable during activity?
• Where are glucose values located mainly during activity for each individual on each session?
• Can differences be seen related to activity intensity? (Can a first hump for low intensity and another hump for high intensity activities be observed or are both states mixed?)

And with the day segment the traditional view of glycemic response to meals and other events that occur in this period can be understood, as the physical activity and resting time values are not taken into account.

Glycemic Matrix (GM)

The clustering of the columns is performed with the standardized values of glucose which adopt values between 55 and 250. The resulting columns constitute the following ranges of glucose values: From 55 to 79 (column 1), from 80 to 92 (column 2), from 93 to 103 (column 3), from 104 to 116 (column 4), from 117 to 134 (column 5), from 135 to 166 (column 6), and from 167 to 250 (column 7).

The clustering of the rows is performed with the column AUC values. AUC values go from 0 to 1. The resulting rows constitute the following ranges of AUC values: From 0 to 0.045 (row F), from 0.045 to 0.119 (row E), from 0.119 to 0.204 (row D), from 0.204 to 0.294 (row C), from 0.294 to 0.412 (row B), and from 0.412 to 1 (row A). The area under the curve of each column gives an estimation of the sum of the densities of that value range.

The Glycemic Matrix, which clusters glucose values and evaluates AUC within different columns and rows, allows for a comprehensive analysis of glycemic behavior. This tool enables both the classification of metabolic types and the visualization of results in an understandable format, making it easier to observe the effects of lifestyle interventions.

As can be seen in figure 2:

Figure 2: Glycemic matrix and metabolic segmentation representation options.

1. Individual overall glucose density curve. Each cell of the Glycemic Matrix is labelled with their glucose and AUC range (in a 2D approach).
2. Individual overall density curve classification in the Glycemic Matrix display. The columns belong to the glucose ranges (left to right from 1 to 6) while the rows belong to the density ranges (up to down from A to F).
3. Individual overall density curve classification in the Glycemic Matrix in one line. AUC range is represented through color and label instead of position.
4. Individual segmented glucose density curves. Each cell of the Glycemic Matrix is labelled with their glucose and AUC range (in a 2D approach).
5. Individual segmented density curves classification in the Glycemic Matrix stacked one above the other. AUC range is represented through color and label.
6. Overall glucose density curves for all the periods. Each cell of the Glycemic Matrix is labelled with their glucose and AUC range (in a 2D approach).
7. Glycemic Matrix for all the period glucose densities. The columns belong to the glucose ranges (left to right from 1 to 6) while the rows belong to the density ranges (up to down from A to F). The shade represents the percentage of curve density values in each density range for each glucose range.
8. Overall density curve classification in the Glycemic Matrix for all the periods stacked horizontally (each column represents the Glycemic Matrix classification for one period and the rows correspond to the Glycemic Matrix glucose ranges up to down from 1 to 6). AUC range is represented through color instead of position.

Starting from the density curve of standardized glucose values and representing these in the Glycemic Matrix, combined with Metabolic Segmentation, provides a powerful tool for studying individual metabolic profiles. This approach enables the observation of glucose variability and progression over time, while visually establishing goals and tracking progress towards them. It also opens the possibility for AI-driven classification of metabolic profiles, which could be a critical step towards glycemic response modelling and prediction.

Biomedical research has often focused on identifying differences between subgroups within the population. However, this approach faces challenges when dealing with complex multifactorial disorders like diabetes, where lifestyle and environmental factors play substantial roles. The advent of biotechnology, particularly real-time data acquisition through tools like CGM, has provided unprecedented opportunities for designing personalised interventions aimed at improving well-being and quality of life. This wealth of data, however, necessitates the development of tools that can extract and present this information in a meaningful way, accessible to both clinicians and patients [1].

The Glycemic Matrix and Metabolic Segmentation represent significant advancements in the interpretation of glucose metabolism. These tools allow for a comprehensive analysis of glucose data within its contextual framework, providing a clearer understanding of how different lifestyle factors impact glycemic control. This holistic approach could lead to more effective management strategies for conditions like diabetes, where precise control of blood glucose levels is crucial [2].

Moreover, these tools have substantial educational value. By offering clear visualizations of the impact that lifestyle choices have on glucose levels, they empower individuals to make informed decisions about their health. This is particularly important in the management of type II diabetes, where lifestyle modifications are often the first line of intervention [23,24].

Future work will focus on refining these tools and expanding their applications. One promising area of research involves integrating machine learning and AI to identify patterns and clusters within the data, allowing for the classification of metabolic types and the prediction of glycemic responses. This could enable the development of more personalised and dynamic intervention strategies, further advancing the field of precision medicine [25].

Additionally, further validation of the Glycemic Matrix and Metabolic Segmentation across diverse populations will be essential to ensure their generalizability and effectiveness. Research will also explore the potential of these tools to predict long-term outcomes, such as the progression of metabolic disorders and the effectiveness of various interventions over time.

In conclusion, the Glycemic Matrix and Metabolic Segmentation offer a novel and comprehensive approach to understanding glucose metabolism. By integrating lifestyle context with continuous glucose data, these tools provide valuable insights that can inform both clinical practice and individual health management. As these tools continue to evolve, they hold the potential to transform the way metabolic health is understood and managed, paving the way for more personalised and effective interventions in the future.

1. Rodbard D. Continuous Glucose Monitoring: A Review of Successes, Challenges, and Opportunities. Diabetes Technol Ther. 2016 Feb;18 Suppl 2(Suppl 2):S3-S13. doi: 10.1089/dia.2015.0417. PMID: 26784127; PMCID: PMC4717493.
2. Battelino T, Danne T, Bergenstal RM, Amiel SA, Beck R, Biester T, Bosi E, Buckingham BA, Cefalu WT, Close KL, Cobelli C, Dassau E, DeVries JH, Donaghue KC, Dovc K, Doyle FJ 3rd, Garg S, Grunberger G, Heller S, Heinemann L, Hirsch IB, Hovorka R, Jia W, Kordonouri O, Kovatchev B, Kowalski A, Laffel L, Levine B, Mayorov A, Mathieu C, Murphy HR, Nimri R, Nørgaard K, Parkin CG, Renard E, Rodbard D, Saboo B, Schatz D, Stoner K, Urakami T, Weinzimer SA, Phillip M. Clinical Targets for Continuous Glucose Monitoring Data Interpretation: Recommendations From the International Consensus on Time in Range. Diabetes Care. 2019 Aug;42(8):1593-1603. doi: 10.2337/dci19-0028. Epub 2019 Jun 8. PMID: 31177185; PMCID: PMC6973648.
3. Hall H, Perelman D, Breschi A, Limcaoco P, Kellogg R, McLaughlin T, Snyder M. Glucotypes reveal new patterns of glucose dysregulation. PLoS Biol. 2018 Jul 24;16(7):e2005143. doi: 10.1371/journal.pbio.2005143. PMID: 30040822; PMCID: PMC6057684.
4. Beck RW, Connor CG, Mullen DM, Wesley DM, Bergenstal RM. The Fallacy of Average: How Using HbA1c Alone to Assess Glycemic Control Can Be Misleading. Diabetes Care. 2017 Aug;40(8):994-999. doi: 10.2337/dc17-0636. PMID: 28733374; PMCID: PMC5521971.
5. Shrom KC, Arbeláez AM. Discrepancy between hemoglobin A1c and average glucose based on continuous glucose monitoring data in non-diabetic adults: Are we still missing postprandial hyperglycemia? Journal of Diabetes Science and Technology. 2010;4(3):543-548.
6. Rodbard D. Interpretation of continuous glucose monitoring data: glycemic variability and quality of glycemic control. Diabetes Technol Ther. 2009 Jun;11 Suppl 1:S55-67. doi: 10.1089/dia.2008.0132. Erratum in: Diabetes Technol Ther. 2018 Apr;20(4):320. doi: 10.1089/dia.2008.0132.correx. PMID: 19469679.
7. Nathan DM, Buse JB, Davidson MB, Ferrannini E, Holman RR, Sherwin R, Zinman B; American Diabetes Association; European Association for Study of Diabetes. Medical management of hyperglycemia in type 2 diabetes: a consensus algorithm for the initiation and adjustment of therapy: a consensus statement of the American Diabetes Association and the European Association for the Study of Diabetes. Diabetes Care. 2009 Jan;32(1):193-203. doi: 10.2337/dc08-9025. Epub 2008 Oct 22. PMID: 18945920; PMCID: PMC2606813.
8. Shah VN, Shoskes A, Tawfik P. Continuous glucose monitoring versus blood glucose monitoring in adults with type 1 diabetes using insulin injections: A systematic review and meta-analysis. Diabetes Care. 2019;42(9):1717-1724.
9. Brod M, Christensen T, Thomsen TL, Bushnell DM. The impact of non-severe hypoglycemic events on work productivity and diabetes management. Value Health. 2011 Jul-Aug;14(5):665-71. doi: 10.1016/j.jval.2011.02.001. PMID: 21839404.
10. Davis SN, Shavers C, Costa F. Prevention of severe hypoglycemia by methylprednisolone during an acute episode of hypoglycemia in patients with type 1 diabetes mellitus. The Journal of Clinical Endocrinology & Metabolism. 2005;90(4):2057-2061.
11. Jauch-Chara K, Oltmanns KM. Hypoglycemia and Cognitive Dysfunction in Diabetic Patients. Diabetes, Obesity and Metabolism. 2007;9(4):523-531.
12. Lu J, Ma X, Shen Y. Time in Range Is Associated with Cardiovascular Outcomes in Type 2 Diabetes Patients. Journal of Clinical Endocrinology & Metabolism. 2020;105(7):161.
13. Monnier L, Mas E, Ginet C, Michel F, Villon L, Cristol JP, Colette C. Activation of oxidative stress by acute glucose fluctuations compared with sustained chronic hyperglycemia in patients with type 2 diabetes. JAMA. 2006 Apr 12;295(14):1681-7. doi: 10.1001/jama.295.14.1681. PMID: 16609090.
14. Hirsch IB. Glycemic Variability and Diabetes Complications: Does It Matter? Of Course It Does! Diabetes Care. 2015 Aug;38(8):1610-4. doi: 10.2337/dc14-2898. PMID: 26207054.
15. Tuomi T, Santoro N, Caprio S, Cai M, Weng J, Groop L. The many faces of diabetes: a disease with increasing heterogeneity. Lancet. 2014 Mar 22;383(9922):1084-94. doi: 10.1016/S0140-6736(13)62219-9. Epub 2013 Dec 3. PMID: 24315621.
16. Matabuena M, Petersen A, Vidal JC, Gude F. Glucodensities: A new representation of glucose profiles using distributional data analysis. Stat Methods Med Res. 2021 Jun;30(6):1445-1464. doi: 10.1177/0962280221998064. Epub 2021 Mar 24. PMID: 33760665; PMCID: PMC8189016.
17. Conn VS, Hafdahl AR, Brown LM. Meta-analysis of quality-of-life outcomes from physical activity interventions. Nurs Res. 2009 May-Jun;58(3):175-83. doi: 10.1097/NNR.0b013e318199b53a. PMID: 19448521; PMCID: PMC3159686.
18. Hulman A, Simmons RK, Brunner EJ. Beyond Glucose: The Role of Physiological Factors in the Diabetes Risk Continuum. The Lancet Diabetes & Endocrinology. 2021;9(3):158-168.
19. Lal RA, Maahs DM. Clinical Use of Continuous Glucose Monitoring in Pediatrics. Diabetes Technol Ther. 2017 May;19(S2):S37-S43. doi: 10.1089/dia.2017.0013. PMID: 28541138; PMCID: PMC5444498.
20. Zeevi D, Korem T, Zmora N, Israeli D, Rothschild D, Weinberger A, Ben-Yacov O, Lador D, Avnit-Sagi T, Lotan-Pompan M, Suez J, Mahdi JA, Matot E, Malka G, Kosower N, Rein M, Zilberman-Schapira G, Dohnalová L, Pevsner-Fischer M, Bikovsky R, Halpern Z, Elinav E, Segal E. Personalized Nutrition by Prediction of Glycemic Responses. Cell. 2015 Nov 19;163(5):1079-1094. doi: 10.1016/j.cell.2015.11.001. PMID: 26590418.
21. Buschur EO, Faulds E, Dungan K. CGM in the Hospital: Is It Ready for Prime Time? Curr Diab Rep. 2022 Sep;22(9):451-460. doi: 10.1007/s11892-022-01484-x. Epub 2022 Jul 7. PMID: 35796882; PMCID: PMC9261155.
22. R documentation: Kernel density estimation.
23. Allen NA, Fain JA, Braun B, Chipkin SR. Continuous glucose monitoring counseling improves physical activity behaviors of individuals with type 2 diabetes: A randomized clinical trial. Diabetes Res Clin Pract. 2008 Jun;80(3):371-9. doi: 10.1016/j.diabres.2008.01.006. Epub 2008 Mar 4. PMID: 18304674; PMCID: PMC2430041.
24. Fitipaldi H, McCarthy MI, Florez JC. Genetic basis of type 2 diabetes and its implications for precision medicine. Nature Medicine. 208;24(11):1793-1800.
25. Dennis JM, Shields BM, Henley WE. Precision Medicine in Diabetes: A Population-Based Study of Individualized Treatment Responses. Diabetes Care. 2019;42(2):362-369.