dkNET New Investigator Pilot Program in Bioinformatics Awardee Webinar Series
Presenter: Joon Ha, PhD. Associate Professor, Department of Mathematics, Howard University, Washington DC.
Abstract
The most common form of diabetes, type 2 diabetes (T2D) is a failure of insulin-secreting pancreatic beta-cells to increase insulin to the level required to maintain normal blood glucose. Thus, identifying beta-cell function and insulin sensitivity in those who are at high risk is crucial to preventing and delaying the disease. Hyper-glycemic clamp and euglycemic hyper- insulinemic clamp are considered to be gold standard measures for these quantities. However, these two methods demand highly skilled labor and thus are cost-prohibitive. Glucose challenge tests have been used to estimate beta-cell function and insulin sensitivity. The product of beta-cell function and insulin sensitivity, termed the disposition index (DI), is of great value because it measures beta-cell function relative to insulin requirements. However, glucose challenge tests are expensive and time-consuming and therefore impractical to implement in large-scale clinical studies. To address this challenge, we developed a model disposition index (mDI estimated without insulin) that does not require insulin measurements during an oral glucose tolerance test (OGTT) (Ha et al., Diabetes 2021 (70) suppl. 1). mDI outperforms the conventional oral disposition index (oDI) at predicting progression to diabetes.
To further increase access and refine the assessments of beta-cell function, we are adapting our model to calculate a model disposition index using continuous glucose monitoring (CGM). CGM has been in the spotlight of diabetes management and has revolutionized the field of medicine as they are approved for glucose monitoring and clinical decision-making in patients with diabetes. CGM devices are relatively inexpensive compared to oral glucose challenge tests, accessible, and simple to use, especially in remote or free-living environments. The CGM device continuously measures interstitial glucose every 5 minutes and provides glucose profiles for 7-14 days. Thus, there are numerous data points compared to glucose challenge tests, but the abundant data points have not previously been used for estimating metabolic parameters. We compared mDI to two widely used CGM-derived metabolic parameters for assessing metabolic status and risk, mean glucose and glycemic excursion. Both mean glucose and glycemic excursion correlated strongly with mDI. The new approach promises to be cost- effective and easy to perform and therefore implementable in large-scale clinical studies. As for specific clinical applications, estimated model parameters during OGTTs identified ethnic differences in common pathways to T2D between Pima Indians and Koreans.
Upcoming webinars schedule: https://dknet.org/about/webinar
WASP-69b’s Escaping Envelope Is Confined to a Tail Extending at Least 7 Rp
dkNET Webinar: Estimating Relative Beta-Cell Function During Continuous Glucose Monitoring and Its Clinical Applications 03/10/2023
1. Joon Ha
Department of Mathematics, Howard University, Washington DC.
1
Estimating Relative Beta-Cell Function During Continuous
Glucose Monitoring and Its Clinical Applications
2. Glucose Pattern Leading to Type 2 Diabetes (T2D)
Glucose does not increase much until reaching the threshold and sharply rises
2
Mason et al., Diabetes 2007;56:2054-2061
mean of 50 Pima Indians
Good Biomarker?
3. What is a good biomarker?
• Robustness to detect progression to the disease; already
substantially changed before onset of the disease (counter
example; blood glucose), leading to an early biomarker
• Prediction of the disease, typically assessed by Receiver Operating
Characteristic Curve (ROC) or Survival Analysis (longitudinal),
leading to a good predictor
• Identifying metabolic characteristics, leading to personalized
therapies
5. Contents
1. Summary of current metabolic parameter surrogates during various glucose
challenge tests
2. Novel marker during standard oral glucose tolerance tests (OGTTs)
3. Novel marker during Continuous Glucose Monitoring (CGM)
4. Clinical applications:
A) Personalized intervention
B) Detecting subjects at high risk on a plane of disposition index
C) Heterogeneity and Homogeneity across Ethnics
6. Gold standard measurements of metabolic parameters
• Insulin resistance and β-cell dysfunction are key pathophysiological factors
for onset of type 2 diabetes (T2D)1
• Two clamp experiments are generally accepted as the “gold standard” to
measure the two risk factors2,3
• β-cell function relative to insulin sensitivity (i.e., Disposition Index [cDI] =
insulin sensitivity x insulin secretion) is considered the strongest metabolic
predictor for T2D4,5
• However, not practical for large-scale studies; up to 4 hours experiment time
and highly skilled labor and thus are cost-prohibitive
1Hannon., Ann. N.Y. Acad. Sci, 2015; 2Arslanian S., Horm Res, 2005; 3Sjaarda L., Diabetes Care, 2013
4Bergman N. R. et al., Diabetes, 2002; 5Utzschneider M.K., Diabetes Care, 2013
Hyperinsulinemic-Euglycemic clamp: Peripheral insulin sensitivity
Hyperglycemic clamp: Beta-cell function
7. Frequently sampled intravenous glucose tolerance tests (FSIGT)
• Blood samples collected every 2 min to 30 minutes during two hours: 22
Glucose and Insulin measurements
• Beta-cell function: Acute insulin response to Glucose (AIRg), AUC I of the first
10 minutes
• Insulin sensitivity: fitting the minimal math to G, MINMOD, R. Bergman;
Banting Medalist
IGT
G
DI
High Risk
Insulin Sensitivity (10-4/(µU/ml)*min)
Insulin
Secretion
(
µ
U/ml)
G constant Normal G
SI =2.0
AIRG =400
DI=800
SI =0.4
AIRG =2000
DI=800
AIR
G
• DI originally derived
from Intravenous
Glucose Tolerance Test
(IVGTT)
8. Current Metabolic Estimates during OGTTs
• Insulin Resistance or Sensitivity: HOMA-IR, QUICKI, Matsuda Index
• Beta-cell Function: HOMA-B, Insulinogenic Index(IGI)
• Relative Beta-Cell Function: fDI(HOMA-B*(1/HOMA-IR), oDI (IGI*Matsuda)
• All are built by algebraic formulae using single and average measurements
rather than glucose and insulin profiles
• Current mathematical model-derived estimates require frequently
sampled OGTTs up to 6 hours (11 G, 11 I, 11 C-peptide points), Cobelli’s
model and Mari’s model; not applicable for large-scale epidemiological
studies.
• Standard OGTTs (G, I at 0,30,60,90,120) that underwent under clinical
settings. However, the current math models (Cobelli and Mari) are not
applicable to estimate metabolic parameters with standard OGTTs
9. Pancreatic b-cells
(Insulin Secretion)
s
(b-cell
function)
Liver
(Hepatic Glucose
Production)
HGP
Plasma Insulin
I(t)
k
Glucose Space
G(t)
EG0
(G Effectiveness)
OGTT
(75g)
Muscle
(Glucose Uptake)
SII
Novel metabolic parameters During Standard OGTTs
𝒅𝑮
𝒅𝒕
= 𝑶𝑮𝑻𝑻(𝒕) + 𝑯𝑮𝑷(𝑰, 𝑺𝑰) − 𝑬𝑮𝑶 + 𝑺𝑰𝑰 𝑮
𝒅𝑰
𝒅𝒕
=
𝜷
𝑽
𝑰𝑺𝑹(𝝈, 𝑮) − 𝒌𝑰 6,7
mDI = SI*s
• Estimation: SI and s
• Data: G and I, at 5 time points:
t=0, 30, 60, 90, 120 min
Ha J., Satin L., Sherman A., Endocrinology 2016; Ha J. and Sherman A., AJP Endo. Metabolism 2020, Ha et al. Diabetes 2021
10. Mathematical DI with (mDI) and without Insulin (mDI-woI)
mDI with insulin
mDI = SI*s
• Uses glucose and insulin at 5 time points of OGTT
t=0, 30, 60, 90, 120 min
• Estimates SI and s separately
mDI-woI = SI*s
mDI without insulin
• Uses glucose only, t=0, 30, 60, 90, 120 min
• Cannot estimate SI and s separately
• Estimates mDI
ADA 2021, Young Investigator Award
Joon Ha 2021
11. Rationale for no requirement of insulin
Courtesy from Max Springer,
U of Maryland
12. Rationale for no requirement of insulin (cont’d)
Courtesy from Max Springer,
U of Maryland
13. mDI vs. mDI-woI
mDI-woI is excellently correlated with mDI
Data: Dr. Sangsoo Kim, the Pusan National University
Hospital, South Korea
N= 137, mean age=50.5, mean BMI=24.2, Male=48%.
14. Mean Glucose and Glycemic Excursion
• mDI-woI detects mean G and Glycemic excursion
• Large change in mDI-woI, but small change in mean G
16. : ROC Analysis to detect T2D and PreDM
T2D PreDM
Good Detector
Cross-sectional data
17. • N=5742 (Non-diabetes) and OGTTs
every two years
• Outcome: DM over the course of
14 year-longitudinal study
Good Predictor
: Longitudinal Confirmation
Korean Genome Epidemiology Study
(KoGes)
Data: Dr. Sangsoo Kim, PNUH, South Korea
18. Robustness of mDI –woI (longitudinal, N=215)
N=215 who developed from NGT at baseline to PDM and T2D
mDI-woI substantially changed
at onset of Prediabetes
19. Summary, so far
• Model insulin sensitivity, b-cell function, and DI agree well with
clamp parameters in obese youth
• mDI-woI correlates with mDI with insulin
• mDI-woI is a good and robust predictor
21. https://time.com/4703099/continuous-glucose-monitor-blood-sugar-diabetes/
Continuous Glucose Monitoring (CGM)
Real-time Glucose Reading
• Real-time glucose monitoring; daily average glucose and glycemic excursion
• glucose measurements collected in interstitial fluid (ISF G)
• Insulin measurements not collected
• Less invasive and relatively cheap
BG
ISF G
Peaks of ISF G tend to delay and
decrease
Diabetes Care. 2003;26(8):2405-2409. doi:10.2337/diacare.26.8.2405
Can we estimate relative beta-cell function with CGM?
• Current math models cannot do this because they require insulin measurements
• Our model can do (Ha and Sherman AJP 2020, Ha et al. Diabetes 2021)
• Goal: Estimate relative beta-cell function, mDI-CGM during CGM
BG vs. ISF G
22. Step 1: Relative Beta-cell function During an OGTT
wearing a CGM device
𝒅𝑮
𝒅𝒕
= 𝑶𝑮𝑻𝑻(𝒕) + 𝑯𝑮𝑷(𝑰, 𝑺𝑰) − 𝑬𝑮𝑶 + 𝑺𝑰𝑰 𝑮
𝒅𝑰
𝒅𝒕
=
𝜷
𝑽
𝑰𝑺𝑹(𝝈, 𝑮) − 𝒌𝑰 6,7
mDI-ISF = SI*s
• Estimation: SI and s
• Data: ISF at 5 time points:
t=0, 30, 60, 90, 120 min
Ha J., Satin L., Sherman A., Endocrinology 2016; Ha J. and Sherman A., AJP Endo. Metabolism 2020, Ha et al. Diabetes 2021
• Assumption: Discrepancy Between BG and ISF G is
uniform
• A data set for Step 1: BG and ISF G during OGTTs;
Patients wear a CGM sensor during an OGTT;
Glucose load same for all patients, 75 g
23. Mean ISF and Glycemic Excursion vs. mDI_ISF
mDI-ISF detects mean G and excursion G
24. Step 2: Relative Beta-cell function During CGM in a free
living environment
𝒅𝑮
𝒅𝒕
= 𝑶𝑮𝑻𝑻(𝒕) + 𝑯𝑮𝑷(𝑰, 𝑺𝑰) − 𝑬𝑮𝑶 + 𝑺𝑰𝑰 𝑮
𝒅𝑰
𝒅𝒕
=
𝜷
𝑽
𝑰𝑺𝑹(𝝈, 𝑮) − 𝒌𝑰 6,7
Output:
• Estimation: Si ,s, and Glucose Appearance
rate
• mDI-CGM = SI*s
Ha J., Satin L., Sherman A., Endocrinology 2016; Ha J. and Sherman A., AJP Endo. Metabolism 2020, Ha et al. Diabetes 2021
Input:
• CGMs at 25 time points during a meal
t=0, 10, 20, 30, … 240 min
• A total amount of carbo intakes during a meal
Carbohydrate intakes: Can-Pro 5.0 (web ver.), The Korean Nutrient Society
• Assumption: Discrepancy Between BG and ISF G is
uniform
• A CGM data set with free-living environments for Step
2; Glucose load estimated with carbo intakes
26. Summary, so far
• The mathematical model enables to estimate relative beta-
cell function during OGTTs and CGM without insulin
measurements
• mDI-CGM predicts mean G
• mDI-woI and mDI-CGM are cost effective for large scale
epidemiological studies and beneficial for patient care.
28. Personalized intervention with estimated SI and BCF
• Model-derived beta-cell function and insulin sensitivity characterize glucose tolerance status
• Insulin resistance is a manageable risk factor
-Lifestyle intervention could be more effective with insulin resistance subjects than weak beta function
• Insulin resistance and weak beta-cell function groups have the same level of mDI
Weak beta-cell function
Insulin resistance
Which group is
at higher risk?
30. Six subtypes of Non-diabetes:
Machine learning algorithm
N=325, Non-diabetes
Age=48.2,
BMI=24.3, PNUH
Normal Function
Insulin Sensitive
(NF IS)
Weak Function
Mild Insulin Resistance
(WF MIR)
Weak Function
Insulin Resistance
(WF IR)
Normal Function
Insulin Resistance
(MDF IR)
Mild DysFunction
Insulin Resistance
(MDF SIR)
Strong Function
Severe Insulin Resistance
(SF SIR)
A B F
E
D
C
G
I
31. Survival Analysis
Insulin sensitivity vs. beta-cell function
Risk Assessment of six subtypes of Non-diabetes
(KoGes, 16 years follow-up)
• Cluster C is at the highest risk and followed by E, B, D, F, A
• A+B+C: 82%, D:8.5%, E:7.9%, F:1.2%
• Weakest beta-cell function class C is at the highest risk
32. Longitudinal Changes of Clusters
Transition Matrix Table
• AàB and BàC are most frequently observed
• AàBàC is the most common pathway to progression to diabetes
A B
Longitudinal Changes of Clusters
33. Most common pathway to T2D:
Koreans vs. Pima Indians
Koreans Pima Indians
Beta-cell function increases and decrease,
as insulin resistance worsens
Population Fit
42. Progressors vs. Non-progressors
AUC G vs. mDI-woI
Progressors have smaller mDI at baseline
Non-progressor
baseline
Progressor
baseline
0 2 4 6
mDI-woI
43. Progressors vs. Non-progressors
2h-PG, 1h-PG, FPG vs. mDI-woI, mDI
Fasting Glucose of the two groups at baseline are not different, suggesting that
Fasting Glucose is not a good predictor to progression to diabetes in Koreans
Progressors have smaller mDI at baseline
mDI
44. Summary, overall
• Model insulin sensitivity, b-cell function, and DI agree well with
conventional surrogates, but outperform to predict dysglycemia
• Model-derived insulin sensitivity and beta-cell function could have
potential to be implemented in personalized therapies.
• mDI-woI and mDI best predict diabetes, compared to current diabetes
criteria, G0, G120, A1c, and oDI, based on a longitudinal study
• A combination of machine learning algorithm and model parameters
identifies 6 subtypes of Non-diabetes and reveals the most common
pathway of progression to diabetes in a cohort of Korean population;
potential to apply for other cohorts
45. Conclusions
• A longitudinal Confirmation of mDI-CGM with
Young African Americans, Howard University
Future Directions
•mDI-woI is a reliable indicator of dysglycemia
(PreDM and DM)
•mDI-woI is suitable for large scale observational
and interventional studies to assess b-cell function
• s and SI are good surrogates for beta-cell function
and insulin sensitivity
46. Minimum Data points for math model derived metabolic surrogates
Outcomes
5G + 5I
5G + 2I or 3G + 3I
5G
3G (t=0, 60, 120)
CGM + Carbo
FPG + FPI + A1c
SI, beta-cell function, DI
SI, beta-cell function, DI
DI
DI
DI (ongoing)
Six subtypes with a machine
learning algorithm (ongoing)
Data
47. Funding Source
• dkNET, Pilot Study of Bioinformatics for a new PI, NIDDK, NIH
with title “Estimating Relative Beta-Cell Function During Cont
inuous Glucose Monitoring”
• Brain Pool Program of South Korea, Department of Endocrin
ology and Metabolism Pusan National University Hospital, Pu
san: May 2022 – Dec 2024 titled “Finding a Robust Early Bio
marker of Progression to Type 2 Diabetes Mellitus Using a No
vel Mathematical Model”
• Howard University Start-up Fund 49
48. Sangsoo Kim, MD, Division of Endocrinology,
Pusan National University Hospital, South Korea
Collaborators and data sources
Jinmi Kim, Ph.D, Department of Biostatistics,
Clinical Trial Center, Biomedical Research Institute,
Pusan National University Hospital, South Korea
Team of Division of Endo. and Meta., PNUH
Wook Yi, MD
Sori Yang, MD
Myoungsoo Lim, MD
Doohwa Kim, MD
Minsoo Kim, Ph.D. Candidate
Hyejung Jae, RN
49. 15-year longitudinal data of Pima Indians
51
Pima Indian Data:
Clifton Bogardus,
Phoenix, NIDDK, NIH,
A community
with very high
rates of obesity
and diabetes
50. Consultants and data sources
Stephanie Chung
Anne Sumner
Diabetes, Endocrinology and Obesity Branch, NIDDK, NIH.
Arthur Sherman, Ph. D.
NIDDK, NIH, MD