ORIGINAL RESEARCH

published: 13 March 2020

doi: 10.3389/fbioe.2020.00195

A Novel Synthetic Model of the

Glucose-Insulin System for
Patient-Wise Inference of
Physiological Parameters From
Small-Size OGTT Data
Sebastián Contreras 1 , David Medina-Ortiz 1,2 , Carlos Conca 1 and Álvaro Olivera-Nappa 1,2*
1
Centre for Biotechnology and Bioengineering (CeBiB), University of Chile, Santiago, Chile, 2 Department of Chemical
Engineering, Biotechnology and Materials, Faculty of Physical and Mathematical Sciences, University of Chile, Santiago, Chile

Existing mathematical models for the glucose-insulin (G-I) dynamics often involve
variables that are not susceptible to direct measurement. Standard clinical tests for
Edited by:
measuring G-I levels for diagnosing potential diseases are simple and relatively cheap,
Julio R. Banga,
Spanish National Research Council, but seldom give enough information to allow the identification of model parameters
Spain within the range in which they have a biological meaning, thus generating a gap
Reviewed by: between mathematical modeling and any possible physiological explanation or clinical
Pavel Loskot,
Swansea University, United Kingdom interpretation. In the present work, we present a synthetic mathematical model to
Pasquale Palumbo, represent the G-I dynamics in an Oral Glucose Tolerance Test (OGTT), which involves
University of Milano-Bicocca, Italy
for the first time for OGTT-related models, Delay Differential Equations. Our model
*Correspondence:
can represent the radically different behaviors observed in a studied cohort of 407
Álvaro Olivera-Nappa
aolivera@ing.uchile.cl normoglycemic patients (the largest analyzed so far in parameter fitting experiments),
all masked under the current threshold-based normality criteria. We also propose a
Specialty section:
novel approach to solve the parameter fitting inverse problem, involving the clustering
This article was submitted to
Bioinformatics and Computational of different G-I profiles, a simulation-based exploration of the feasible set, and the
Biology, construction of an information function which reshapes it, based on the clinical
a section of the journal
Frontiers in Bioengineering and
records, experimental uncertainties, and physiological criteria. This method allowed an
Biotechnology individual-wise recognition of the parameters of our model using small size OGTT data (5
Received: 20 December 2019 measurements) directly, without modifying the routine procedures or requiring particular
Accepted: 27 February 2020
clinical setups. Therefore, our methodology can be easily applied to gain parametric
Published: 13 March 2020
insights to complement the existing tools for the diagnosis of G-I dysregulations. We
Citation:
Contreras S, Medina-Ortiz D, Conca C tested the parameter stability and sensitivity for individual subjects, and an empirical
and Olivera-Nappa Á (2020) A Novel relationship between such indexes and curve shapes was spotted. Since different G-I
Synthetic Model of the Glucose-Insulin
System for Patient-Wise Inference of
profiles, under the light of our model, are related to different physiological mechanisms,
Physiological Parameters From the present method offers a tool for personally-oriented diagnosis and treatment and to
Small-Size OGTT Data.
better define new health criteria.
Front. Bioeng. Biotechnol. 8:195.
doi: 10.3389/fbioe.2020.00195 Keywords: mathematical modeling, glucose-insulin control, OGTT, inverse problems, personalized medicine

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 1 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

INTRODUCTION relevant parameters and perform large scale studies. Caumo et al.
(2000) and Dalla Man et al. (2002, 2004, 2006) focused their
Disequilibriums in the glucose-insulin (G-I) dynamics, as in efforts on modeling glucose absorption from the digestive system
diabetes, insulin resistance, glucose intolerance, among others, into the bloodstream, making differences between OGTT and
are a widespread condition in modern society (Cobelli et al., 2009; MT. However, parametric identifiability in their models required
Ajmera et al., 2013; Hu et al., 2015; Toniolo et al., 2018). For this the a priori knowledge of average values of some parameters
reason, the mathematical modeling of the G-I control system has of the sample, assumed equal and constant for all individuals.
been frequently visited, as shown by the wide variety of models The number of patients (88) in Dalla Man et al. (2004) allowed
presented in numerous reviews published to date (Bergman, to determine the non-Gaussian distributions of some of the
2005; Boutayeb and Chetouani, 2006; Makroglou et al., 2006; parameters. Later, Dalla Man et al. (2007) presented a nested
Palumbo et al., 2013; Cobelli et al., 2014). sub-systems model, fitting its parameters to a population of 204
For modeling purposes, we may understand the G-I dynamics clinically healthy individuals that underwent a MT. Reflecting on
as follows. The digestion of macronutrients generates glucose its complexity, the authors suggested to use this model only as a
(among others nutrients), which enterocytes absorb into the simulator. Following the trend of previous minimal models, this
bloodstream in the upper intestinal tract. The blood glucose model also leaves out equations for other regulatory hormones
concentration increase caused by glucose absorption induces such as glucagon, epinephrine, growth hormone, and incretins,
pancreatic β-cells to secrete insulin in different timescales: early which regain importance in other works (Brubaker et al., 2007;
insulin release, signaled by the incretin hormones (secreted from Silber et al., 2010; Mari et al., 2013).
intestinal enterocytes), is achieved by emptying the contents Salinari et al. (2011) presented a model in which the
of the vacuoles in the β-cells. After that source is depleted, intestinal absorption of glucose is obtained as a solution
insulin production follows a saturable dynamics. Insulin signals of a transport partial differential equation, where glucose is
the uptake of glucose from peripheral tissues (mainly muscle progressively absorbed while passing through the intestine, and
and adipose tissue), which metabolize it to obtain energy or to stomach emptying is assumed to be exponential. Subsequently,
synthesize storage macromolecules, and the decrease of glucose De Gaetano et al. (2013) presented an extension of the classic
release from the liver. If blood glucose levels are too low, minimal model, modeling the gastrointestinal tract as four
pancreatic glucagon-induced hepatic glucose release restores the compartments, coupled with first-order kinetics. Besides, the
steady-state homeostatic level. authors proposed fixed forms for hepatic glucose production
Some typical routine tests employed for the diagnosis of G- and incretin action, without clear physiological justification or
I-related dysregulations are the Oral Glucose Tolerance Test supporting background, and in disagreement with the state
(OGTT) and the Meal Test (MT), which are widely used of the art (Silber et al., 2010). The patient sample analyzed
given their clinical simplicity and low cost. In OGTT, a in De Gaetano et al. (2013) comprised 78 patients with
fasting patient ingests a 75 g-controlled dose of liquid glucose different clinical classifications, and parameters were fitted to
(Bartoli et al., 2011), while in MT a controlled meal with whole groups of patients according to each clinical criterion,
known glycemic index (such as rice) is given to a patient. In reporting statistical differences for insulin sensitivity between
both tests, glycemia is measured at different times. Typically, different groups.
such measurements take place at time 0 (fasting) and 2 h The importance of the development of mathematical models
after the ingestion of glucose, but more temporal resolution for the study of the G-I control system lies in the potential to
might be required, or other physiological variables measured, serve as support for clinical diagnostic tools in the detection of
depending on how strict clinical criteria are. Moreover, efforts type II diabetes, insulin resistance, glucose intolerance, among
have been made to modify and standardize the temporal other dysregulations of the G-I control system. However, at
resolution and duration of the tests (Bergman et al., 2018). present, these models do not manage to effectively capture the
Besides the clinical interpretation of the OGTT and MT values, differences that exist between patients in the way of achieving
some model-derived indexes can be obtained, as is the case of glycemic control (reflected as morphological variations in OGTT
the Insulin Sensitivity SI , derived from the very well known response curves), and do not represent it as a difference in
Minimal Model (Bergman et al., 1979). The insulin sensitivity the involved parameters that can be interpreted according
SI quantifies the ability of insulin to increase the effect of to clinical criteria. Additionally, the more complex models
glucose on its own disappearance in a steady state. From mentioned above have not been used to infer physiological
Bergman’s Minimal Model, mathematical models of the G-I parameters for an individual patient from an OGTT, since
dynamics have evolved, increasing their complexity and aiming they have only been used to calculate average parameters in
to different objectives. a group of patients, or required special clinical setups to
Stumvoll et al. (2000) presented an empirical approach based obtain the data needed for fitting them. This is partly due
on correlations to determine SI from OGTT curves. Mari et al. to the intrinsic complexity of the models and the number
(2001) proposed a parametric approach to obtain this index, of parameters involved, but also to the lack of a numerical
studying a population of 104 individuals with different clinical procedure for parameter fitting. Therefore, a mathematical model
classification. High correlations between results for SI calculated capable of accounting for different physiological states and
from the model vs. direct measurement for the full patient sample applicable as a tool for clinical diagnosis becomes necessary for
suggested the applicability of the OGTT to obtain clinically personalized medicine.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 2 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

In the present work, we present a synthetic mathematical masking pre-disease conditions under the concept of normality.
model to represent the G-I dynamics in OGTT using Delay Therefore, as currently defined, to identify the physiological
Differential Equations (DDE), in which each parameter describes background behind a clinically healthy or unhealthy individual
a single physiological phenomenon. To the best knowledge of seems to be an ill-posed inverse problem.
the authors, this is the first DDE model involved in describing The implementation of the mathematical model and
OGTT dynamics, including the mutual interrelation between resolution of the parameter-fitting inverse problem was
glucose and insulin. The structure of the model allowed for performed in Matlab version R2017a, using the Global
representation of every observed qualitative dynamic behavior Optimization Toolbox. All scripts and routines for parameter
in our cohort, regardless of the number, height or location of fitting were run on the Chilean National Laboratory for High-
the glucose and insulin peaks. Using a novel information-based Performance Computation (NLHPC) servers, using BASH-based
approach we achieved an individual-wise parameter fitting using control scripts.
the five glycemia and insulinemia points of a routine OGTT
directly, for a cohort of 407 patients that underwent a 75-g
OGTT, which is the largest cohort analyzed so far for this RESULTS
end. We also show that there are different controlling behaviors Synthetic OGTT Glycemia-Insulinemia
within the clinical normality thresholds, accounting for different
physiological mechanisms to achieve glycemic control. Given
Model
The synthetic model proposed in this work considers five main
that parameter fitting is individual, it would be possible to classify
variables. Four of them represent the amount or concentration of
each patient within different groups, suggesting that different
glucose in different compartments: in the stomach S, in the upper
dysregulated mechanisms require different corrections, thus
intestinal tract J (jejunum) and L (ileum), and in the bloodstream
transforming the proposed model and fitting procedure into a
(glycemia) G. The last variable accounts for the insulinemia I.
tool for preventive clinical diagnosis and personalized medicine.
These different variables interact as illustrated in the box diagram
of Figure 2, in ways that will be further detailed in this section.
We follow the notation of De Gaetano et al. (2013), but our model
MATERIALS AND METHODS also considers the contributions of other works (Dalla Man et al.,
2007; Salinari et al., 2011; De Gaetano et al., 2013), together with
A cohort of 407 volunteers was used for testing the capabilities
our own developments.
of our model and our patient-wise parameter recognition
Figure 2 shows the interdependence of the different variables
methodology. All volunteers gave their informed consent to use
of the model. The gastrointestinal sections are decoupled from
their OGTT data in this study. All volunteers in this sample
the blood G-I dynamics, which, as we will see, are connected in a
were clinically healthy according to criteria used in Chile, in
very non-linear fashion. To obtain the G-I profiles of a particular
strict clinical settings to characterize normoglycemic patients.
individual, given all the parameters involved, it is necessary to
According to these criteria, a patient who has basal (stable
solve the system of differential equations that we propose to
overnight) glycemia lower than 100 mg/dL, basal insulinemia
model it. All the proposed constitutive equations may be found
lower than 15 µU/mL, glycemia values not exceeding 160
in this section, together with their physiological background and
mg/dL at any time and not persisting at values higher than
the reasons behind its mathematical formulation. A summary of
140 mg/dL over 2 h, and insulinemia not persisting at values
the model is presented in the last subsection and a detailed list of
higher than 60 µU/mL for a continuous time period of 120
its parameters and variables can be found in Tables S2, S3.
min, would be classified as normal. Approximately 80% of
the cohort corresponded to female patients, with ages ranging
Gastrointestinal System
between 18 and 65 years-old. Nevertheless, no statistical sex-
To mathematically represent the stomach emptying for a liquid
related difference was found within the cohort. Every patient
underwent an OGTT with five measurements for both glycemia bolus, we assume its rate vs = dS dt to be directly proportional to
and insulinemia: fasting (basal state) and every 30 min, for the content of glucose at every time, S(t), as Equation (1) shows:
2 h. Further description of the data, as statistical properties,
histograms of the measurements at every time and of the age dS
= −kjs S, S(0) = D, (1)
distribution are presented in Supplementary Material (Table S1 dt
and Figures S1, S2, respectively).
We identified many radically different G-I profiles among where kjs is a first-order kinetic constant, D the ingested glucose
the population. Examples of them are hypoglycemic individuals, bolus in an OGTT (75 g), and the minus sign accounts for the
single/double peak patients, and those with practically invariant disappearance of glucose, hence the emptying of the stomach.
G-I profiles, as shown in Figure 1. In Figure 1, the solid The glucose that leaves the stomach appears in the jejunum J,
lines correspond to a spline interpolation of the experimental being a source term in the rate (Equation 2).
measurements, which are marked with solid diamonds of the As absorption of glucose may take place in this section of
same color. These different OGTT profiles could account for the small intestine, we may close the mass balance noting that
diverse physiological states related to gastric emptying, intestinal the glucose that is not absorbed will continue its way, being
absorption, or other components of the glycemic control system, transported to the ileum L.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 3 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

FIGURE 1 | Some of the different G-I profiles encompassed under the clinical normality criterion. Continuous curves are spline interpolations of the diamond marked
experimental trends. Note that similar glycemia curves are not necessarily associated with similar insulinemia curves, and a single-peak profile in glycemia is not
necessarily associated with a single-peak insulinemia profile. Upper clinical normality criteria are represented as red limits for the basal and last points and for the entire
OGTT time span.

FIGURE 2 | Box diagram of the proposed model and the interactions between the different compartments and variables. The gastrointestinal tract, namely variables
S, J, and L, are decoupled from the blood G-I dynamics.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 4 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

dJ
= kjs S − kgj J − kjl J , J(0) = 0 (2)
dt |{z} |{z} |{z}
Stomach delivery Absorption into the bloodstream Delivery to L

The structure of our model, from this point and further, is where kxg is the insulin-independent glucose uptake rate, kxgi is
significantly different from the one presented in De Gaetano the uptake rate of insulin-sensitive tissues, η is the bioavailability
et al. (2013). De Gaetano et al. (2013) suggested that to of the intestinal absorbed glucose, and Gprod is the rate of hepatic
represent the effect of intestinal transit between jejunum glucose production.
and ileum, a virtual compartment R can be added in
between, where absorption does not occur. Nevertheless, Hepatic Glucose Production Function
the equations proposed to reach such objective are not We implicitly incorporated the effect of glucagon into a
entirely appropriate for representing different curve shapes, hepatic glucose production function, which is an always
especially those curves with delayed 2-peak dynamics with positive term that contributes to the equation of blood
equal peak height and width or a higher second peak glucose dynamics. Previous works model this contribution with
(see Figure 1). exponential functions (De Gaetano et al., 2013; Erlandsen
Here we decided to follow the formalism presented by et al., 2018), without a solid physiological background but
Salinari et al. (2011) to represent intestinal transit. Taking rather a mathematical convenience. However, when analyzing
into account the peristalsis-driven intestinal transit mechanism, the functional nature of such expressions, we realize that the
we can assume there is no mixing in the axial axis. differential mechanism is not entirely clear since the function
Hence, we model the flow through the intestine as a derivative cannot be written as a function of itself. Therefore, no
plug-flow reactor with uniform velocity U. To represent reliable physiological mechanism supports the form of such rate
the distribution of glucose transporters along the small functions. We propose Gprod as the complement of a Monod-like
intestine, we assume that absorption occurs in two separated equation, which is typically used to model problems of saturable
areas of the intestine, the jejunum and the ileum, located growth, production, enzymatic reaction and receptor/ligand
at a distance l, so the time it takes for the ileum to interaction or transport:
receive glucose transiting from the jejunum is τ := l/U.
Consequently, instead of having a spatial partial differential kλ
Gprod = , (5)
equation for intestinal absorption, we have two ordinary k2 + G
differential equations, one for the jejunum (Equation 2) and one
As defined above, Gprod is also the solution of the mechanism
for the ileum, the last one having a τ −delayed forcing function
given by Equation (6), which represents a logistic hyperbolic
(Equation 3).

dL 0, if t < τ
= kjl ϕ(t) − kgl L(t) , ϕ(t) = (3)
dt | {z } | {z } J(t − τ ), if t ≥ τ
Delayed contribution from J Absorption into the bloodstream

where kjl and kgl are first-order kinetic constants, respectively, production with an asymptotic maximum rate (see derivation in
accounting for the rate of jejunal glucose delivery and the glucose Supplementary Material).
absorption into the bloodstream.
 
Blood Glucose Dynamics dGprod Gprod k2
=− 1− Gprod , (6)
To represent variations in glycemia, our model contemplates dG G kλ
the following control mechanisms. As source terms in the
By imposing steady-state conditions Gprod (Gb ) = G0prod ,
glycemia equation, we considered the intestinal absorption
of glucose adjusted by glucose bioavailability (η), following k2 can be written as a function of the other variables,
the form presented in Dalla Man et al. (2002), and the resulting in Equation (7), i.e., a saturable Michaelis-
hepatic contribution to glucose homeostasis (Gprod ), which Menten-like kinetics much more representative of the
indirectly accounts for the action of glucagon. The sink physiological background of cellular processes. Noticeably,
terms in the equation represent glucose uptake by insulin- even though the rate Equation (6) does not include
insensitive tissues and renal excretion, which is proportional any set point, the integrated steady-state hepatic glucose
to G, and insulin-driven consumption of glucose, taking production (Equation 7) explicitly depends on the difference
place in insulin-sensitive tissues, which is proportional to GI. between glycemia and its base level (see derivation in
Equation (4) represents mathematically the glycemia dynamics. Supplementary Material).
dG

= −kxg G − kxgi GI + Gprod + η kgj J + kgl L , G(0) = Gb (4)
dt | {z } | {z } | {z } | {z }
Basal uptake Insulin-sensitive uptake Hepatic delivery Intestinal absorption

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 5 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

kλ bloodstream. Application of these models to experimental data

Gprod = . (7)
kλ demonstrated a minimal delay, in the range of a few minutes,
+ (G − Gb )
G0prod between concentrations in the bloodstream and the intermediate
active compartment. Taking into account that usual OGTT
Blood-Insulin Dynamics experiments take measures every 30 min and the registered G-
For the blood-insulin system, following the model in De Gaetano I dynamics occur in the order of hours, this small delay was
et al. (2013), we propose a Hill’s dynamics for pancreatic secretion not included in the formulation of our model, because the time
(Goutelle et al., 2008), and an I-proportional degradation term. resolution might result too coarse to accurately calibrate such
Noteworthy, these dynamics exhibit a saturation behavior since it parameters. In this way, we only consider a direct action of blood
is formulated exclusively for OGTT circumstances. However, we insulin on target tissues. From a mathematical and practical point
corrected the mathematical form of the incretin action suggested of view, this decision also resulted in a more compact model with
in De Gaetano et al. (2013). De Gaetano et al. (2013) assumed fewer parameters to fit experimental data.
that incretin secretion is proportional to glucose levels within Insulin sensitivity SI , formally introduced by Bergman et al.
the intestinal lumen. Nonetheless, it has been shown that it (1979) and mathematically defined by Equation (10), accounts
rather depends on the rate of absorption of glucose from the for the quantitative influence of insulin to increase the effect of
intestinal lumen (Silber et al., 2010). Therefore, we corrected this glucose on its own disappearance, in steady state.
in Equation (8):

∂E ∂ dG/dt
SI = , E=− (10)
∂I ∂G
G̃ = G + fgj (kgj J + kgl L). (8)
In our mathematical model, we adopted the term kxgi GI and
Equation (8) makes sense from a physiological point of view since kxg G of De Gaetano et al. (2013) to represent the glucose-
intestinal epithelial cells are not able to sense the absolute amount mediated effect of insulin on glucose disappearance from the
of glucose in the intestine due to the lack of glucose sensor bloodstream and glucose uptake by insulin-independent tissues,
proteins, but their internal metabolic rates are directly dependent respectively. However, contrary to the equations of De Gaetano
on the steady-state cytoplasmic concentration of glucose, which et al. (2013), in our model parameter kxgi is a true insulin
is proportional to glucose membrane transport through the cell. sensitivity value. This was achieved by the redefinition of Gprod ,
In this equation, fgj is a conversion factor that indirectly links resulting in such a way that no additional terms appeared in the
glucose absorption rate to insulin secretion rate through incretin mathematical calculation of the insulin sensitivity according to
action, thus representing the relative power of incretin action vs. Equation (10).
direct glycemic action on the pancreas. Under these assumptions,
Equation (9) gives the final expression for the insulin dynamics: Summary of the Model
Collecting the different expressions derived in the previous
  sections for the constitutive compartments of our model, we can
dI  βγ + 1 I summarize it in the following system of differential equations:
= kxi Ib   γ −  (9)
dt β γ Gb
+1 Ib
G̃ dS
= −kjs S, S(0) = D
dt
where kxi is a first-order kinetic constant for the insulin dJ
= kjs S − kgj J − kjl J, J(0) = 0
degradation in target tissues, β and γ are parameters for half dt
saturation and acceleration of the insulin production, which

dL 0, if t < τ
account for first and second phase pancreatic secretion, G̃ the = kjl ϕ(t) − kgl L(t), ϕ(t) =
dt J(t − τ ), if t ≥ τ
apparent G, enhanced by incretin action, and Gb , Ib the steady- dG

state value of such variables. = −(kxg + kxgi I)G + Gprod + η kgj J + kgl L , G(0) = Gb
dt
 
Insulin Action on Glycemia and Insulin Sensitivity dI  βγ + 1 I
Former mathematical models including insulin action were = kxi Ib   γ − 
dt β γ Gb
+1 Ib
divided in models that considered a direct action of blood G̃
insulin on tissues to regulate glucose uptake and more complex
models that considered an additional intermediate compartment.
Such compartment represented the interstitial fluid in peripheral
tissues, into which insulin was absorbed from the bloodstream Parameter Fitting Strategy
following a first-order kinetics, and only then could exert its During a 5-point OGTT, five experimental measures of both
action. Mathematically, the effect of this formulation in the glucose and insulin are captured, generating vectors Gexp and
more complex models was the introduction of a small delay Iexp . By protocol, measurements are taken at 0, 30, 60, 90, and
and a proportionality constant between concentrations in the 120 min after glucose ingestion, giving a time vector Texp =
bloodstream and the interstitial fluid, which caused a small [0 30 60 90 120 ]. The model detailed in the previous sections
decrease in peak height and slight broadening of peak width was used to represent and interpolate these points continuously.
for insulin in the intermediate compartment compared to the Let Gnum (θE), Inum (θE) be its solution for the G-I dynamics, with

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 6 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

parameters θE. The traditional way of formulating the parametric Gnum , following the structure of Equation (11). It is important
fitting problem is by minimization of a cost function Jexp to point out that this contribution has no greater effect than
that accounts for the difference between the modeled curve favoring those solutions that are smooth and regular. Introducing
and experimental measurements. Generally, this function is this component into the curve fitting procedure adds information
proportional to the mean squared error MSE, because the optimizer would not only look for those solutions
whose numerical profiles match the experimental data points, but
5 for those whose profiles also do not drift considerably from the
1 X
2
Jexp (θE, α) = 2
Gexp (i) − Gnum (Ti ) expected trend.
5 (Gmax − Gmin ) i=1
(11)
5 Simulation-Based Regularization for Clusters of
α X
2
+ 2
Iexp (i) − Inum (Ti ) , Similar Curves
5 (Imax − Imin ) i=1 Taking into account the physiological and molecular mechanisms
involved in the G-I dynamics and the experimental values
where α is a constant that connects the contribution of the obtained in typical OGTT measurements, we determined
insulin curve to MSE and (Gmax − Gmin )2 is a scaling factor. plausible lower and upper bounds (θi,min and θi,max , respectively)
This problem consists of finding the values of all 13 parameters for each parameter θi . In a first stage, we may define the feasible
of the model from 10 experimental measures obtained during set F0 considering such thresholds,
a routine 5-point OGTT for a given patient, which is a
slightly underdetermined problem. However, we can exploit the
13
knowledge we have about the nature of the physiological G-I O  
F0 = θi,min , θi,max , (13)
response, accumulated in more than 40 years of routine testing
i=1
and modeling, to gain in robustness and identifiability of the
parameter set for each patient. We identified and used the where the operator ⊗ represents the Cartesian product between
following strategies for improvement: the intervals defined by the lower and upper thresholds of each
• Increasing data density through the use of interpolators, to parameter. We performed a simulation stage to explore the
favor smoothness and regularity of the solutions, and to nature of F0 , in which we simulated 108 values of θE∗ ∈ F0 ,
penalize nonphysiological oscillations. and studied the Gnum (θE∗ ) and Inum (θE∗ ) profiles obtained. If such
• Simulation-based regularization for clusters of similar curves. profiles fulfilled the clinical normality criteria, they were assigned
• Nested sub-problems and sequential approximations to build to groups of experimental profiles based on their similarities,
a robust initial guess. aiming to build a set of initial guesses for the parameters
• Incorporation of information in the cost function and the corresponding to such individuals. Once we had enough θE∗
delimitation of the feasible set. for each group of curves, a fourth component for the global
• Algorithm choice for the parameter recognition problem and cost function (Equation 16) was added, accounting for the
final shaping of the feasible set. contribution of local regularization near θEj∗ characteristic of the
j’th group.
Increasing Data Density
Given the nature of the equations presented in our model, non- Nested Sub-problems and Sequential
physiological high-frequency oscillatory solutions might appear. Approximations
Taking into account (a) the nature of the physiological G- We followed a staged approach for the construction of
I control system, (b) the oscillations measured experimentally initial guesses for the inverse problem, summarized in the
in the literature, and (c) the 30-min apart measurements following algorithm:
taken during a 5-point OGTT, we know that high-frequency Steps one and two aim to build an appropriate initial guess for
oscillations –relative to the sampling time– should not be the parameter fitting problem, solved in step three. This step does
observable in our solution. Therefore, we propose to favor those not include further assumptions on the nature of the solutions,
solutions that only have low-frequency oscillations, somehow but only the model equations.
forcing the it to resemble the experimental data in a smooth way.
For this, we propose to increase the density of putative measured Incorporation of Information in the Cost Function
points using a soft interpolant to connect the experimental Given that the a priori identifiability of the model is not
measurement points. Without loss of generality, for G, the guaranteed (De Gaetano et al., 2013), we corrected the cost
interpolant Ĝ used to increase the data density is defined by function presented in the previous section to incorporate more
Equation (12): information. Without making any assumption concerning the
parameters, we incorporated information from clinical records
Ĝ = φGspline + (1 − φ)Gpol , (12) and from the test itself. In particular. we added terms that account
for: (i) experimental errors, associated with sampling time and
that is a convex combination between a cubic spline and a low- laboratory techniques, and (ii) expected extreme values (maxima
degree polynomial interpolator. We can define a new component and minima), inferred from the experimental measurements and
of the error function, Jspline , based on the MSE between Ĝ and based on clinical criteria.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 7 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

Algorithm 1: Sequential approximation to the initial guess Algorithm 2: Incorporation of experimental errors in the
Data: cost function
θEj∗ : Initial guess, built from the simulation-based Data:
parametric pool of group j; xexp : experimental glycemia or insulinemia value at time
Gnum , Inum : Model derived G-I profiles, with temporal texp ;
resolution tnum ; texp : sampling time, element from Texp ;
λ : penalizing factor, λ > 1;
Ĝ, Î : Interpolation of the experimental G-I profiles at tnum ;
Result: Jerror
Result: Initial guess θE0 for the parametric recognition Set Jerror = 0;
inverse problem of a particular patient foreach texp ∈ Texp , x = {G, I} do
foreach individual in the cohort do Calculate E, ellipse centered in (texp , xexp (texp )), with
Find the most similar simulated curve, and set θE0 = θEj∗ . semi-axes a = 21t , b = 21x .;
Fit the glycemia-related parameters, using I = Î Calculate c = xnum ([texp − 1t , texp + 1t ]), image of the
(Equation 12) as a known forcing function on major semi-axe through function xnum .;
Equation (4), minimizing Jexp (θE, 0).; Calculate d, the distance between E and c.;
Update the glycemia-related components of θE0 . if d = 0 then
Fit the insulinemia-related parameters, using G = Ĝ Calculate d′ , the distance between c and the center of
(Equation 12) as a known forcing function on E;
Equation (4), minimizing Jexp (θE, α), considering only Find θ ′ , angle of the point of minimum distance;
insulinemia (α → ∞).; Calculate r(θ ′ ), ellipse radius at θ = θ ′ ;
 ′ 
Update the insulinemia-related components of θE0 . Update Jerror = Jerror + λ r(θd ′ ) ;
Set θE0 as initial guess for the parameter fitting inverse else
problem. Update Jerror = Jerror + λ(d + 1);
end end
end

Regarding the first term, associated with experimental

errors in both time and measurement, such 2− D variability measurements so that we may have more equations of the form
transformed each data point into an ellipse in the (t, G) or (t, I) h′ (tj∗ ) = 0. For our case, as the OGTT data consists of five
space, centered on the experimentally determined value Gexp glucose and five insulin measurements, the maximum number of
or Iexp , at time Texp . We used a constant 1t = ±3 min as a additional equations we can have is 6.
scale for the temporal uncertainty (horizontal semi-axis of the The classical way of solving the direct problem of finding
ellipse), meanwhile a proportional contribution (to the Gexp or maximum/minimum values of a function h(t) is finding a time
Iexp values) was chosen for the vertical axis. The weight of this t ∗ in which h′ (t ∗ ) = 0 and a sign change occurs in the same
error source in the total cost function was calculated based on point. Then, hopt = h(t ∗ ). This same reasoning can be applied
a polar probability density of ρ(r, θ ), as described in algorithm 2.
to glycemia, using Equation (4) and imposing dGdt = 0 to obtain
Note that when considering together Jexp and Jerror , the algorithm
Equation (15):
can be significantly simplified, as the calculations for the case d =
0 may be skipped only giving a greater weight to the contribution


of Jexp . G∗prod + η kgj J ∗ + kgl L∗
opt
Regarding the second term, its derivation was—in G = , (15)
a mathematical sense—more complicated. Given the kxg + kxgi I ∗
characteristics of the 5-point OGTT considered on this
where Gopt is the maximum glycemia, and I ∗ , G∗prod , J ∗ , L∗
work, it seems reasonable to expect that the maximum value of
the modeled glycemia curve should be of the same order of the are, respectively, the insulin concentration, hepatic glucose
maximum of the experimental profile. Consequently, we will production rate, and the amount of glucose in the jejunum and
have an idea of the time t ∗ ∼ Texp (i) when our model reaches ileum at time t ∗ (when Gopt is reached). Note that the formulation
such value. Furthermore, considering a sequence of experimental is general, and it serves for any point of derivative equal to
measurements Hi recorded at times Ti , which approximate a C1 zero (which we can estimate from the clinical exam trends).
function h(t). If for certain i0 it is fulfilled that We can indeed follow and apply the same idea for the insulin
equation (Equation 9). If we do this for all critical points we
will have n linearly independent equations, since the times t ∗



Hi0 − Hi0 −1 Hi0 +1 − Hi0 < 0, (14) and optimal values of G, I will be different. We can think of
  each
Pn equation as the i− th component of a function F(θE) =
there exists a time t ∗ ∈ Ti0 −1 , Ti0 +1 where h′ (t ∗ ) = 0. Note that f E
e
i=1 i i . Note that around θE∗ where F(θE∗ ) = 0, the parameters

this can happen more than once for a sequence of experimental can be expressed as implicit functions of the known variables

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 8 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

FIGURE 3 | Scatter plot of model-predicted glycemia and insulinemia vs. experimental measurements for the entire cohort of 407 patients. The data cloud lies in the
identity zone, without significant deviations from it, accounting for a good fit.

FIGURE 4 | Probability histograms of relative errors of the model predictions for the cohort of 407 patients and five data points per patient. Given the Gaussian nature
of the residues of both variables, we can safely discard the presence of bias.

(such as Gi,max and Ii,max ). These equations will define different set, E E
Pn tagging as unfeasible the solutions θ such that F(θ ) =
loci for the fitted parameters. The problem that appears directly E
i=1 fi (θ )E
E
ei 6= 0. Thus, we will have a new feasible set for our
after this definition is that known variables, in reality, are only solutions and a new approach to the problem.
approximately known within a confidence interval Ic,i . Then,
when incorporating them, we will have manifolds (of loci). More Settings of the Inverse Problem: Functions,
specifically, we say that θE satisfies the equation i, if there are Algorithms, and Thresholds
G∗i,max , Ii,max
∗ , t ∗ ∈ Ic,i such that fi (θE) = 0E. We then build an Finally, including all the contributions mentioned above, the
information function, which conditions the shape of the feasible parameter-fitting problem can be formulated as:

min E α) + λ2 Jspline (θE, α) + λ3 Jerror (θE, α) +

λ1 Jexp (θ, ǫkθE − θEj∗ k , (16)
θE∈F0 |F(θ)=
E 0E | {z } | {z } | {z } | {z }
| {z } Experimental data Interpolator Experimental errors Local regularization
Information function

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 9 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

FIGURE 5 | Sensitivity and stability analysis of fitted parameters for patient 180. The figure shows the glycemia and insulinemia profiles of the patient in the frame of
patients that have similar glycemic curves (upper and lower left plots, respectively). A volcano plot, as described in the text, is presented in the upper right plot, and a
representation of the width of the 95% confidence interval for each parameter is presented in the lower right plot. In this example, even though some parameters have
a low slope over the studied interval, the variability of their values, expressed as the width of the 95% confidence interval, is almost negligible.

FIGURE 6 | Sensitivity and stability analysis of fitted parameters for patient 144. The figure shows the glycemia and insulinemia profiles of the patient in the frame of
patients that have similar glycemic curves (upper and lower left plots, respectively). A volcano plot, as described in the text, is presented in the upper right plot, and a
representation of the width of the 95% confidence interval for each parameter is presented in the lower right plot. In this example, the variability of the kλ parameter is
considerably higher than in the case of Figure 5, probably because of differences in the experimental profiles.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 10 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

FIGURE 7 | Sensitivity and stability analysis of fitted parameters for patient 31. The figure shows the glycemia and insulinemia profiles of the patient in the frame of
patients that have similar glycemic curves (upper and lower left plots, respectively). A volcano plot, as described in the text, is presented in the upper right plot, and a
representation of the width of the 95% confidence interval for each parameter is presented in the lower right plot. In this example, the variability of some parameter
values is not negligible.

with ǫ arbitrarily small, and which solution is the set of highly correlated to experimental measures, which demonstrates
parameters that characterize the glycemic-insulinemic control for the goodness of fit of our model and method.
each patient. The resolution of the minimization problem 16 at
every stage was achieved by combining deterministic methods
(gradient search) and heuristic methods such as simulated Sensitivity Analysis
annealing and pattern search, available in the Matlab Global Parameter sensitivity was evaluated for each patient by analyzing
Optimization Toolbox. Parameters were obtained for the whole how variations in parameter values affected the error between
studied cohort. predicted and experimental values (the cost function defined
by Equation 16). Some particular examples are shown by the
volcano plots in Figures 5–7. In these figures, the greater the
DISCUSSION slope around the central point, the more sensitive the patient
is to variations of that parameter. Therefore, a more sensitive
Goodness of Fit parameter suggests higher confidence in its fitted value. We
Applying the parameter recognition procedure presented in also performed ten different parameter fitting experiments for
previous sections, it was possible to fit our model to the each patient, starting from different initial values and using
experimental OGTT glycemia and insulinemia profiles of the deterministic and heuristic procedures described above.
the whole studied cohort, thus obtaining the physiological With these results, we calculated a 95% confidence interval
parameters that control the observed trends for each patient. for the parameters of each patient, as shown in Figures 5–7.
To evaluate the performance of our model and the proposed Following our former reasoning, all parameters with high slopes
parameter recognition procedure, we studied the quality of around the central point in the volcano plot have very narrow
their predictions for both glycemia and insulinemia. Figure 3 confidence intervals, but unexpectedly some parameters with low
shows the scatter plots of experimental vs. predicted values sensitivity in volcano plots have also very narrow confidence
for glycemia and insulinemia for all 407 patients, where both intervals in fitting experiments. This observation demonstrates
predicted variables follow the expected trend. The data point that parameter accuracy following our fitting method is higher
cloud lies in the identity zone, without significant deviations, and than expected from analyzing parameter sensitivity around a
the variables show high correlation. The probability-normalized central point.
residues histogram of both variables had a Gaussian nature, as Analyzing differences between patients, our results show that
depicted in Figure 4, with low variance. In sum, prediction errors this sensitivity analysis differs for each case, since different curve
are normally distributed and unbiased, while predictions are shapes suggest different physiological ways to achieve glycemic

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 11 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

FIGURE 8 | Parameter stability in response to small perturbations in the experimental glucose measurements. Random noise of at most 10% of the reported
measurement was added to the original values, and the inverse problem was solved. Results show that, even for the maximum percentage of variation, the difference
between the original and final values remained non significant.

control. Hence, the relative importance of each parameter on experimental measures. Starting from the experimental OGTT
the control mechanism, and the reliability of their values, can points for a fixed individual (Gexp , Iexp ), we simulated a
be considered proportional to their sensitivity or the extension set of virtual patients with OGTT curves (Gsim , Isim ) whose
of their confidence interval. As the most sensitive parameters measurements resulted of adding random noise to (Gexp , Iexp ),
are those that determine the shape of the control profile, and solved the parameter fitting problem, considering as a
understanding their individual and collective meaning may give starting point the parameters fitted to the original patient.
valuable information about the health status of each patient, Different amounts of noise were added to the experimental
highlighting the physiological background of the observed trend. points in all patients. The effects of these variations on all
We present some examples of these observations in parameter values are shown in Figure 8 for one patient example.
Figures 5–7, which feature a special kind of graph that we will All parameters in all generated variations show no significant
address as a volcano plot. In a volcano plot, we analyze the differences in comparison to the original data, showing that the
impact that small variations in the parameters, relative to their model allows margins of error associated to both experimental
optimal value, have on the value of the error functional JT . The error or time lags at the moment of taking samples without
x-axis represents the percent variation of the optimal value of the compromising the accuracy of the solution. This invariability is
parameter, and the y-axis, the variation of the error, normalized related to the inclusion in the fitting procedure of a component
by its minimum value, which is the optimal. Therefore, the point accounting for experimental error, acting as a mathematical
(100, 100) is the center and global minimum of all curves in buffer, which demonstrates the utility of this approach for
the plot. individual parameter fitting.
Figure 5 shows the sensitivity and stability analysis of the
parameters for patient 180. Out of the whole set, the parameter
kλ shows the more significant variability, which is very small and CONCLUSIONS
almost negligible. On the other hand, Figure 6 shows the same
analysis for patient 144, which has a very similar glycemic profile, We presented a new synthetic mathematical model for the G-
but whose kλ is somehow unreliable. Despite sharing almost the I dynamics in an Oral Glucose Tolerance Test. Our model
same glycemic profile, they have essential differences in their only includes parameters with a coherent physiological meaning
insulinemic trends. The above demonstrates the importance of and, to the best knowledge of the authors, is the first approach
analyzing both glycemia and insulinemia for having a reliable which involves Delay Differential Equations (DDE) presented in
indicator of the health status of an individual, given that glycemia the literature. The new model can represent radically different
alone might not be enough. Furthermore, patient 31, who shares G-I behaviors observed in the studied population, including
little or no properties with the glycemic and insulinemic profile hypoglycemiant individuals, single/double peak patients, and
of the other two patients studied above (see Figure 7), has a those with practically invariant G-I profiles. In this way,
different sensitivity footprint, proving the relationship between we demonstrate that these unexpected G-I profiles are not
curve shape and parameter reliability. experimental errors, but are the result of particular combinations
of physiological processes.
We have proposed and numerically implemented a novel
Stability of the Solutions in Response to strategy for the resolution of the curve-fitting problem, exploiting
Experimental Errors existing knowledge about the function of the G-I control
After obtaining final values for every parameter, we evaluated system, leading to the correct recognition of the individual
their stability in response to small perturbations in the fitted parameters for the studied cohort of 407 individuals. Our

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 12 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

methodology showed to be robust, as the dimensionality of DATA AVAILABILITY STATEMENT

the problem can be dramatically decreased by reshaping the
feasible set with the incorporation of an information function The datasets generated for this study are available on request to
and splitting the problem into sequential approaches, thus the corresponding author.
allowing a correct fitting of the model parameters for each
patient. As suggested by the simulations, and afterwards verified ETHICS STATEMENT
by our results, we observed consistency between differences
in glycemic and insulinemic OGTT curves and differences The studies involving human participants were reviewed and
in parameter values. The parameters obtained for each and approved by the Ethics Committee of the San Cristobal Medical
every patient showed to be stable under small perturbations Center, Santiago, Chile. The patients/participants provided their
of the experimental measurements, and their sensitivity varied written informed consent to participate in this study.
from one patient to another, giving physiological and patient-
wise insights of the mechanistic background of the observed AUTHOR CONTRIBUTIONS
trends. This can be asserted particularly by the fact that
each parameter in the proposed model represents a unique SC, DM-O, ÁO-N, and CC: conceptualization, writing, review,
physiological phenomenon. and editing. SC, ÁO-N, and CC: methodology. SC and DM-O:
Finally, since our model can represent the different forms of validation. SC, DM-O, and ÁO-N: investigation. SC and ÁO-N:
control observed so far, to characterize them through parameters supervision. ÁO-N: project administration and funding.
with a physiological meaning, and to identify those parameters
using a robust methodology for the inverse problem solving, FUNDING
we propose it as a tool for patient evaluation, review of health
criteria, and re-definition of clinical normality. Understanding This research has been financed mainly by the Centre for
that under the current clinical normality definition there are Biotechnology and Bioengineering - CeBiB (PIA project FB0001,
different ways to achieve glycemic control, parametric analysis Conicyt, Chile). DM-O gratefully acknowledges Conicyt, Chile,
of patients would allow the development of individual-oriented for Ph.D. fellowship 21181435. Powered@NLHPC: This research
treatments, contributing significantly to the preventive and was partially supported by the supercomputing infrastructure
personalized diagnosis of relevant pathophysiological events in of the National Laboratory for High-Performance Computing,
the control system. In accordance to this, our future work NLHPC (ECM-02), Chile.
will include a thorough analysis of parameter values in normal
and pathological patients under current definitions and their SUPPLEMENTARY MATERIAL
statistical distributions in broader patient samples, in order to
define clinical meaning, normality and pathological criteria for The Supplementary Material for this article can be found
each parameter, which falls out of the scope and length of the online at: https://www.frontiersin.org/articles/10.3389/fbioe.
present work. 2020.00195/full#supplementary-material

REFERENCES Caumo, A., Bergman, R. N., and Cobelli, C. (2000). Insulin sensitivity from meal
tolerance tests in normal subjects: a minimal model index. J. Clin. Endocrinol.
Ajmera, I., Swat, M., Laibe, C., Le Novere, N., and Chelliah, V. (2013). Metab. 85, 4396–4402. doi: 10.1210/jcem.85.11.6982
The impact of mathematical modeling on the understanding of diabetes Cobelli, C., Dalla Man, C., Pedersen, M. G., Bertoldo, A., and Toffolo,
and related complications. CPT Pharmacometr. Syst. Pharmacol. 2, 1–14. G. (2014). Advancing our understanding of the glucose system via
doi: 10.1038/psp.2013.30 modeling: a perspective. IEEE Trans. Biomed. Eng. 61, 1577–1592.
Bartoli, E., Fra, G., and Schianca, G. C. (2011). The oral glucose tolerance test doi: 10.1109/TBME.2014.2310514
(OGTT) revisited. Eur. J. Int. Med. 22, 8–12. doi: 10.1016/j.ejim.2010.07.008 Cobelli, C., Dalla Man, C., Sparacino, G., Magni, L., De Nicolao, G., and Kovatchev,
Bergman, M., Jagannathan, R., Buysschaert, M., Pareek, M., Olsen, M. H., Nilsson, B. P. (2009). Diabetes: models, signals, and control. IEEE Rev. Biomed. Eng.
P. M., et al. (2018). Lessons learned from the 1-hour post-load glucose level 2:54. doi: 10.1109/RBME.2009.2036073
during OGTT: current screening recommendations for dysglycaemia should Dalla Man, C., Camilleri, M., and Cobelli, C. (2006). A system model
be revised. Diabetes Metab. Res. Rev. 34:e2992. doi: 10.1002/dmrr.2992 of oral glucose absorption: validation on gold standard data. IEEE
Bergman, R. N. (2005). Minimal model: perspective from 2005. Hormone Res. Trans. Biomed. Eng. 53, 2472–2478. doi: 10.1109/TBME.2006.8
Paediatr. 64(Suppl. 3):8–15. doi: 10.1159/000089312 83792
Bergman, R. N., Ider, Y. Z., Bowden, C. R., and Cobelli, C. (1979). Quantitative Dalla Man, C., Caumo, A., Basu, R., Rizza, R., Toffolo, G., and Cobelli, C. (2004).
estimation of insulin sensitivity. Amer. J. Physiol. Endocrinol. Metab. 236:E667. Minimal model estimation of glucose absorption and insulin sensitivity from
doi: 10.1152/ajpendo.1979.236.6.E667 oral test: validation with a tracer method. Am. J. Physiol. Endocrinol. Metab.
Boutayeb, A., and Chetouani, A. (2006). A critical review of mathematical 287, E637–E643. doi: 10.1152/ajpendo.00319.2003
models and data used in diabetology. Biomed. Eng. Online 5:43. Dalla Man, C., Caumo, A., and Cobelli, C. (2002). The oral glucose minimal model:
doi: 10.1186/1475-925X-5-43 estimation of insulin sensitivity from a meal test. IEEE Trans. Biomed. Eng. 49,
Brubaker, P. L., Ohayon, E. L., D’Alessandro, L. M., and Norwich, K. H. 419–429. doi: 10.1109/10.995680
(2007). A mathematical model of the oral glucose tolerance test Dalla Man, C., Rizza, R. A., and Cobelli, C. (2007). Meal simulation model
illustrating the effects of the incretins. Ann. Biomed. Eng. 35, 1286–1300. of the glucose-insulin system. IEEE Trans. Biomed. Eng. 54, 1740–1749.
doi: 10.1007/s10439-007-9274-1 doi: 10.1109/TBME.2007.893506

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 13 March 2020 | Volume 8 | Article 195
Contreras et al. Patient-Wise Inference of Glycemic Control Parameters

De Gaetano, A., Panunzi, S., Matone, A., Samson, A., Vrbikova, J., Bendlova, B., Salinari, S., Bertuzzi, A., and Mingrone, G. (2011). Intestinal transit of a glucose
et al. (2013). Routine OGTT: a robust model including incretin effect for precise bolus and incretin kinetics: a mathematical model with application to the
identification of insulin sensitivity and secretion in a single individual. PLoS oral glucose tolerance test. Am. J. Physiol. Endocrinol. Metab. 300, E955–E965.
ONE 8:e70875. doi: 10.1371/journal.pone.0070875 doi: 10.1152/ajpendo.00451.2010
Erlandsen, M., Martinussen, C., and Gravholt, C. H. (2018). Integrated Silber, H. E., Frey, N., and Karlsson, M. O. (2010). An integrated glucose-
model of insulin and glucose kinetics describing both hepatic glucose and insulin model to describe oral glucose tolerance test data in healthy
pancreatic insulin regulation. Comput. Methods Prog. Biomed. 156, 121–131. volunteers. J. Clin. Pharmacol. 50, 246–256. doi: 10.1177/00912700093
doi: 10.1016/j.cmpb.2017.12.009 41185
Goutelle, S., Maurin, M., Rougier, F., Barbaut, X., Bourguignon, L., Ducher, Stumvoll, M., Mitrakou, A., Pimenta, W., Jenssen, T., Yki-Järvinen, H.,
M., et al. (2008). The hill equation: a review of its capabilities in Van Haeften, T., et al. (2000). Use of the oral glucose tolerance test
pharmacological modelling. Fundament. Clin. Pharmacol. 22, 633–648. to assess insulin release and insulin sensitivity. Diab. Care 23, 295–301.
doi: 10.1111/j.1472-8206.2008.00633.x doi: 10.2337/diacare.23.3.295
Hu, F. B., Satija, A., and Manson, J. E. (2015). Curbing the diabetes Toniolo, A., Cassani, G., Puggioni, A., Rossi, A., Colombo, A., Onodera,
pandemic: the need for global policy solutions. JAMA 313, 2319–2320. T., et al. (2018). The diabetes pandemic and associated infections:
doi: 10.1001/jama.2015.5287 suggestions for clinical microbiology. Rev. Med. Microbiol. 30, 1–17.
Makroglou, A., Li, J., and Kuang, Y. (2006). Mathematical models and software doi: 10.1097/MRM.0000000000000155
tools for the glucose-insulin regulatory system and diabetes: an overview. Appl.
Numer. Math. 56, 559–573. doi: 10.1016/j.apnum.2005.04.023 Conflict of Interest: The authors declare that the research was conducted in the
Mari, A., Bagger, J. I., Ferrannini, E., Holst, J. J., Knop, F. K., and Vilsbøll, absence of any commercial or financial relationships that could be construed as a
T. (2013). Mechanisms of the incretin effect in subjects with normal potential conflict of interest.
glucose tolerance and patients with type 2 diabetes. PLoS ONE 8:e73154.
doi: 10.1371/journal.pone.0073154 Copyright © 2020 Contreras, Medina-Ortiz, Conca and Olivera-Nappa. This is an
Mari, A., Pacini, G., Murphy, E., Ludvik, B., and Nolan, J. J. (2001). A model- open-access article distributed under the terms of the Creative Commons Attribution
based method for assessing insulin sensitivity from the oral glucose tolerance License (CC BY). The use, distribution or reproduction in other forums is permitted,
test. Diab. Care 24, 539–548. doi: 10.2337/diacare.24.3.539 provided the original author(s) and the copyright owner(s) are credited and that the
Palumbo, P., Ditlevsen, S., Bertuzzi, A., and De Gaetano, A. (2013). Mathematical original publication in this journal is cited, in accordance with accepted academic
modeling of the glucose–insulin system: a review. Math. Biosci. 244, 69–81. practice. No use, distribution or reproduction is permitted which does not comply
doi: 10.1016/j.mbs.2013.05.006 with these terms.

Frontiers in Bioengineering and Biotechnology | www.frontiersin.org 14 March 2020 | Volume 8 | Article 195