Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 53-2 (2020) 16112–16118
Global
Global Sensitivity Analysis
Sensitivity Analysis onon thethe Bergman
Bergman
Global Sensitivity
MinimalAnalysis
Model on  the
 Bergman
Global Minimal
Sensitivity
Minimal Model
Analysis on the Bergman
Minimal Model
Model 
Minimal
Souransu
Souransu Nandi
Model Singh
Nandi Tarunraj
Tarunraj Singh
Souransu
Souransu Nandi
Souransu Nandi Tarunraj
Nandi Tarunraj Singh
Tarunraj Singh
Singh
Control,
Control, Dynamics
Dynamics Souransu
and
and Nandi
Estimation
Estimation Tarunraj
Laboratory,
Laboratory, Singh
University at
University at Buffalo,
Buffalo,
Control,
Control, Dynamics
Buffalo,
Dynamics
Control,Buffalo,
Dynamics NY and
and
and14221 Estimation
USA
Estimation
Estimation Laboratory,
(e-mail:
Laboratory,
Laboratory, University
tsingh@buffalo.edu).
University
University at
at
at Buffalo,
Buffalo,
Buffalo,
Buffalo, NY
NY 14221
14221 USA
USA (e-mail:
(e-mail: tsingh@buffalo.edu).
tsingh@buffalo.edu).
Buffalo,
Control,Buffalo,
Dynamics NY
NYand 14221
14221 USA
Estimation (e-mail: tsingh@buffalo.edu).
Laboratory,
USA (e-mail: University at Buffalo,
tsingh@buffalo.edu).
Buffalo, NY 14221 USA (e-mail: tsingh@buffalo.edu).
Abstract: In
Abstract: In this
this paper,
paper, aa novel novel Global Global Sensitivity
Sensitivity Analysis
Analysis method method is is developed
developed and and is is
Abstract:
illustrated
Abstract:
Abstract:on In
onIn the
Inthe this
this
this paper,
popular
paper, aa
Bergman
paper,Bergman novel
a novel
novelminimal Global
minimal
Global
Global model Sensitivity
model for Type
Sensitivity
Sensitivity TypeAnalysis
1 Diabetes
Analysis
Analysis method
Diabetes
method
method (T1D).is
is developed
Four parameters
is developed
developed and
parameters
and
and is is
is
illustrated
illustrated on the popular
popular Bergman minimal model for
for Type 1 (T1D). Four
are assumed
illustrated
Abstract:
illustrated
are assumed on
onIntothe
tothebe
this
be uncertain
popular
paper,Bergman
popular
uncertain in novel
ain
Bergman the model
the model
minimal
Global
minimal to mimic
to mimic
model patient
for
for Type
Sensitivity
model Type
patient 111 Diabetes
Analysis Diabetes
variability. method
Diabetes
variability. An
An
(T1D).
algorithm
(T1D).
(T1D). Four
Four
is developed
algorithm Fouris parameters
isparameters
presented
and is
parameters
presented
are
to
are assumed
evaluate
assumed
illustrated
areevaluate on to
assumedsensitivity thebe
sensitivity
to be uncertain
popular
to be uncertain metrics
uncertain in
in
Bergman
in bythe
by
the model
which
model
thewhich minimal
modelthe to
the
to mimic
mimic
model
to uncertain patient
uncertain
patient
for
mimic patient Type variability.
parameters
variability.
1 can
Diabetes
variability. An
be
An
An algorithm
ranked.
algorithm
(T1D).
algorithm Four is
Results
is presented
reveal
presented
parameters
is presented
to
to evaluate sensitivity metrics
metrics by which the uncertain parameters
parameters can
can be
be ranked.
ranked. Results
Results reveal
reveal
that
to
are
to
that for
evaluate
assumed
evaluate
for a single meal
sensitivity
to be
sensitivity
aa single meal scenario,
metrics
uncertain
metrics
scenario, in by
byinsulin
the which
model
which
insulin sensitivity
the
to
the mimic
sensitivity is
uncertain one
patient
uncertain
is one of
of the
parameters most
variability.
parameters
the mostcan
can important
be
An
be
importantranked.
algorithm
ranked. factors
Results
is
Results
factors after
presented
after the
reveal
reveal
the
that
that
to for
consumption
for
evaluate a
that for a single single
singleof meal
meal
meal
sensitivity scenario,
that influences
scenario,
meal thatmetrics
scenario, byinsulin
insulin
which
insulinthe thesensitivity
blood
sensitivity
the is
is
uncertain
sensitivity one
glucose
one of
of the
concentration
the
parameters
is oneconcentration most
mostcan
of the most important important
in people
important
be ranked. factors
with
factors
Results
factors after
T1D.
after
after thethe
reveal
the
consumption
consumption of
of meal
meal that influences
influences the blood
blood glucose
glucose concentration in
in people
people with
with T1D.
T1D.
Copyright
consumption
that for
consumption a © 2020
singleof The
meal Authors.
that
scenario, This
influences is
insulin an theopen access
blood
sensitivity article
glucose
is one under the
concentration
of meal that influences the blood glucose concentration in people with T1D. the
meal of the CC
most BY-NC-ND
in people
important license
with
factors T1D.
after
(http://creativecommons.org/
consumption
Keywords: of meal that
Sensitivity licenses/by-nc-nd/4.0)
influences
Analysis, Polynomial the blood glucose
Chaos, Sobol concentration
Indices, Insulin in people
Insulin Sensitivity with T1D.
Keywords:
Keywords: Sensitivity
Sensitivity Analysis,
Analysis, Polynomial
Polynomial Chaos,
Chaos, Sobol
Sobol Indices,
Indices, Insulin Sensitivity
Sensitivity
Keywords: Sensitivity Analysis, Polynomial Chaos,
Keywords: Sensitivity Analysis, Polynomial Chaos, Sobol Indices, Insulin Sensitivity Sobol Indices, Insulin Sensitivity
Keywords: Sensitivity Analysis, Polynomial Chaos,
1. INTRODUCTION
INTRODUCTION Sobol thereby
tration: Indices, motivating
Insulin Sensitivity researchers to to further
further study study
1.
1. INTRODUCTION tration: thereby motivating researchers
1. INTRODUCTION
1. INTRODUCTION tration:
those
tration:
tration: thereby
parameters
thereby
thereby motivating
and
motivating
motivating their researchers
interactions.
researchers
researchers to
to
to further
It would
further
further study
also
study
study
those
those parameters
parameters and
and their
their interactions.
interactions. It
It would
would also
also
Type
Type 1 Diabetes
11 Diabetes (T1D)
1.
(T1D) is a
INTRODUCTION
is chronic
aa chronic ailment
ailment which
which without
without help
those
tration:
help us figure
parameters
thereby
those usparameters
figure out
out andthe
motivating
andthe non-significant
their
their interactions.
researchers
interactions.
non-significant parameters.
to It would
further
It would This
parameters. This
also
study
also
Type
careful
Type 1 Diabetes
regulation
Diabetes (T1D)
results
(T1D) is
is in
a chronic ailment
micro-vascular
chronic ailment which without
complications
which without help
those us
information
help us figure
figure
parameters
help us figure couldout
out be
and
outbethe the
used
the non-significant
in a couple
non-significant
their of
interactions.
non-significant parameters.
ways. It
parameters.
It would
would
parameters. This
give
This
also
This
Type 1 regulation
careful Diabetes (T1D) results is ina chronic ailment which
micro-vascular without information
complications could used in a couple of ways. It would give
careful
such
careful
Type as1 regulation
retinopathy,
regulation
Diabetes results
(T1D) neuropathy,
results is in
in
a micro-vascular
chronic nephropathy
micro-vascular
ailment complications
which and
complications macro-
without information
scientists
information
help us
information figure could
working
could
couldout be
on
be
be used
the
used in aaa couple
understanding
used in couple
non-significant
in couple of
the
of
of ways.
human
ways. It
It
parameters.
ways. It would
physiology
would
would give
give
This
give
careful
such as regulation
retinopathy, results
neuropathy,in micro-vascular
nephropathy complications
and macro- scientists working on understanding the human physiology
such
vascular
such
carefulas retinopathy,
complications
as regulation
retinopathy, neuropathy,
such
neuropathy,
results as
in as nephropathy
cardiovascular
nephropathy
micro-vascular and
disease macro-
and macro-
complications and the
macro- scientists
the addedworking
scientists
information
scientists information
working
could be
working on
on understanding
on used that
understanding couplethe
mathematically
in amathematically
understanding the human
human
of ways.
the human It physiology
the influence
physiology
would
physiology give
such
vascular as retinopathy,
complications neuropathy,
such nephropathy
cardiovascular and
disease and added information that the influence
vascular
strokes.
vascular
such as complications
The American
complications
retinopathy, such
such as
Diabetes
neuropathy, as cardiovascular
Association
cardiovascular
nephropathy disease
estimates
disease
and and
and
macro- the
of
the a added
certain
added
scientists
thea added information
parameter
information
working
information on that
seems
that
understanding mathematically
to have
mathematically
that mathematically thelittle
human tothe
no
the
the influence
effect on
influence
physiology
influence on
vascular The
strokes. complications
American such as cardiovascular
Diabetes Association disease
estimates and of certain parameter seems to have little to no effect
strokes.
that
strokes.the
vascular The
cost
The American
associated
American
complications suchDiabetes
with
Diabetes
as the Association
treatment
Association
cardiovascular and estimates
produc-
estimates
disease and of
of
the a
a certain
glucose
certain
added
of a glucose parameter
concentration:
parameter
information
certain concentration:
parameter seems seems
seems
that hence to
to have
mathematically
to have little
concluding
have little to
to no
that
no
the
little tothat effect
the
effect
influence
no effect on
par-
on
on
strokes.
that the The
cost American with
associated Diabetes the Association
treatment and estimates
produc- the hence concluding the par-
that
tivity
that the
loss
the
strokes. cost
of
cost
The associated
patients
associated
American (with with
with the
diabetes)
Diabetes the treatment
has
treatment
Association risen and
and fromproduc-
$245
produc-
estimates the
the
of
the a glucose
ticular parameter
glucose
certain
glucose concentration:
parameter may
concentration:
concentration: not
seems hence
be
hence
hence to concluding
significant
concluding
have little
concluding in tothat
the
that
no
that the
biological
the
effect
the par-
par-
on
par-
that the
tivity loss cost
of associated
patients (with with the treatment
diabetes) has risen and from produc-
$245 ticular parameter may not be significant in the biological
tivity
billion
tivity
that loss
in
loss
the of
2012
of
cost patients
to
patients
associated$327 (with
(with diabetes)
billion
with in
diabetes)
the 2017,has
has
treatment arisen
34
risen and% from
from $245
increase
$245
produc- ticular
process.
ticular
the parameter
glucose
ticular The
parameter
parameter may
insignificance
may
concentration:
may not
not
not be
of
be
hence
be significant
the parameters
significant
concluding
significant in
in
in the
could
the
that
the biological
also
biological
the
biological be
par-
tivity
billion loss
in of
2012 patients
to $327 (with diabetes)
billion in 2017,has risen
aa 34 % from $245
increase process. The insignificance of the parameters could also be
billion
(Association
billion
tivity in
in
loss 2012
2012
of et to
toal.
patients $327
(2018)).
$327 (withbillion
billion in 2017,
Therefore
in
diabetes) 2017,has 34
there
a 34
risen %
%is
fromincrease
a clear
increase
$245 process.
interpreted
process.
ticular The
The
parameterby insignificance
the
process. Thebyinsignificance control
insignificance
may not of
and
of
be
of the the parameters
optimization
the parameters
significant in could
community
could
the
parameterscommunity also
could also be
also
biological be
as
be
billion in 2012
(Association et toal.$327
(2018)).billionTherefore
in 2017, there a 34 % is increase
aa clear interpreted the control and optimization as
(Association
motivation
(Association
billion in from
2012 et
et toal.
a
al. (2018)).
quality
(2018)).
$327 of
billion Therefore
life and
Therefore
in 2017, from there
there
a 34 a is
health
%is a clear
care
clear
increase interpreted
an opportunity
interpreted
process.
interpreted The by
by the
to
the control
reduce
control
insignificance
by the control and
the
and
of
and optimization
uncertain
optimization
the parameters
optimization community
dimension
community
could
community of
also as
the
as
be
as
(Associationfrom
motivation et al. (2018)).
aa quality of Therefore
life and from there is a clear
aa health care an opportunity to reduce the uncertain dimension of the
motivation
economics
motivation
(Association from
point
fromet ofa
al. quality
view,
quality
(2018)). to of life
develop
of life and
and
Thereforean from
Artificial
from there a health
Pancreas
health
is a care
care
clear an
model
an opportunity
interpreted
an when
opportunity
opportunity by to
solving
to
the
to reduce
stochastic
reduce
control
reduce the
the
and
the uncertain
optimal
uncertain
optimization
uncertain dimension
control.
dimension Other
community
dimension of
of
of the
ben-
the
as
the
motivationpoint
economics from of a view,
quality to of life and
develop an from
Artificial a healthPancreascare model when solving stochastic optimal control. Other ben-
economics
(AP)
economics which
motivation point
pointcan
from of
ofa view,
emulate
view,
quality to
to develop
the
of human
develop
life andan
an Artificial
pancreas
Artificial
from a Pancreas
as closely
Pancreas
health care model
efits
model
an
model of when
SA
when
opportunity
when cansolving
be
solving
to
solving stochastic
found
reduce in
stochastic
stochastic articles
the optimal
optimal
uncertain
optimal by control.
Saltelli
control. et
dimension
control. Other
al.
Other
Other of ben-
(2004,
ben-
the
ben-
economics
(AP) which pointcan of view,
emulate tothedevelop
human an Artificial
pancreas Pancreas
as closely efits of SA can be found in articles by Saltelli et al. (2004,
(AP)
as
(AP) which
possible.
economicswhichpointTocan that
canthat
ofemulate
end, to
emulate
view, the human
researchers
the human
develop pancreas
have
pancreas
anhave developed
Artificial as
as closely efits
2000);
efits
sev- 2000);
closely
Pancreas model of
efits of SA
ofwhen can
Campolongo
SA
SA can can be
be
solving found
be found in
et al.
al.
in
stochastic
found articles
(2007);
in articles
articlesoptimal by
by Saltelli
Saltelli
by Saltelli
control.
Saltelli et
etet al.
al.al.
etOther (2004,
(2010);
(2004,
ben-
al.(2010);
(2004,
as(AP) which
possible. Tocan emulate
end, the human
researchers pancreas developed as closely
sev- Campolongo et (2007); Saltelli et al.
aseral
as
(AP)possible.
mathematical
possible.
which To
Tocan that
that end,
models
end,
emulate researchers
for predicting
researchers
the human have
have the
pancreas developed
blood
developed as sev-
glucose
sev-
closely 2000);
Borgonovo
2000);
efits of Campolongo
SA
2000); Campolongo and
Campolongo
can Plischke
be found et
et al.
(2016)
al.
in
et al. (2007);
(2007);
articles
(2007); and Saltelli
references
Saltelli
by Saltelli
Saltelli et
et etal.
there
al.
et thereal.(2010);
in.
(2010);
(2004,
al. (2010);
as possible.
eral mathematical To that end, researchers
models for predicting have the developed
blood sev- Borgonovo
glucose and Plischke (2016) and references in.
eral
levels
eral
eral
levels
mathematical
in T1D
T1DTopatients.
mathematical
as possible.
mathematical
in patients.
that models
models for
for predicting
end, researchers
models for predicting
predicting have the blood
blood glucose
thedeveloped
the blood sev- Borgonovo
glucose
glucose Borgonovo
2000);
Borgonovo
This paper
and
and
Campolongo
and Plischke
Plischke
Plischke
presents a
(2016)
(2016)
et al.
(2016)
novel
(2007); and
and
sensitivity
references
references
and Saltelli
references
analysis
et there
there
there
method
in.
in.
al. (2010);
in. to
levels
levels in T1D
in T1D patients.
T1D patients.
eral mathematical
levels in patients.models for predicting the blood glucose This This paper
Borgonovo presents
and Plischkea novel
(2016)sensitivity
and analysis
references method
there in. to
Over the the past
past few few decades,
decades, models models for for T1D T1D have have in- rank
This
in- rank paper
uncertain
This paper
paper presents
presents a
parameters
presents novel
aa novel
novel of sensitivity
of a T1D
T1D model.
sensitivity
sensitivity analysis
model.
analysis
analysis The method
method
method
method to
to
to
Over levels in T1D patients. rank uncertain
uncertain parameters
parameters of a
a T1D model. The
The method
method
Over
creased
Over
Over the
the
the past
their
past
past few
fidelity
few
few decades,
in
decades,
decades,terms models
of
models
models for
prediction
for
for T1D
T1D
T1D and have
have
have in-
perfor-
in-
in- presented
rank
This
rank uncertain
paper
uncertainis a Global
parameters
presents
parametersaSensitivity
novel of a Analysis
T1D
sensitivity
of a T1D (GSA)
model.
analysis
model. The
The technique
method
method
method to
creased their fidelity in terms of prediction and perfor- presented
presented is
is a Global Sensitivity Analysis (GSA) technique
creased
mance
creased
Over the
creased
mance
their
pastfidelity
Palumbo
their
their
Palumbo et al.
fidelity
few
fidelity
et
in
in
in terms
al.decades,
(2013). of
of prediction
However,
termsHowever,
terms
(2013). models
of prediction
for
prediction thisT1D
this
and
development
andhave
and
development
perfor-
perfor- that
presented
rank
in- that
perfor- presented
is is aaa Global
isuncertain
dependent
is
dependent
Global
parameters
Global on
Sensitivity
on Sensitivity
the probability
probability
Sensitivity
the
Analysis
of a Analysis
Analysis (GSA)
distribution
T1Ddistribution (GSA)
model.
(GSA) technique
function
Thetechnique
method
technique
function
mance
takes
mance
creased Palumbo
place
Palumbo et
at fidelity
theirat the al.
price
et price (2013).
al. in of
(2013).
terms However,
model this
this development
complexities.
However,
of complexities.
prediction Inclusion
development
and perfor- (pdf)that
(pdf)
that
that isis
presented
isof dependent
ofdependent
the
is a evolving
dependent Global on
evolving
on the
the probability
blood
probability
on Sensitivity
the probabilityglucose
Analysis distribution
concentration.
distribution function
function
(GSA) technique
distribution functionThe
mance
takes Palumbo
place the et al. (2013).
of However,
model this development
Inclusion the blood glucose concentration. The
takes
of
takes
mance place
systems
place
Palumbo at
of
at the price
differential
the et price
al. of
of
(2013). model
equations
model complexities.
capturing
complexities.
However, this Inclusion
more sub-
Inclusion
development (pdf)
Bergman
(pdf)
that is of
of
(pdf) of the the
model
the
dependent evolving
for
evolving
on
evolving theblood
glucose
blood –
probability
blood– glucoseglucose
insulin
glucose concentration.
dynamics
concentration.
distribution
concentration. Bergman
functionThe
The
The
oftakes place of
systems at differential
the price ofequations model complexities.
capturing Inclusion
more sub- Bergman model for glucose insulin dynamics Bergman
of of systems
systems
systems
takes of
place of
the
of
at differential
human
differential
the price equations
physiology,
of equations
model capturing
although
capturing
complexities. leads more
moreto sub-
better
sub-
Inclusion Bergman
et
(pdf)al.
Bergman
Bergman of model
(1981),
model
the
model is for
chosen
for
evolving
for glucose
glucoseto
blood
glucose –
– insulin
illustrate
– insulin
glucose
insulin dynamics
the proposed
dynamics
concentration.
dynamics Bergman
Bergman
BergmanGSA
The
of
systems systems of of
the differential
human equations
physiology, capturing
although leads moreto sub-
better et al. (1981), is chosen to illustrate the proposed GSA
systems
model
systems
of systems of
of the
prediction,
the human
human physiology,
increases
of differential physiology, although
the number
equationsnumber
although
capturing of leads
states
leads moreto as
toas better et al.
well formulation.
better
sub- Bergman (1981),
formulation.
et
et al.
al. (1981), model
(1981), is chosen
Parameters
is chosen
is for
chosenglucoseto
to illustrate
in illustrate
to the
– insulin
illustrate the
modeldynamics
the
the proposed
areproposed
considered
proposed BergmanGSA
GSA
GSA to
systems
model of the
prediction, human physiology,
increases the although of leads
states to better
well Parameters in the model are considered to
model
as
modelthe
systems prediction,
parameters.
prediction,
of the human increases
Typically,
increases
physiology, the
the number
measurements
number
although of
of states
from
states
leads to as
real
as well
pa-
well
better formulation.
be
et random
formulation.
al. (1981),
formulation. Parameters
variables
Parameters
is chosen
Parameters in
(including
in
to the
the model
the
model
illustrate
in the the model theare
initial
are considered
blood
considered
proposed
are blood
considered glucose
GSA to
to
to
asmodel the prediction, increases
parameters. Typically, the number of states
measurements from as well
real pa- be random variables (including initial glucose
astients
as the
the
model parameters.
are used
parameters.
prediction, to Typically,
estimate
Typically,
increases thesemeasurements
the parameters.
measurements
number of from
Errors
from
states real pa-
associ-
real
as pa-
well be
be random
level) to
random
formulation.
be random variables
represent
variables the
Parameters
variables (including
inter-
(including
in
(including and
the the
model
the initial
intra-patient
the initial
are blood
blood
considered
initial blood glucose
variability.
glucose
glucose to
as theare
tients parameters.
used to Typically,
estimate thesemeasurements
parameters. from real
Errors pa- level)
associ- to represent the inter- and intra-patient variability.
tients
ated
tients
as the are
with
are used
the
used
parameters. to
to estimate
experimental
estimate
Typically, these
process
these parameters.
of measurement
parameters.
measurements Errors
Errors
from associ-
as well
associ-
real pa- level)
level)
be to
to
random
level) to represent
represent
variables
represent the
the
the inter-
inter-
(including
inter- and
and
and intra-patient
intra-patient
the initial
intra-patient bloodvariability.
variability.
glucose
variability.
tients
ated are
with used
the to estimate
experimental process these parameters.
of measurement as Errors associ-
well To propagate the uncertainties of the model through time,
ated
as
ated
tients
ated with
un-modeled
with the
the experimental
are used
with the physiological
experimental
to estimate these
experimental process
behavior,
process
process of
of measurement
lead to
measurement
parameters.
of measurement to parameters
Errors as
as well
well To
associ-
as well propagate
level) to represent the uncertainties
the inter- and of the model
intra-patient through time,
variability.
as un-modeled physiological behavior, lead parameters To
To propagate
the
To popular uncertainty
propagate
propagate the
the uncertainties
uncertainty
the uncertainties
uncertainties of
of the
quantification
of the model
the model
modeltool:through
Polynomial
through
through time,
time,
time,
asbeing
as
atedun-modeled
identified
un-modeled
with
as un-modeled the physiological
as random
physiological
experimental
physiological behavior,
variables
behavior,
process
behavior, with
of lead
lead
measurement
lead to
finite
to parameters
support
parameters
to parameters as wellor the popular quantification tool: Polynomial
being identified as random variables with finite support or the
Chaos
the
To
the popular
(PC)
popular
propagate
popular isuncertainty
used.
uncertainty
the The
uncertainties
uncertainty quantification
time evolving
quantification
of
quantification the modeltool:
pdfs
tool:
tool:are Polynomial
throughthen
Polynomial
Polynomial used
time,
being
abeing identified
probability
identified
asprobability
un-modeled as random
distribution.
as random
random variables
physiological variables
variables
behavior, with
with lead finite
finite support
support or
to parameters or
or Chaos (PC) is used. The time evolving pdfs are then used
Chaos (PC) is used. The time evolving pdfs are then
aabeing identified
probability
as
distribution.
distribution.
with finite support to
the rank
Chaos
Chaos popularthe
(PC)
(PC) parameters
is used.
uncertainty The
is used. Theoftime of the
time model.
evolving
quantification
evolving Using
pdfs
tool:a
are
pdfs aare quantitative
then
then used
Polynomial used
used
abeing
aEvenprobability
identified
probability as random variables with finite support or to
distribution.
distribution. to rank
rank
measure,
the
theit
parameters
parameters
is determined of the
the
which
model.
model. Using
Using
parameter a quantitative
quantitative
has the most
Even with
with the
the computational
computational resources
resources available
available today,
today, the
the to rank
Chaos
to rank the
measure, the
(PC) it parameters
is used.
parameters
is determined The of the
time
of the
which model.
evolving
model. Using
pdfs
Using ahas
parameter a
arequantitative
then
quantitative
the used
most
Evena probability
with the distribution.
computational resources available today, the measure,
influence it
on is determined
the pdf of which blood parameter
glucose has the
level. most
After
increase
Even with
Even withinthe
increase inthethe
the number
computational
computational
number of
of parameters
resources makes
available
resources available
parameters makes it difficult
today,
ittoday, the measure,
to rank
measure, theit
it is
is determined
parameters
determined of which
the
which
the influence on the pdf of the blood glucose level. After
difficult model. parameter
Using
parameter ahas
has the
quantitative
the most
most
increase
toEven within
quantify
increase inthethe
the
the number of
of parameters
state uncertainty
number
computational uncertainty
parametersin the
resources the makes
T1D it
makes
available difficult
models
it difficult
today, influence
comparing
influence
influence iton
as comparing
the measure, on the
onthe the
is the pdf
measures,
pdf
pdf of
determined
the of the
of thewhich
the blood
an evaluation
evaluation
blood
blood glucose
parameter is made
glucose
glucose level.
made
level.
has theon
level. After
onAfter
the
most
After
toincrease
quantify in the
the number
state of parametersin makes
T1D it
modelsdifficultas measures, an is the
toato quantify
function
quantify
increase in ofthe
the
the state
parametric
state
number uncertainty
of in
uncertainties.
uncertainty
parametersin the
the A T1D models
comprehensive
T1D
makes models
it difficultas
as comparing
relative
comparing
influence
comparing on the
importance
the themeasures,
pdf
the measures, of
measures, the
of an
uncertain
an
the evaluation
evaluation
blood
an evaluation is
parameters.
is
glucose made
made
level.
is madeCompar- on
Compar-
on
on thethe
After
the
aatofunction
quantifyofthe
function of
state uncertainty
parametric
parametric
in theAT1D
uncertainties.
uncertainties. A
models as relative
comprehensive
comprehensive relative importance
importance of
of the
the uncertain
uncertain parameters.
parameters. Compar-
Sensitivity
ato function
quantify
a function of
Sensitivity Analysis
oftheparametric
Analysis state
parametric (SA)
(SA) would
uncertainty have
uncertainties.
in
uncertainties.
would have a
the
a A
Aprofound
comprehensive
T1D models
comprehensive
profound impact
impact as isons
relative
comparing
relative
isons are
are also
importance
the
importance
also made
made to
of
measures,
of
to the
the
the
the variance
uncertain
an
uncertain
variance evaluationbased isGSA
parameters.
parameters.
based GSAmade technique
Compar-
on
Compar-
technique the
Sensitivity
ona several
Sensitivity
function Analysis
avenues
Analysis
of of
parametric (SA)
research
(SA) would
would and have
uncertainties. a
development.
have a Aprofound
profound It
comprehensive impact
would
impact isons
using
isons
relative
isons are
Sobol
are
are also
also
importance
also made
indices.
made
made to
Final
to
of
to the
the
the
the variance
results
variance
uncertain
variance are based
seen
based GSA
to
GSA
parameters.
based GSAbe technique
extremely
technique
Compar-
technique
onSensitivity
several Analysisof(SA)
avenues research would and have a profoundItimpact
development. would using Sobol indices. Final results are seen to be extremely
onhelp
on several
us better
several
Sensitivity avenues
better
avenues
Analysis of
understand research
of(SA)
research and
would and development.
rank
andhave which
development.
awhichprofound It
parameters would
Itimpact using
isons
using are
would interesting Sobol
interesting
using Sobol
Sobol also indices.
in made
indices. Final
the context
indices. context
Final
Final results
to theresults
variance
results are
of understanding seen
understanding
are seen
seen to
arebased to
GSA
to be
be extremely
T1D.
extremely
beT1D.technique
extremely It is
is
on
help several
us avenues of
understand research andand development.
rank It
parameters would in the of It
help
help
on us
contribute
us
several better
better
avenues understand
(jointly) to
understand
of the
research and rank
variation
andand rank ofwhich
which
development. glucose parameters
concen-
parameters
It would interesting
observed
interesting
using Sobol
interestingthat that in
in the
a
the
indices.
in athe context
couple
context
Final
context of of
results understanding
parameters
of understanding
are
of understanding seenwhich
to be T1D.
are
T1D.
extremely
T1D. It
directly
It is
It is
is
help us better
contribute understand
(jointly) to the and rank ofwhich
variation glucose parameters
concen- observed couple of parameters which are directly
contribute
contribute
help us (jointly)
(jointly)
better to
to
understand the
the variation
variation
and rank of
ofwhich glucose
glucose concen-
concen-
parameters observed
linked
observed
interesting
observed to that
the
that in a
the
thatinsulin couple
insulin
a couple
a couple context of
sensitivity
of parameters
of
of parametersof
parameters a patient
understanding which
which
whichhave are
are
T1D.
are directly
significant
directly
It
directly is
contribute
 (jointly) by
Research Sponsored to Juvenile
the variation
Diabetes of glucoseFoundation
Research concen- linked to the sensitivity of a patient have significant


Research Sponsored
contribute (jointly) byto Juvenile
the Diabetes of
variation Research
glucose Foundation
concen- linked
impact
linked
observed
linked to
to
toonthe
thetheir
that
the insulin
a blood
insulin
couple
insulin sensitivity
glucose
sensitivity
of
sensitivity of
parameters
of aaa patient
level.
of patient
patientwhichhave
have
have aresignificant
significant
directly
significant
(JDRF)
(JDRF)
Research
 Research
Research
under
under
Sponsored
JDRF Grant
Sponsored
JDRF Grant
Sponsored
by
by Juvenile
by Juvenile
Diabetes
Key: 1-SRA-2019-823-S-B.
Diabetes Research
Diabetes Research
Key: 1-SRA-2019-823-S-B.
Juvenile
Foundation
Research Foundation
Foundation impact
impact on
on their
their blood
blood glucose
glucose level.
level.
(JDRF)

(JDRF) under
under JDRF Grant
JDRF Grant Key: 1-SRA-2019-823-S-B.
Key: 1-SRA-2019-823-S-B. impact
impacttoon
linked onthetheir
their blood
insulin
blood glucose
glucose level.
sensitivity of a patient have significant
level.
(JDRF)Research underSponsored
JDRF Grant by Juvenile Diabetes Research Foundation
Key: 1-SRA-2019-823-S-B.
(JDRF) under JDRF Grant Key: 1-SRA-2019-823-S-B. impact on their blood glucose level.
2405-8963 © 2020, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2020.12.431
 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118
Table 1. Statistical Distances and their math-
ematical expressions
Statistical A Description
Distance (D)
Wasserstein W
Ω
|PY (y) − QY (y)|dy
 y    12
Hellinger H 2 1− pY (y)qY (y)dy
 Ωy
Total Variation T 0.5 |pY (y) − qY (y)|dy
Ωy
Kolmogorov K supy∈Ωy |PY (y) − QY (y)|
Bhattacharya B −log(
Ωy
pY (y)qY (y)dy)
  12
Cramer-von Mises C |PY (y) − QY (y)|2 dy
Ωy
The paper has been structured in the following way. Sec-
tion 2 introduces the moment independent metrics that
have been proposed. Section 3 presents an efficient algo-
rithm to determine the afore-mentioned metrics. Section
4 presents the results from the implementation of the
GSA metrics on the Bergman model; before ending with
concluding remarks in section 5.
2. MOMENT INDEPENDENT METRICS
Similar to Sobol’ indices, several sets of metrics are pre-
sented (also referred to as classes) which capture the con-
tributions of uncertain variables varied individually (first
order effect) and when varied concurrently. The classes are
based on observing the disparity between the output pdf
of a function and the conditional output pdf evaluated at
certain specific locations over the input sub-space.
2.1 Statistical Distance measures
In probability theory, the disparity between two proba-
bility measures is often quantitatively determined using
metrics called statistical distances. Larger the value of
the distance, more distinct are the pdfs. Based on the
penalty levied on the disparity measures, there exists more
than one type of distance. The quantitative value (of
distance) changes with the choice of statistical distance
used. However, they retain certain basic properties such
as non-negativity, symmetry, positive definiteness among
others. A list of important or popular Statistical Distances
(D) have been presented in Table 1. The first and second
columns present the name and a convenient abbreviation
of different Ds respectively. The final column presents
the mathematical expression, where y is used to repre-
sent a realization of the random variable Y , pY (y) and
qY (y) are two distinct pdfs over Y , PY (y) and QY (y) are
the corresponding cdfs and ΩY is the support of Y (i.e.
y ∈ ΩY ). The descriptions for Wasserstein, Hellinger, Total
Variation, Kolmogorov distance can be found in Gibbs
and Su (2002), Bhattacharya distance in Bhattacharyya
(1943) and Cramer-von Mises distance in Baringhaus and
Henze (2017). Henceforth in this article, DA (p, q) is used
to represent the statistical distance with abbreviation A.
For example, DW would refer to the Wasserstein distance.
We present 6 different measures for GSA (refer to Table
1). The measures are different based on the nature of their
penalty with disparity and computational effort required
to compute them.
16113
The Ds can be used to define certain global sensitivity
metrics by comparing disparities between output pdfs
and conditional output pdfs. Depending on the nature of
conditional pdfs, the metrics are divided into classes. For
functional relationships of the form
Y = g(X) (1)
where X ∈ R is a vector of n random inputs, Y ∈ R is
n
a scalar output and g : X → Y is a function which maps
the inputs to the output; class 1 considers the conditional
output pdf fY |Xi (y, xi ) which is the pdf of the output Y
for a given fixed input value of Xi . Class 2 considers the
conditional pdf fY |Xi ,Xj (y, xi , xj ) which is the pdf of the
output Y given fixed values of Xi and Xj . Following a
similar pattern we get the higher class metrics. Finally,
a total effect metric for each individual input Xi is also
defined which sums up all the class order metrics which
are relevant to Xi . The total effect metric is analogous to
the total effect Sobol’ indices. Details about the metrics
have been elaborated in the following subsections.
2.2 Class 1 metrics
Consider the equation

Di = D(fY (y), fY |Xi (y, xi ))fXi (xi )dxi (2)
ΩX i
where Di is the averaged statistical distance over the
input subspace ΩXi , ΩXi represents the domain of Xi (i.e.
xi ∈ ΩXi ),

fY (y) = fX,Y (x, y)dx, (3)
ΩX

fY |Xi (y, xi ) = fY,X̃i |Xi (x, y)dx̃i , (4)
ΩX̃
i
fX,Y (x, y) is the joint input output pdf, x is a realization
of X, X̃i is the joint variable X̃i = [X1 , ..., Xj , ...Xn ]T for
j = i, x̃i is a realization of X̃i , ΩX̃i is the domain of X̃i
and ΩX is the domain of X.
D(fY (y), fY |Xi (y, xi )) in equation (2) quantifies the dis-
parity between the output pdf and the output pdf condi-
tioned on a single input parameter. If the particular input
Xi (on its own) is unimportant, it would contribute mini-
mally to the marginalized output pdf: which would mean
fY (y) and fY |Xi (y, xi ) are in close proximity: resulting
in low values of D. In contrast, if the input parameter
(Xi ) is indeed influential, it would contribute significantly
to the output pdf. This means that the conditioned pdf
fY |Xi (y, xi ) would be far from the marginalized pdf fY (y)
leading to higher values of D. Hence, on observing what the
values of D are on average, one can estimate the relative
importance of inputs (i.e. larger the value of Di , more is the
significance of Xi ). Note that the metrics Di only represent
the first order effects of the inputs: which means the metric
only captures the disparity between the output pdf and the
conditioned output pdf when only a single input variable
Xi is varied across its domain ΩXi . It does not account for
the effects of the uncertainties when the input variables
are varied simultaneously. The effects from varying inputs
simultaneously are quantified only via higher order effect
metrics (i.e. Class 2 and higher).
16114 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118

2.3 Class 2 metrics Note that class n metric is not used during normaliza-
Similar to Class 1 metrics, consider the equation tion since it is a constant that would get added to the
 denominator of all the N S metrics. As only the relative
Di,j = D(fY (y), fY |Xi ,Xj (y, xi , xj ))× values of N S are significant, removing a constant from
ΩXi ,Xj the common denominator does not effect the ranking of
the magnitudes of the N S metrics. Evident from the
fXi ,Xj (xi , xj )dxi dxj (5)
definitions, N S metrics vary between 0 and 1. Closer the
where Di,j is the averaged statistical distance over the value of N S to 1, more is the significance of the joint
input subspace ΩXi ,Xj and input uncertainty corresponding to the metric. It should
 be noted that the Borgonovo metric δi is a specific Di
fY |Xi ,Xj (y, xi , xj ) = fY,X̃ij |Xi ,Xj (x, y)dx̃ij . (6) where the chosen D type is the Total Variation distance
ΩX̃
ij and the Gamboa metric is analogous to the case where D
If the combination of Xi and Xj is an important contribu- type is the Cramer-von Mises distance. However, they only
tor to the uncertainty in the output, then the conditioned consider the first order effects and the effect of the joint
pdf, fY |Xi ,Xj (y, xi , xj ) would be further apart from the variation of the uncertain inputs are not investigated. To
marginalized pdf fY (y) since Xi and Xj have been fixed the authors’ best knowledge, a metric based on statistical
and are not being varied. This would yield higher values distances to capture the first order as well as higher order
of the statistical distance D in equation (5). Similar to the effects of contributing input random variables on the na-
class 1 metric, averaging this distance over the subspace ture of the output uncertainty is investigated for the first
of ΩXi ,Xj would be indicative of the joint contribution of time in this work. To quantify the total effect a single input
Xi and Xj . Analogous arguments can be made about the variable has on the uncertainty of the output, (similar to
joint contribution of m input variables, quantified as class the total effect Sobol’ indices), total effect N STi metrics
m metic. If there are a total of n inputs, the final metric are also defined. The metrics
 are evaluated as
would be a class n metric which would account for the N ST i = N Si1 ,··· ,in−1 (12)
variation of the output due to the joint variation of all the Pi
n inputs. where Pj = {(i1 , · · · , in−1 )∃k, 1  k  n, ik = j}. The
N STi are simply the sum of all the partial sensitivity
2.4 Class n Metric measures N S where the influence of Xi has been accounted
The class n metric for either in part or in whole.
 is defined as
Di1 ,i2 ,...,in = D(fY (y), fY |X (y, x))fX (x)dx. (7) 3. EFFICIENT EVALUATION OF N S INDICES
ΩX
Note that fY |X (y, x) is a Dirac delta function because We often encounter problems in engineering where even a
fixing all the inputs would make the output deterministic single function evaluation is computationally expensive.
with a definite value. Therefore we would get For those functions, deriving the aforementioned GSA
fY |X (y, x) = δ(y − g(X)) (8) measures can become impractical especially when output
and D would measure the statistical distance between the pdfs and conditioned output pdfs need evaluation. Either
output pdf and a delta function. Averaging that distance analytical expressions for these pdfs do not exist, cannot
over the entire ΩX would yield the final class n metric. be evaluated or can only be approximated from a large
number of sample realizations (eg. Monte Carlo methods):
2.5 Normalized Metrics making the calculation of the N S metrics through tradi-
tional techniques extremely difficult. Methods to reduce
Considering that the values of D can vary largely depend- the computational cost and provide tractable alternatives
ing on the type of statistical distance chosen to be imple- to approximate the N S metrics are presented next.
mented, a normalization of all the metrics are exercised to
facilitate comparisons. The normalized metrics are referred 3.1 Polynomial Chaos based Surrogate Model
to as N S and represent a shorthand for Non-moment based
Polynomial Chaos (PC) is a probabilistic modeling tool
Sensitivity indices. They are defined in the following way:
to approximate a stochastic function with a polynomial
Class 1 N S:
function of the random variables. First introduced by
Di Wiener (1938), to expand a Gaussian process with the help
N Si =    (9)
Di + Di,j + · · · + Di1 ,··· ,in−1 of an infinite series using Hermite polynomials, PC has
Class 2 N S: been subsequently thoroughly investigated by a number
Di,j of researchers Ghanem and Spanos (1991); Cameron and
N Sij =    (10) Martin (1947); Xiu and Karniadakis (2002). In this paper,
Di + Di,j + · · · + Di1 ,··· ,in−1 we determine a surrogate model Ŷ using PC for the true
.. system Y = g(X) such that instead of sampling Y (which
.
can be expensive), we can sample Ŷ (the surrogate model)
and eventually instead, relatively cheaply. From PC theory, it is well
Class n − 1 N S: known that a stochastic function Y can be written as an
Di1 ,··· ,in−1 infinite polynomial series expansion in the form
N Si1 ,··· ,in−1 =    . ∞
Di + Di,j + · · · + Di1 ,··· ,in−1 Y = Yi Φi (X) (13)
(11) i=1
 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118
where Φi are certain specific orthogonal set of basis func-
tions in X and Yi ∈ R are their corresponding coefficients.
The nature of Φi is determined by probability measures of
X. The orthogonal bases required for some of the popular
random variable types can be found in the Wiener-Askey
scheme provided in Xiu and Karniadakis (2002). Equation
(13) is typically truncated to a finite number of terms as
an approximation to yield the surrogate model
N

Y ≈ Ŷ = Ŷi Φi (X).
i=1
The objective is to determine the coefficients Ŷi of the
surrogate model Ŷ so that one can have a simple model of
the true stochastic system as a polynomial function of the
input stochastic variables.
Traditionally there have been two broad category of meth-
ods to find those coefficients: namely Intrusive methods
and Non-Intrusive methods (Kim et al. (2013)). In this
paper, we briefly review a method from each of those
categories.
Intrusive PC In intrusive PC, we look for coefficients
which minimize the mean value of the square of the model
error, leading to the following closed form expression for
the coefficients
gΦi fX (x)dx
Ŷi = ΩX 2 . (15)
Φ f (x)dx
ΩX i X
This method of evaluating the coefficients is also popularly
known as the Galerkin projection method.
Although calculating the denominator of equation (15)
is trivial, the numerator can be extremely difficult to
evaluate for generic non-linear functions and places where
g is merely a computer code. Hence in spite of obtaining
a clean expression for the coefficients in equation (15) the
need for the evaluation of a multivariate integral forms the
most profound limitation of this method.
Non-Intrusive PC In order to circumvent multidimen-
sional integrals, the coefficients can also be determined
using function evaluations and least squares. Consider the
expression: e =
   
Φ1 (x(1) ) Φ2 (x(1) ) · · · ΦN (x(1) ) Ŷ1 y(x(1) )
 Φ1 (x(2) ) Φ2 (x(2) ) · · · ΦN (x(2) )   Ŷ2   y(x(2) ) 
   
 .. .. ..  . − ..
 . . ··· .   .
.   . 
Φ1 (x(m) ) Φ2 (x(m) ) · · · ΦN (x(m) ) ŶN y(x(m) )
        
A YP C
(16)
where x(i) represents the ith sample point from a total of m
samples in the input space, A ∈ Rm×N is a matrix whose
rows represent the bases evaluated at x(i) , y is a vector
assimilated using the transformation g over x(i) samples
and e is a vector of the approximation errors at each
sample point. The coefficients YP C can be determined by
minimizing the 2-norm of e and is given by
YP C = (AT A)−1 AT y. (17)
Equation (17) now provides a simple expression to de-
termine the coefficients of the PC surrogate model from
system realizations of the original function.
16115
Since sampling the surrogate is much cheaper than sam-
pling the true system, we now have a tractable way
to determine the computationally expensive pdfs fY (y),
fY |Xi (y, xi ) and fY |X̃i (y, x̃i ) using MC methods.
3.2 Efficient Evaluation of the Expectation Integral
Expectation integrals of the form

I= k(x)fX (x)dx (18)
(14) ΩX
turn up in numerous applications of basic and applied
sciences. As a result, many researchers over the years
have endeavoured to efficiently evaluate this integral. Some
of the most popular methods used have been Monte
Carlo (MC) sampling methods (Stroud (1971)), Gauss
quadrature (GQ) rules (Stroud and Secrest (1966)), Sparse
quadrature rules (Gerstner and Griebel (1998)) and Con-
jugate Unscented Transform (CUT) rules (Adurthi et al.
(2018)). In this paper, the methods are not highlighted (as
there is considerable existing literature). However, we do
provide commentary on which method to use for the N S
metrics.
In all of these methods, the integral I is approximated by
a weighted sum of function evaluations
Nmethod
I ≈ Iˆ = wi k(x(i) ) (19)
i=1
where x(i) are certain samples from the input space,
Nmethod denotes the number of such samples and wi are
specific weights. The aforementioned methods primarily
differ in the manner by which the location of the sample
points x(i) and their corresponding weights wi are deter-
mined. The benefit of representing an integral with func-
tion evaluations lies in the fact that sample realizations
are independent of each other and parallel computing tech-
niques can be adopted to evaluate the system realizations
simultaneously.
Remarks on the use of MC, GQ and CUT The objective
of discussing expectation integrals has been to highlight
the fact that in order to determine the values of N S, we
first need to evaluate equations (2), (5) and (7) which are
 expectation integrals.
Di results from a univariate integral while the other class
 metrics are outcomes of multidimensional integrals. These

integrals are approximated as
nmethod
(j)
Di ≈ wj D(fŶ (ŷ), fŶ |Xi (ŷ, xi )) (20)
y j=1
nmethod
Di,j ≈
(k)
wk D(fŶ (ŷ), fŶ |Xi ,Xj (ŷ, xi , xj ))
(k)
(21)
k=1
..
.
nmethod
Di1 ,i2 ,··· ,in ≈ wj D(fŶ (ŷ), fŶ |X (ŷ, x(j) )) (22)
j=1
(j)
where xi is the j th sample point out of a total of nmethod
samples in the ΩXi . Note that y has been replaced by ŷ
in the equations to represent the surrogate model instead
of the true system. nmethod represents the total number
16116 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118
of sample points in the used sampling algorithm. In the
following section, commentary on which sampling method
to use for the metrics is provided.
Since the convergence of MC is slow and requires an
enormous number of samples to evaluate equations (20)
through (22), it is never used to determine Di . GQ is
always used to evaluate class 1 metrics Di since it is an
univariate integral and GQ provides the minimal set of
points and weights to integrate a polynomial of any order
for univariate integrals. For input vectors X which have
a uniform distribution or a Gaussian distribution, CUT
is preferred to evaluate higher order classes of metrics
considering it requires fewer number of points than GQ.
However, if the input variables have distributions which
are not uniform or Gaussian, GQ should be adopted.
Although other sparse quadrature rules have not been
discussed, they can also be used instead of GQ to calculate
the higher order classes.
3.3 Summarized Review of N S evaluation
This subsection now elaborates the step by step process
needed from start to finish to yield the desired N S metrics
using results from all the previous sections.
Step 1: Develop the surrogate model Ŷ from the model
equation Y = g(X) using PC.
Step 2: Determine the output pdf: fŶ (ŷ). This is done
by sampling the input space ΩX , evaluating the surro-
gate function at each of those samples and plotting the
histogram of the outputs ŷ.
Di as well as Di1 ,··· ,in are approximated by a weighted
sum of Ds evaluated at strategic points as represented by
equations (20) through (22). The j th sample point for Di
(j) (j)
is denoted by xi . For each xi we need
(j)
D(fŶ (ŷ), fŶ |Xi (ŷ, xi )) and for each
(j)
D(fŶ (ŷ), fŶ |Xi (ŷ, xi )) we need the pdf: fŶ |Xi (ŷ, xi ).
This pdf is approximated by sampling the ΩX̃i space,
evaluating the surrogate model at the samples by holding
(j)
Xi = xi and finally plotting the histogram of ŷ. Each
(j)
value of D obtained from fŶ |Xi (ŷ, xi ) is then stored for
assimilation later. In a similar way, for the higher order
class of metrics, the surrogate model is randomly sampled
while holding the joint input variables at specific values.
Subsequently plotting the histogram yields the conditional
pdfs.
Step 3: Obtain the weighted sum of all the stored Ds to
yield the nth class metrics.
Step 4: Determine the normalized metrics N S using equa-
tions (9) through (11).
4. GSA OF T1D MODEL
This section presents the global sensitivity analysis study
on the Bergman minimal model Bergman et al. (1981),
using the N S metrics. Comparisons are also made with
total effect Sobol’ indices. Finally, based on the relative
ranking of the input uncertain parameters, inferences are
drawn about the most influential factor that impacts the
blood glucose concentration.
Table 2. Type 1 Diabetes Model Parameter
Parameter Value Parameter Value
pnominal
1 0.0287 (±0.0086) Gb 119.1858
pnominal
2 0.0283 (±0.0085) Ib 15.3872
pnominal
3 5.035(±1.51)E − 5 d 0.05
p4 5/54 Gnominal 119.18 (±35.75)
0
4.1 Model and Simulation Environment
In this section, we look at the simple Bergman model
for our numerical analysis regarding T1D. The following
equations for the Bergman model are considered after
accounting for a basal insulin input rate:
Ġ(t) = −(Xb (t) + p1 )G(t) + p1 Gb + D(t) (23)
Ẋb (t) = −p2 X(t) + p3 (I(t) − Ib ) (24)
˙ = −p4 (I(t) − Ib ) + U (t)
I(t) (25)
where Xb is an intermediate state to capture the inter-
action between the blood glucose (G) and insulin (I)
concentration. The meal disturbance D(t) to the glucose
concentration term is determined using the Fisher model
(Fisher (1991)):

0 t < tm
D(t) = (26)
Be−d(t−tm ) t ≥ tm .
For this study, the meal time tm is considered to be
constant at tm = 15 minutes. B which characterizes the
meal quantity is also assumed to be constant with a value
of B = 28.98 corresponding to a 45gm CHO meal (the
value of B is derived identical to Nandi et al. (2017)).
To compensate for the meal at time tm , an insulin bolus
at the start of the simulation t = 0 is administered.
This action is simulated by making the insulin input term
(U (t)) an impulse function lasting for a minute (between
t = 0 to t = 1). The magnitude of the impulse function is
determined using the following formula.
(j) 
 1000 × (CHO Amount in g)
0<t≤1
U (t) = CR × Vi (27)
0 1<t
where CR is the insulin-to-carb ratio and Vi is the distri-
bution volume of insulin. For a meal comprising 45gm of
CHO, a patient CR = 18.477 and Vi = 12, the impulse
magnitude turns out to be 202.96 mU
L .
We assume four sources of uncertainties in this sensitivity
analysis study. The uncertainties lie with the initial glucose
concentration (i.e. G(t = 0)) and parameters p1 , p2 and p3 .
Each uncertain parameter is assumed to have a uniform
distribution with a nominal mean and a support which
varies ±30% about its mean (as an illustrative variation.
A more realistic distribution derived from clinical trial
data would ideally be more appropriate). Table 2 lists
all the parameters used for the simulations. The mean
values of the uncertain parameters are marked by the word
nominal. In terms of the convention used in the paper, the
uncertainties are grouped as
X = [X1 , X2 , X3 , X4 ]T = [p1 , p2 , p3 , G0 ]T . (28)
The objective now is to observe the contribution of each
of these uncertainties (as well their joint contribution)
towards the uncertainty in the blood glucose concentration
over time.
 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118
Fig. 1. Variation of N STi with time
4.2 Uncertainty Quantification of the Bergman Model
Since there are only four uncertain parameters in our
model of study, PC is sufficient in quantifying the prop-
agation of uncertainties through time. Consequently, a
PC surrogate model is developed with the idea that the
surrogate would be far cheaper to sample as compared to
the dynamic system. The basis functions chosen are that
of multivariate Legendre polynomials (as recommended
by the Wiener-Askey scheme for uniformly distributed
inputs) where the multivariate bases are derived from the
tensor product of univariate Legendre polynomials. The
PC order in each univariate direction is chosen to be
NXi = 5. Since the final set of bases is derived from a
tensor product of the univariate bases set, the total num-
ber of bases become N = NX1 NX2 NX3 NX4 = 54 = 625.
For this problem, the PC coefficients Yi (t) of the surrogate
model
625

Ĝ(t) = Ŷ (t) = Yi (t)Φi (X1 , X2 , X3 , X4 ) (29)
i=1
are determined using the non-intrusive least squares ap-
proach (refer to equation 17). The number of samples
used for the method were m = 100000. To verify that
the surrogate model is of acceptable accuracy, the first
couple of moments determined from the PC coefficients
and 10000 MC samples of the original stochastic dynamic
system were compared and were found to be consistent.
4.3 Computation of N S metrics
Since there are four uncertain input variables, i.e. n = 4,
we have a total of 15 N S metrics, namely: the class 1 met-
rics (N S1 , N S2 , N S3 and N S4 ), class 2 metrics (N S1,2 ,
N S1,3 , N S1,4 , N S2,3 , N S2,4 and N S3,4 ), class 3 metrics
(N S1,2,3 , N S1,2,4 , N S1,3,4 and N S2,3,4 ) and class 4 metric
(N S1,2,3,4 ). Each of these metrics are functions of time and
16117
hence need evaluation at every instant. The objective is to
observe the evolution of the relative importance of each
uncertain input with time.
For the class 1 metrics, a 30 point Gauss-quadrature rule
is used to average all the statistical distance measures D.
As all the input variables are uniformly distributed, we
can adhere to CUT for evaluating the average integrals
during evaluation of the higher class metrics. There is no
need to resort to Gauss quadrature rules. To determine the
output pdf and the necessary conditional pdfs, a total of
10000 MC samples are drawn and the histograms of their
outputs are recorded.
After each N S metric from every class is derived, the total
effect N S metrics are calculated: and the results have
been shown (in Figure 1) for all the statistical distance
measures. The total time of simulation has been considered
to be 100 minutes. The order of ranking estimated from
all the statistical distances are seen to be largely con-
sistent. The minimal variations in switching times could
be attributed to the fact that the different measures ex-
ercise varying penalties on pdf disparities. We see from
the figures that initially the most significant parameter is
X4 which is G0 or the initial blood glucose concentration
level. Such an observation is quite intuitive since G(t) is
our output of choice and at t = 0, G0 should be most
influential. It is comforting to note that the N S metrics are
successful at capturing that. However, as time progresses,
the influence of G0 declines as the effect of the insulin bolus
kicks in. This is evident from the fact that X3 and X2 start
dominating. X3 and X2 which are essentially parameters
p3 and p2 are associated with the insulin sensitivity of a
person with T1D; where formally, the insulin sensitivity
has been defined to be p3 /p2 (refer to Bergman et al.
(1981)). It is interesting to note that for the most part
of the simulation (at least until steady state starts setting
in), it is the sensitivity to insulin which is most significant.
Eventually, for all the statistical measures, the ranking of
importance settles to p2 , p3 , p1 and G0 .
4.4 Computation of Sobol’ Indices
To compare the performance of the N S metrics with stan-
dard practices of the GSA community, Sobol’ indices are
derived for the Bergman model under the same simulation
setup. It is rather trivial to approximate the Sobol’ indices
if a PC surrogate model for the system is available. Mere
algebraic and polynomial evaluations of the PC coefficients
yield the Sobol’ indices Sudret (2008). Since the considered
model is a dynamic system, we have Sobol’ indices which
are functions of time which estimate the fraction of the
output variance contributed by each input (or their con-
tributions). For a fair comparison with the total effect N S
metrics, the total effect Sobol’ indices STi are presented in
Figure 2.
It is interesting to note that a similar pattern emerges
from the total effect Sobol’ indices when compared to the
N S metrics. We observe a decline in the influence of G0
from an initial position of dominance followed by the rising
influence of p3 and p2 . The settling order of ranking is
also similar, given by: p2 , p3 , p1 and G0 . As a result, we
can conclude that quantifying up to the second moment is
enough for the Begman model to make inferences on the
influence of parameters.
16118 >Souransu Nandi et al. / IFAC PapersOnLine 53-2 (2020) 16112–16118
Fig. 2. Variation of the total effect Sobol’ indices with time
5. CONCLUSION
This work presents a comprehensive global sensitivity
analysis of the Bergman model for type 1 diabetes. The
problem is setup for the more traditional mode of treat-
ment where an insulin bolus is administered prior to a meal
on top of a basal rate. Certain model parameters have been
assumed to be uncertain and the influence of these input
uncertainties on the output has been quantified.
As the model is a set of ordinary differential equations,
sampling the stochastic system is computationally expen-
sive. As an alternative, polynomial chaos is used to develop
a surrogate model to ease the burden of sampling and
performing multivariate integrals.
Variance based Sobol’ indices as well as non-moment based
N S metrics are presented to rank the input variables de-
pending on their contributions to the output uncertainty.
Since the model is a dynamic system which predicts the
blood glucose concentration of a person with type 1 dia-
betes over time under the effect of a meal and insulin, the
GSA measures are evaluated as functions of time. It is seen
that the influence each parameter has on the output keeps
changing and after a meal, parameters associated with the
insulin sensitivity of a person matter the most.
ACKNOWLEDGEMENTS
This material is based upon work supported through the
Juvenile Diabetes Research Foundation (JDRF) under
JDRF Grant Key: 1-SRA-2019-823-S-B.
REFERENCES
Adurthi, N., Singla, P., and Singh, T. (2018). Conjugate
unscented transformation: applications to estimation
and control. Journal of Dynamic Systems, Measure-
ment, and Control, 140(3), 030907.
Association, A.D. et al. (2018). Economic costs of diabetes
in the us in 2017. Diabetes care, 41(5), 917–928.
Baringhaus, L. and Henze, N. (2017). Cramér–von mises
distance: probabilistic interpretation, confidence inter-
vals, and neighbourhood-of-model validation. Journal
of Nonparametric Statistics, 29(2), 167–188.
Bergman, R.N., Phillips, L.S., and Cobelli, C. (1981).
Physiologic evaluation of factors controlling glucose tol-
erance in man: measurement of insulin sensitivity and
beta-cell glucose sensitivity from the response to intra-
venous glucose. The Journal of clinical investigation,
68(6), 1456–1467.
Bhattacharyya, A. (1943). On a measure of divergence
between two statistical populations defined by their
probability distributions. Bull. Calcutta Math. Soc., 35,
99–109.
Borgonovo, E. and Plischke, E. (2016). Sensitivity anal-
ysis: a review of recent advances. European Journal of
Operational Research, 248(3), 869–887.
Cameron, R.H. and Martin, W.T. (1947). The orthogonal
development of non-linear functionals in series of fourier-
hermite functionals. Annals of Mathematics, 48(2), 385–
392. URL http://www.jstor.org/stable/1969178.
Campolongo, F., Cariboni, J., and Saltelli, A. (2007). An
effective screening design for sensitivity analysis of large
models. Environmental modelling & software, 22(10),
1509–1518.
Fisher, M.E. (1991). A semiclosed-loop algorithm for
the control of blood glucose levels in diabetics. IEEE
transactions on biomedical engineering, 38(1), 57–61.
Gerstner, T. and Griebel, M. (1998). Numerical integra-
tion using sparse grids. Numerical algorithms, 18(3-4),
209.
Ghanem, R.G. and Spanos, P.D. (1991). Stochastic Finite
Elements: A spectral Approach. Springer New York.
Gibbs, A.L. and Su, F.E. (2002). On choosing and
bounding probability metrics. International statistical
review, 70(3), 419–435.
Kim, K.K.K., Shen, D.E., Nagy, Z.K., and Braatz, R.D.
(2013). Wiener’s polynomial chaos for the analysis and
control of nonlinear dynamical systems with probabilis-
tic uncertainties [historical perspectives]. IEEE Control
Systems, 33(5), 58–67.
Nandi, S., Singh, T., Mastrandrea, L.D., and Singla, P.
(2017). Optimal meal time after bolusing for type 1 di-
abetes patients under meal uncertainties. In 2017 Amer-
ican Control Conference (ACC), 4412–4417. IEEE.
Palumbo, P., Ditlevsen, S., Bertuzzi, A., and De Gaetano,
A. (2013). Mathematical modeling of the glucose–insulin
system: A review. Mathematical biosciences, 244(2), 69–
81.
Saltelli, A., Annoni, P., Azzini, I., Campolongo, F., Ratto,
M., and Tarantola, S. (2010). Variance based sensitivity
analysis of model output. design and estimator for the
total sensitivity index. Computer Physics Communica-
tions, 181(2), 259–270.
Saltelli, A., Chan, K., Scott, E.M., et al. (2000). Sensitivity
analysis, volume 1. Wiley New York.
Saltelli, A., Tarantola, S., Campolongo, F., and Ratto,
M. (2004). Sensitivity Analysis in Practice: A Guide
to Assessing Scientific Models. John Wiley & Sons.
Stroud, A.H. (1971). Approximate calculation of multiple
integrals. Prentice-Hall.
Stroud, A.H. and Secrest, D. (1966). Gaussian quadrature
formulas. Prentice-Hall.
Sudret, B. (2008). Global sensitivity analysis using polyno-
mial chaos expansions. Reliability engineering & system
safety, 93(7), 964–979.
Wiener, N. (1938). The homogeneous chaos. Amer-
ican Journal of Mathematics, 60(4), 897–936. URL
http://www.jstor.org/stable/2371268.
Xiu, D. and Karniadakis, G.E. (2002). The Wiener–Askey
polynomial chaos for stochastic differential equations.
SIAM J. Sci. Comput., 24(2), 619–644.