You are on page 1of 15

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN BIOMEDICAL ENGINEERING

Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/cnm.1494


Coupled electromechanical model of the heart: Parallel nite
element formulation
Pierre Lafortune
1
, Ruth Ars
1
, Mariano Vzquez
1,
*
,
and Guillaume Houzeaux
1
1
BSC-CNS, Barcelona Supercomputing Center, Barcelona, Spain
2
IIIA-CSIC, Barcelona, Spain
SUMMARY
In this paper, a highly parallel coupled electromechanical model of the heart is presented and assessed.
The parallel-coupled model is thoroughly discussed, with scalability proven up to hundreds of cores. This
work focuses on the mechanical part, including the constitutive model (proposing some modications to
pre-existent models), the numerical scheme and the coupling strategy. The model is next assessed through
two examples. First, the simulation of a small piece of cardiac tissue is used to introduce the main features of
the coupled model and calibrate its parameters against experimental evidence. Then, a more realistic prob-
lem is solved using those parameters, with a mesh of the Oxford ventricular rabbit model. The results of
both examples demonstrate the capability of the model to run efciently in hundreds of processors and to
Received 24 June 2011; Revised 25 November 2011; Accepted 30 November 2011
KEY WORDS: cardiac mechanics; electrophysiology; parallelization; nite element methods
1. INTRODUCTION
Computational mechanics plays an important role in many engineering elds, and cardiac mechan-
ics is no exception. Nowadays, some models allow realistic simulations of the heart beat, and could
be used to study different pathologies or to understand better the physiological mechanisms reg-
ulating the contraction of the heart [1, 2]. Cardiac mechanics simulations represent a challenge in
terms of numerical algorithms, programming techniques and computational resources. If todays
models can take several weeks on single core computers with coarse meshes [3], the use of parallel
resources will be essential for tomorrows models.
Electromechanical nite element models developed during the last decades improved quickly
in terms of precision and the way they represent the geometry of the heart. However, very few
models are prepared to run on large supercomputers. To cite some of them, models developed by
Sainte-Marie et al. [4], Stevens et al. [5], Gktepe et al. [6], Kerckhoffs et al. [7], Nobile et al. [8]
and Gurev et al. [9] have been used successfully to study different aspects of cardiac functions, but
are made of relatively small meshes. This lack of resolution is an obstacle when trying to reproduce
the complex bers distribution. Note, on the other hand, that recent work of Hosoi et al. [10] and
Reumann et al. [11] show very promising results using large meshes and running in parallel on
thousands of cores. This paper is an effort in this direction.
In previous work, we have introduced the electrophysiology model [12] and the parallel environ-
ment [13] that we use in our framework. This paper addresses the computational mechanics side
of the problem, introducing a highly parallel model capable of running in thousands of processors,
*Correspondence to: Mariano Vzquez, Barcelona Supercomputing Center, Campus Nord UPC, Barcelona, Spain.

E-mail: mariano.vazquez@bsc.es
Int. J. Numer. Meth. Biomed. Engng. (2012)
reproduce some basic characteristic of cardiac deformation. Copyright 2012 John Wiley & Sons, Ltd.
Copyright 2012 John Wiley & Sons, Ltd.
P. LAFORTUNE ET AL.
which allows to perform cardiac simulations on realistic conditions, using meshes of millions of
elements with turnaround times of hours.
The computational model is implemented in Alya System, a high performance computational
mechanics code developed in Barcelona Supercomputing Center BSC-CNS. It is a exible platform
for running in parallel coupled problems using non-structured meshes. The nite element space dis-
cretization and the parallelization based on automatic domain decomposition for distributed memory
facilities are the main features of this strategy. The automatic domain partition is carried out using
the widely known open source software METIS. Based in a masterslave strategy, one core is used
to partition the problem and send each subdomain mesh to each of the other cores taking part of
the run. METIS partitions the original mesh by minimizing communication surface between subdo-
mains and maximizing load balance. Communication is carried out using message passing interface
(MPI). Parallelization is carried out transparently, so, if properly coded, each of the coupled prob-
lems retains the scalability properties: the fact that all the coupled problems are solved onto the same
mesh avoids interpolation, mesh imbalance, and so on. For a thorough description, refer to [13, 14].
Fully implicit and explicit schemes are implemented. The resulting code is able to run using ele-
ments of different space order (P1, P2,. . . , Q1, Q2,. . . ) and higher order time schemes, in 2D and
3D domains of a non-homogeneous anisotropic excitable media.
This paper is divided as follows. First, the electrophysiological model is briey resumed. Then,
the mechanical model used, including the constitutive law, is described in detail. This description
includes numerical tests to validate the material law. The description of the coupling model follows.
Finally, results of the contraction of a simple geometry and a complete heart model are presented
and compared qualitatively with general physiological motions.
The novelties of this work include: (1) proving the efciency of a highly parallel-coupled elec-
tromechanical model of the heart up to 500 cores, using a realistic geometry; (2) the use of a
constitutive law that is based on a general and exible framework and that reects better the behav-
ior in compression of the myocardium, and the characterization of this law; and (3) the study and
characterization of the damping properties of the heart.
2. PHYSIOLOGICAL MODELS
At the organ level, Figure 1 describes the heart as a coupled system: the electrical action poten-
tial propagation activates the mechanical deformation that, in turn, produces the pumping function
against the blood. Coupling is in both ways because blood exerts a pressure on the endocardium,
and tissue deformation acts upon electrical propagation. In the next sections, electrophysiology and
mechanical parts are described separately, and a nal section describes the coupling strategy.
2.1. Electrophysiology: action potential propagation
The electrophysiology potential (.
i
, t ) is modeled using a propagation differential equation
combined with different microscopic models projected to the large scales

,
d

dt
=
d
d.
i
_
D
ij
d

d.
j
_
1(

). (1)
Latin subscripts and greek subscripts count the space dimension of the problem and the number
of activation potentials involved, respectively, being = 1 for monodomain models and = 1, 2
for bidomain ones. In this paper, we will focus on monodomain approaches. The diffusion term is
determined by the diffusion tensor D
ij
, which describes the cardiac tissue bers orientation. The
nonlinear term 1(

) is the term of the total membrane ionic current, modeled in different ways
of different complexities. All along this paper, we will use alternatively FitzHughNagumo (FHN)

Einstein convention on repeated indexes is used for the general eld equations. For the remaining model equations, bold
letters are used for tensors
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
Figure 1. The heart as a physical system.
[15] or FentonKarma (FK) [16] models. For a thorough discussion, see [12]. The weak form of
Equation (1) is: nd V such that V[ W; the following equation veries:
_

[
d
dt
dV =
_
@
[
D
ij
C
m
S
v
d
d.
j
n
j
dS
_

D
ij
C
m
S
v
d[
d.
i
d
d.
j
dV
_

[1
ion
dV (2)
where C
m
and S
v
are some model constants and 1
ion
is the total membrane ionic current.
2.2. Mechanical deformation: the cardiac muscle
2.2.1. Governing equations. Let C
o
be a xed reference conguration of a body and C the
deformed one. The position vector of a particle of that body is expressed by X
I
in the rst con-
guration and by .
i
in the second one. Note that capital letters and subscripts are used for scalars
or tensors dened in the reference conguration, and lowercase letters and subscripts, when dened
in the deformed conguration. The relation between both measures is given by the deformation
gradient tensor:
J
iJ
=
d.
i
dX
J
=
iJ

du
i
dX
J
(3)
where the displacement is u
i
= .
i
X
I
. The local form of the linear momentum balance, to be
solved using a total Lagrangian nite element formulation in the conguration C
o
, is written as
follows:
j
o
d
2
u
i
dt
2
=
d1
iJ
dX
J
j
o
T
i
, (4)
which in its weak form is
_

o
j
o
d
2
u
i
dt
2
=
_

o
d
d.
J
1
iJ

_
@
o
N
J
1
iJ

o
j
o
T
i
(5)
where is a test function, j
o
is the initial density of the body, N
J
is the exterior normal of C
o
on dC
o
, T
i
is the body force and 1
iJ
is the rst PiolaKirchhoff stress tensor. After discretization,
Equation (5) can be expressed in matrix form. Space is discretized using the nite element method
(FEM), using the usual compact support space. Time is discretized with nite differences method
(FD) using a 0-generalized scheme. Alya includes explicit and implicit formulations, but we will
focus in the explicit ones in this paper.
Timespace discretization of this kind of equations leads to a matrix equation of the form
M R u C(u) P u K(u)u =R (6)
where R is the residual forces vector, M is the mass matrix and K(u) is the stiffness matrix, which
typically depends on the unknown. The second term, proportional to the speed, is the Rayleigh
damping, where C(u) is the damping matrix. This matrix is added to damp high frequencies and
avoid unphysical oscillations. Usually, its is written as C(u) = M K(u), where parameters
and are computed depending on the frequency range to be damped [17]. After some numerical
experiments, we have observed that the best choice is to take = 0 and = c o, where c is a
constant between 2 and 10. The frequency o is computed by running a problem without damping,
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
P. LAFORTUNE ET AL.
plotting the displacement for one point in the domain and analyzing its time series. It is worth to
remark that when =0, the matrix damping becomes independent of the unknown, which certainly
simplies its computation.
2.2.2. Constitutive model. Heart muscle is mainly composed of a thick layer of myocytes (parallels
muscle bers) assuming the pumping function of the heart called the myocardium. This muscle is
often simulated as an incompressible material. However, in vivo measurements proved that vessels
perfusing the myocardium make the structure compressible ([18, 19]). The myocardium presents a
particular orthotropic-layered architecture. Muscle bers form layers (three to four bers in thick-
ness, connected by endomysial collagen), and the sheets are linked one to the other with perimysial
collagen and sparsely distributed muscle bridges. Figure 2 illustrates this structure. The ber ori-
entation has a complex distribution that varies in the transmural direction and from base to apex
([20, 21]).
Orthotropy have been proven to be important by the shear tests performed by Dokos [22]. How-
ever, common imaging techniques only provide the information on the orientation of the bers.
For that reason, transversally isotropic models based on biaxial test are commonly used when the
geometry is built from anatomical data ([2325]). Such a model is used in this study.
The biaxial test is a 2D traction test performed on excised slices of muscle. Although this tech-
nique allows a general and direct characterization, the preparation of the sample alter the complex
3D structure of the muscle. For a review of existing constitutive models, see [26] and [27]. To depict
the constitutive models (leaving aside indicial notation), the Cauchy stress is used, which is related
to the rst PiolaKirchoff stress as follows:
=J
1
PF
T
(7)
where J is the determinant of the deformation gradient F. Suppose a compressible, homogeneous,
hyperelastic material under quasi-static isothermal conditions. The constitutive behavior of such
material is expressed in terms of a strain energy function W(b) that relates the Cauchy stress to
the right CauchyGreen deformation b :
=2J
1
dW(b)
db
b (8)
where b =FF
T
. It can be useful, to develop general expressions of stress, to write this equation in
terms of the strain invariants of b. The three rst invariants are:
1
1
=trb, 1
2
=
_
(trb)
2
trb
2
_
, 1
3
=det(b) (9)
Let us now consider the orthonormal vectors dening the local coordinate system associated with
the myocardium structure (Figure 2). Then, f
0
is the mean ber orientation, s
0
is the sheet axis that
lies in plane perpendicular to the bers and n
0
is the normal to the plane formed by the two other
Figure 2. Left: Scanning electron microscopy of the myocardium sectioned perpendicular to the bers
(from LeGrice et al. [20]). Right: Local coordinate system. Fibers, muscle bridges and perimysial collagen
are visible.
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
vectors. If such a material is modelized as a transversally isotropic one, the fourth strain invariant
will be also needed:
1
4
=f
0
bf
0
(10)
The invariants are of particular interest when developing constitutive laws because they remain
unchanged when expressed in a different basis. Using the chain rule,
dW(b)
db
=

a
dW
d1
a
d1
a
db
(11)
and the denitions of the strain invariant derivatives, Equation (8) is reformulated as follows:
=2J
1
W
1
b 2J
1
W
2
_
1
1
b b
2
_
2JW
3
I 2J
1
W
4
N
f
N
f (12)
where the notation W
i
= dW,d1
a
is used and
N
f = Ff
0
. This notation is used to underline the
fact that
N
f is not a unitary vector. See [28] for a more detailed derivation of those equations. The
last equation is a general framework valid for any hyperelastic material having one preferred ori-
entation. The form of W is to be determined. Holzapfel and Ogden [29] suggested a form of W
based on physiological considerations. The invariant 1
1
represents the non-collagenous material,
whereas invariant 1
4
represents the stiffness of the muscle bers. These authors proposed a more
complete orthotropic law. In this paper, a reduced transversally isotropic version is presented. We
then propose an energy function given by
W =
a
2b
e
b.I
1
3/

a
2
(1
1
3)
a
f
2b
f
_
e
b
f
.I
4
1/
2
1
_

1
2
(J 1)
2
(13)
This equation is slightly different from that of [29]. First, a volumetric energy term function of J
have been added to make the material compressible. Then, the term
a
2
(1
1
3) is necessary to
respect the initial conditions (zero energy and stress in the reference conguration). This modica-
tion is required owing to the fact that the original incompressible law has a term I that makes
the energy equal to zero when no deformations are present. Finally, introducing the expression of
the derivatives of the energy in the general stress equation gives the following:
J =
_
ae
b.I
1
3/
a
_
b 2a
f
(1
4
1)e
b
f
.I
4
1/
2
N
f
N
f 1(J 1)I (14)
where a, b, a
f
, b
f
are parameters to be determined experimentally. The value of 1 is related to the
compressibility and a value of 100 kPa is used as a starting point [30].
To nd the constants and validate the constitutive model, the biaxial test conducted by Yin and
coworkers [31] is reproduced. A thin square sheet of tissue of about 4 4 0.15 cm
3
is extracted
from the myocardium. The traction axis are aligned with the ber (f ) and cross-ber (or sheet)
direction (s). The third direction perpendicular to this plane is called the normal direction (n)
(see Figure 2). First, the stress equation (14) is solved numerically using the gradient deforma-
tion tensor F =diag
_
z
f
z
s
z
n
_
. The subscripts refer to the directions described earlier, and z is the
stretch in the indicated direction. z
f
and z
s
are given (supposing a displacement-controlled test)
and z
n
is found solving the following:
o
n
=
_
ae
.b.I
1
3//
a
_
z
2
n
1(J 1) =0 (15)
because no constraint is applied in this direction. This equation is solved numerically using the
bisection method. The stress in the two other directions are then calculated with
o
f
=
_
ae
.b.I
1
3//
a
_
z
2
f
2a
f
(1
4
1)e
b
f
.I
4
1/
2
z
2
f
1(J 1) (16)
o
s
=
_
ae
.b.I
1
3//
a
_
z
2
s
1(J 1) (17)
DOI: 10.1002/cnm
Int. J. Numer. Meth. Biomed. Engng. (2012) Copyright 2012 John Wiley & Sons, Ltd.
P. LAFORTUNE ET AL.
where the constant values are a = 4.3 kPa, b = 9.7, a
f
= 1.69 kPa, b
f
= 15.78. In addition to
the numerical solution used to nd the parameters, the model is tested reproducing the biaxial tests
with a simple FE model containing 50 linear hexahedral elements. The results of those tests, along
with experimental data, are shown on Figure 3.
During systole, the bers contract and the material is in compression. This behavior is hardly
never studied, neither numerically nor experimentally. It is still a challenge to characterize a material
in compression, even more if it is a complex biological structure. This difculty probably explains
the lack of data and numerical test present in the literature. Nevertheless, we can have a general
idea using the solution of a uniaxial traction test. The deformation gradient tensor is, in this case,
simplied to F = diag
_
z
f
z
s
z
s
_
, where is the ber direction aligned with the traction, and s
the two other directions forming a perpendicular plan to . Once again, we suppose here that z
f
is the displacement of the traction machine, and we nd z
s
by solving the stress equation in the s
directions:
o
s
=
_
ae
.b.I
1
3//
a
_
(z
2
s
) 1(J 1) =0 (18)
Then, the stress in the direction can be computed as follows:
o
f
=
_
ae
.b.I
1
3//
a
_ _
z
2
f
_
2a
f
(1
4
1)e
b
f
.I
4
1/
2
z
2
f
1(J 1) (19)
Figure 4 shows the result of the uniaxial traction test. First, the behavior of the material with the
parameters previously found is plotted. A very low stiffness is observed in compression, owing to
the nature of the exponential functions used in the constitutive equation. To stiffen the material in
compression, the constant a
f
is increased to a
comp
f
= 6.5 a
f
when compression is detected in the
direction of the bers (z
f
< 1.0). The result using the modied coefcient is plotted on the same
graphic. Even if only acting on the term that controlled the bers, this modication must be seen as
acting on the whole structure. As mentioned by Holzapfel and Ogden [29], the bers are most likely
to buckle when compressed, and they have a minor contribution.
Unfortunately, to the authors knowledge, there is no experimental data available for the
myocardium in compression. Nevertheless, numerical results published by Hunter [32] are also
presented.
22
20
18
16
14
12
10
8
6
4
2
0
0 0.05 0.1 0.15
22
20
18
16
14
12
10
8
6
4
2
0
S
s
s
(
k
P
a
)
0 0.05 0.1 0.15
E
ff
E
ss
S
f
f
(
k
P
a
)
Figure 3. Biaxial test. Experimental values extracted from Yin et al. ([31]) for strain ratio 1

,1
ss
of 0.48
(_), 1.02 () and 2.05 (.). The full curves represent the solution of the analytical solution and the dash
curves the values obtained with the nite element method.
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
50
40
30
20
10
0
-10
-20
-30
-40
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2 1.25

f

(
k
P
a
)

original
a
f
modified in compression
Hunter et al.

f
Figure 4. Solution of the stress equations for a uniaxial tensioncompression test. The response given by
the original parameters and when the value of a
f
is modied are shown. Values published by Hunter et al.
[32] are also shown.
2.3. Electromechanical coupling: active tension development
So far, the electrical and mechanical models are described independently and assessed through
individual tests. The excitationcontraction (EC) coupling is the phenomenon by which the bers
contract after a wave of electrical activation propagates through the myocardium. EC coupling is
the result of a series of complex reactions that can be summarized as follows. Ca
2C
enters the
cells during the plateau phase of the action potential caused by an increase of the permeability of
the cell membrane (sarcolemma). This extracellular Ca
2C
triggers the release of a larger amount
of intracellular Ca
2C
from the sarcoplasmic reticulum. Ca
2C
then binds to the troponin C, which
interacts in turn with tropomyosin, allowing the myosin to bind with the actin laments. Myosin
head pull the actin lament toward the center of the sarcomere, producing the contractile force. To
simplify, we can say that the free intracellular concentration of Ca
2C
mediates the contraction of the
myocardium, explaining why many models are based on this quantity.
A large amount of EC coupling models have been developed (see [33] for a comprehensive
review). The model used in this study is a one-way coupling, in which the displacements have
no inuence on electrophysiology. Also, it is assumed that the active stress is produced only in the
direction of the ber. The active contribution is simply summed to passive stress. Then, the total
Cauchy stress is expressed by [34]:
=
pas
o
act
_
z,
_
Ca
2C
__
f f (20)
where
pas
is the passive stress dened by Equation (14), and f is a unit vector aligned with the
ber in the current conguration. If z
f
is the muscle ber stretch and f
0
is the ber in the reference
conguration, it can be expressed by z
f
f =Ff
0
. Hunter and co-workers [32] suggested a simple
model based on steady state data at different constant levels of activation and extension ratios to
calculate the active stress (see also [35]):
o
act
=
_
Ca
2C
_
n
_
Ca
2C
_
n
C
n
50
o
max
_
1 j
_
z
f
1
__
(21)
where C
50
is the value of the intracellular calcium concentration for 50% of o
max
, n is a coefcient
controlling the shape of the curve and o
max
is the maximum isometric active tensile stress developed
at z
f
=1. Experiments show that o
max
is a linear function of z with j being the non-dimensional
slope dT,dz. This parameter represents the dependence of the force on the sarcomere length. The
force increases with the stretch of the ber, because the number of overlapping myosin/actin bending
site increases. The uctuation of the free calcium ions concentration follows this relation [36]:
_
Ca
2C
_
(t
loc
) =
_
Ca
2C
_
max
(t
loc
,t
Ca
) exp (1 t
loc
,t
Ca
) (22)
DOI: 10.1002/cnm
Int. J. Numer. Meth. Biomed. Engng. (2012) Copyright 2012 John Wiley & Sons, Ltd.
P. LAFORTUNE ET AL.
where t
Ca
is a parameter and t
loc
is the local activation time of a certain integration point of the
mesh. The t
loc
is initiate to 0.0 s at the beginning of the simulation for the entire mesh. When the
membrane potential overcomes a threshold value (xed to 50 mV), EC coupling is triggered and
t
loc
starts running at this specic location. Then,
_
Ca
2C
_
(and consequently o
act
) also vary.
Figure 5 shows the relation between the local activation time and the action potential, the con-
centration of intracellular Ca
2C
, and the normalized active force produced. Parameters values are
C
50
=0.5 pm, o
max
=100 kPa, n =3, j =2.5,
_
Ca
2C
_
max
=1.0 pm and t
Ca
=0.06 s.
3. NUMERICAL RESULTS
The coupled electromechanical model is implemented in Alya, the BSC in-house computational
mechanics code. For both examples, the summarizing implementation facts are as follows:
v Two coupled problems (electrical activation potential and mechanical deformation) are solved
in the same parallel code, Alya.
v A monodomain model is used for describing the electrical behavior of the system.
v Problems are solved in a single mesh, composed of linear tetrahedral elements.
v An explicit time scheme is used.
v A fractional step coupled strategy is used (mechanical and electrical problems are solved
sequentially at each time step).
v For explicit schemes, no special linearization is required and no matrix is stored. As the time
step is small, we have observed that no internal supplementary iterations are needed.
To assess the computational model presented here, two examples of the coupled electromechani-
cal problem are shown. The rst one is a simple geometry used to illustrate the basis of the problem,
including the action of the Rayleigh damping. The second one is a much larger one obtained from a
rabbit heart mesh run in BSCs Marenostrum supercomputer using up to 500 processors.
3.1. Small piece of cardiac tissue
This example is an electromechanical simulation of a small block of cardiac tissue used to cal-
ibrate the damping parameters. It is made of 2250 tetrahedral elements (576 nodes) and mea-
sures (0.25 0.25 0.75) cm
3
. The electrical propagation is solved using a monodomain model
with bers in the z-direction and a FitzHughNagumo ionic current model. In this problem,
the electrical conductivity is isotropic, equal to 0.012 ms cm
1
and the membrane capacitance is
C
m
=0.001 pFcm
2
. The constitutive and coupling models are those described earlier.
-110
-75
-50
-25
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
0.2
0.4
0.6
0.8
1
1.2
E

(
m
V
)
[
C
a
+
+
]

(

M
)
time (sec)
[Ca
++
]
Membrane potential
Normalized active force
Figure 5. Time scale comparison of the normalized active force, the membrane potential and the
_
Ca
2C
_
.
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
The problem is solved explicitly, using appropriate schemes for both problems. In the cases
presented, forward Euler for the electric one (rst order time derivative) and Newmark for the
mechanical one (second order time derivative) were used. Time step is computed from the most
restrictive problem, which turns out to be the mechanical one, whose time step is computed from
the speed of sound and the size of the elements. Figure 6 shows a snapshot of the activation potential
wave contracting the tissue.
Figure 7 plots, for one point in the middle of the domain, the time evolution of the activation
potential together with the displacement at the same point. The gure clearly demonstrates the
action of the Rayleigh damping, whose coefcients are =1000 and = 0. It is worth to remark
that the results are not very sensitive to changes in the mass matrix coefcient , which can vary
between 1000 and 5000 with very similar damping action. To compute it, we analyze the non-
damped problem (blue line) to obtain an estimation of the high frequency to be damped. Then,
is chosen to be of the same order of magnitude than the inverse of this frequency. For this set of
problem parameters, Figure 7 shows that the peak of the activation potential is slightly ahead of the
peak of the displacement, concurrently with Figure 5.
The boundary conditions for this problem are set by xing the displacement at the z=0 plane,
whereas the rest of the geometry is free to move in all directions. Once the activation potential wave
is gone, the tissue returns slowly to its original resting position. In the case of the heart, the interior
pressure of the ventricles helps it to come back, making this transition relatively faster. Observe
also the small displacement overshoot before the activation potential wave. We believe that this
is due to the sound wave in the material (recall that it is compressible) that somewhat warns the
Figure 6. Small piece of contracting cardiac tissue. Color scale represents the activation potential propaga-
tion. Longitudinal direction is aligned with the z-axis
-0.08
-0.06
-0.04
-0.02
0
0 0.1 0.2 0.3 0.4
Z
-
D
i
s
p
l
a
c
e
m
e
n
t
(
c
m
)
Rayleigh damping
No Rayleigh damping
-0.078
-0.077
-0.076
-0.075
-0.074
0.05 0.06 0.07 0.08 0.09 0.1
Z
-
D
i
s
p
l
a
c
e
m
e
n
t
(
c
m
)
Rayleigh dumping
No Rayleigh dumping
Time(sec) Time(sec)
-0.073 0.02
Figure 7. The Rayleigh damping action on the mechanical contraction time evolution for a point in the
middle of the domain (the undamped case is the dashed line). Right: close-up.
DOI: 10.1002/cnm
Int. J. Numer. Meth. Biomed. Engng. (2012) Copyright 2012 John Wiley & Sons, Ltd.
P. LAFORTUNE ET AL.
material points slightly before the activation potential wave arrives. This is a normal behavior of
compressible ows or deformations.
3.2. Two ventricles of a rabbit heart
To test the feasibility of the models presented, a contraction of the left and right ventricles of a
rabbit heart is simulated. The mesh was kindly provided by M. Bernabeu and M. Bishop from the
Computing Laboratory of Oxford University [37]. The geometry comes from high-resolution mag-
netic resonance imaging (raw data resolution: 43 43 36 pm). The detailed reconstruction
procedure can be found in [38]. The mesh consist of 432 000 linear tetrahedral elements and also
provides the ber architecture based on a theoretical model. Again here, the constitutive and cou-
pling models are those described earlier. The electrical conductivity is anisotropic. A value of
0.0025 ms cm
1
is used in the ber direction, and 0.000833 ms cm
1
in the other directions. The
membrane capacitance is C
m
= 0.001 pFcm
2
. The depolarization is initiated on the superior
section of the interventricular septum. To have a rst idea on how the model works, this simulation
is simplied in terms of boundary conditions: no pressure is imposed in the endocardium, and the
heart is free to relax after the contraction. This gives a good qualitative idea of the contraction pat-
tern for one cycle. The results shown here were run in 200 processors at a total turnaround time of
10 h approximately.
Figure 8 reproduced the deformed heart at different times of the simulation. Figures 9 and 10
show respectively the temporal evolution of a longitudinal and transversal section of the heart at
equatorial level.
This simulation lacks the effect of the intra ventricular pressure. However, we believe that it
should be valid for a qualitative behavior of the ventricles action. In fact, the ejection-like defor-
mation is clearly seen, and some features can be compared with the basic clinical observations of
real hearts. For instance, the longitudinal displacement evolution during systole is expected to be
in the apical direction [39]. Figure 9 shows that the model here presented reproduces this motion:
during the contraction, the apex is moving downward. It is also observed that the wall of the left
ventricle became thicker during systole, as shown in Figure 10. Finally, the displacement vectors at
equatorial level are shown in Figure 11. The direction of the displacement reveals the complexity of
the myocardium motion.
Figure 8. Deformed mesh of the heart at time 0, 0.04, 0.172, 0.26, 0.348 and 0.44 s. The contour represents
the membrane potential (mV).
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
Figure 9. Longitudinal deformation (cm) at time 0, 0.136 and 0.44 s.
Figure 10. Norm of the deformation (cm) at time 0, 0.136 and 0.44 s.
3.3. Parallel efciency
A key feature of our computational model is its parallel efciency. This allows either to solve very
large problems in a reasonably turnaround time or medium size problems very quickly. It has been
assessed using the bi-ventricular model of the rabbit heart presented on section 3.2. In this case, the
turnaround wall clock time for running one heart cycle using 200 processors was of around 10 h.
Parallel efciency is measured by analyzing scalability. Strong scalability means how faster a
problem of xed size is solved when the number of processors increase progressively. This scalabil-
ity measure is linear when the speedup increases linearly with the number of processors involved.
To measure it, we have used this problem as a benchmark, with runs of 50, 100, 300 and 500 pro-
cessors, taking the smallest one as the reference value. Then, perfect linear scalability means that
Figure 11. Displacement vectors at a transversal section of the heart at equatorial level.
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
P. LAFORTUNE ET AL.
0
100
200
300
400
500
0 100 200 300 400 500
900 107000 215000 322500 430000
S
p
e
e
d

U
p
Elements/CPU
simulations
ideal
CPU
Figure 12. Scalability up to 500 cores on Marenostrum supercomputer.
the 500 run goes 10 times faster than the reference. The tests were carried out in Marenostrum
supercomputer. Marenostrum is a distributed memory cluster made of 10 240 IBM PPC processors.
Figure 12 shows the strong scalability as the number of cores used versus the speed up, where the
red dashed line is the ideal. The top horizontal coordinate label is the mean number of elements per
core. In this case and being the mesh rather coarse, the scalability for 500 cores is 80% of the ideal,
owing to the very low amount of elements (900) per core and the proportionally large amount of
communications per iteration.
4. DISCUSSION AND CONCLUSIONS
In this paper, we have presented a parallel computational model of the coupled electromechanical
behavior of the heart. The main conclusion is that, this model is capable of running simulations of
very large meshes in hours in hundreds of processors. Succeedingly, we discuss each of the main
points of the model.
4.1. Constitutive modeling
Some modications have been made to the model recently published by Holzapfel and Odgen [29].
Compressibility is taken into account, and a transversally isotropic version is created with a new
set of parameters. This version of the law is useful when the bers eld is only characterized by its
direction, with no information on the sheet direction. This is the most likely situation when data is
provided by medical imaging techniques. The unidimensional traction/compression numerical test
performed underlines the fact the constitutive models not always behave symmetrically if used in
tension or compression. This aspect is very important because the muscle is contracted when eject-
ing the blood. Unfortunately, very few publications study this aspect. Note also the critical lack of
clinical data in compression to validate the constitutive models.
4.2. Coupling
The model presented in this study still must be calibrated to better reproduce the timing commonly
seen between the action potential and the force (Figure 5). According to most authors, the peak of
the force should coincide with the completion of repolarization cycle (see, e.g. [40]). This could be
set by adding parameters to the equation, but the present form is judged sufcient for this study.
Even if this simple relation is based on steady-state properties and will not be able to reproduce
DOI: 10.1002/cnm
Int. J. Numer. Meth. Biomed. Engng. (2012) Copyright 2012 John Wiley & Sons, Ltd.
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
the realistic velocity-dependent behavior of the myocardium, it is commonly used in nite element
model, and it is suitable to simulate the healthy beating heart. If more complex behaviors have to be
simulated, a full dynamic model is necessary ([36, 40]).
Coupling from the mechanical to the electrical part is caused by the deformation of the physical
space, producing a difference in the travel time of the electrical propagation. In this version of the
model, a one-way coupling is used (from electrical to mechanical), and this effect is not considered.
It will be taken into account in future works.
4.3. Numerical and implementation issues
All along this paper, we have used explicit schemes for both the mechanical and electrical prob-
lems (see [17] for a detailed description of the explicit schemes). Although they are programmed
in Alya, owing to the small physical time scales and the high non-linearities of both problems, we
have preferred the all-explicit strategy. From our experience, the staggered loosely coupled strat-
egy is specially well-suited for large scale coupled problems. It avoids storing a large matrix with
a block for each of the problem and coupling off-diagonal blocks. Besides, it allows the use of the
most appropriate numerical solution scheme for each problem block. Regarding the computational
efciency, it is worth to note that we are far from the optimal. Parallel efciency is correct, but
sequential efciency must be highly improved. Non-linear solid mechanics of complex materials is
expensive with a large amount of inner nodal or Gauss point loops. Then, one must be very careful
on how these loops are written to stay efcient. As this model is coming out from the initial devel-
opment stage, a hard code cleaning must be faced. Based on previous experience, we expect around
an order of magnitude of improvement.
Another important issue to consider is how much is the relative computational effort of electro-
physiolgy and solid mechanics is, especially for large problems. Some researchers propose to use
a ne low-order unstructured mesh for solving electrophysiology and a much coarser high-order
structured one (but not higher than Q2 or Q3) for solid mechanics. In this case, it is usually reported
that electrophysiology takes more CPU time than solid mechanics (see reviews [2, 41, 42]). On
the other hand, we propose to use the same unstructured mesh for both problems, which leads to
the opposite ratio. With the models we used for the rabbit ventricles, the gures are around 30%
electrophysiology and 70% solid mechanics.
To obtain the results presented here, we have previously performed a set of runs to calibrate the
parameters, which can be rather hard. However, some of this parameters have not been explored,
but xed from the beginning; notably, the bers eld and the electrical initial conditions. In the
Oxford rabbit model, the bers eld comes given based on certain semi-empirical model. On the
other hand, we have xed the initial activation potential to two localized impulses in the basal
region. In this paper, we have preferred to focus on the model itself than studying more advanced
physiological aspects.
4.4. Future lines
Most of the future lines are outlined in the precedent paragraphs. Better and more sophisticate phys-
iological models must be included, in the constitutive law, the coupling, the bers eld and the
initial and boundary conditions. Some of these improvements have already been carried out, but are
not the object of this paper. Some others depend on the state of collaboration with bioengineers and
medical doctors. For instance, the model is being used to study infarction and resynchronization.
As the model allows to run simulations with high time and space resolution, it is very appealing for
direct validation against experimental measurements on real cases obtained from medical images,
such as ejection rate, torsion, wall thickening, deformation time sequence and so on.
ACKNOWLEDGEMENTS
This work is made possible, thanks to the close collaboration with bio-engineers and medical doctors.
Among them are D. Gil and J. Garcia-Barns (Computer Vision Center CVC-UAB, Spain), F. Carreras
(Htal. de Sant Pau, Spain), M. Ballester Rods (U. Lleida, Spain), A. Jerusalm and D. Tjahjanto (IMDEA
Copyright 2012 John Wiley & Sons, Ltd.
DOI: 10.1002/cnm
Int. J. Numer. Meth. Biomed. Engng. (2012)
P. LAFORTUNE ET AL.
Materials, Spain), D. Auger and T. Franz (Univ. of Cape Town, South Africa) and M. Bernabeu and M.
Bishop (Oxford University, UK).
REFERENCES
1. Nordsletten DA, Niederer SA, Nash MP, Hunter PJ, Smith NP. Coupling multi-physics models to cardiac mechanics.
Progress in Biophysics and Molecular Biology 2011; 104(1-3):7788.
2. Trayanova N, Winslow R. Whole heart modeling. Applications to cardiac electrophysiology and electromechanics.
Circulation Research 2011; 108(4):113128.
3. Nickerson D, Smith N, Hunter PJ. New developments in a strongly coupled cardiac electromechanical model.
Europace 2005; 7:118127.
4. Sainte-Marie J, Chapelle D, Cimrman R, Sorine M. Modeling and estimation of the cardiac electromechanical
activity. Computers and Structures 2006; 84:17431759.
5. Stevens C, Remme E, LeGrice I, Hunter P. Ventricular mechanics in diastole: Material parameter sensitivity. Journal
of Biomechanics 2003; 36(5):737748.
6. Gktepe S, Kuhl E. Electromechanics of the heart: A unied approach to the strongly coupled excitation-contraction
problem. Computational Mechanics 2010; 45:227243.
7. Kerckhoffs RCP, Neal M, Gu Q, Bassingthwaighte JB, Omens J, McCulloch A. Coupling of a 3D nite element
model of cardiac ventricular mechanics to lumped systems models of the systemic and pulmonic circulation. Annals
of Biomedical Engineering 2007; 35:118.
8. Nobile F, Quarteroni A, Ruiz-Baier R. An active strain electromechanical model for cardiac tissue. International
Journal for Numerical Methods in Biomedical Engineering 2011. DOI: 10.1002/cnm.1468.
9. Gurev V, Constantino J, Rice JJ, Trayanova NA. Distribution of electromechanical delay in the heart: Insights from
a three-dimensional electromechanical model. Biophysical Journal 2010; 99(3):745754.
10. Hosoi A, Washio T, Okada J, Kadooka Y, Nakajima K, Hisada T. A multi-scale heart simulation on massively par-
allel computers. In Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing,
Networking, Storage and Analysis, SC 10. IEEE Computer Society: Washington, DC, USA, 2010; 111.
11. Reumann M, Fitch BG, Rayshubskiy A. et al. Strong scaling and speedup to 16,384 processors in cardiac electro
mechanical simulations. Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International
Conference of the IEEE, Minneapolis, Minnesota, USA, september 2009; 27952798.
12. Vzquez M, Ars R, Houzeaux G. et al.A massively parallel computational electrophysiology model of the heart.
International Journal for Numerical Methods in Biomedical Engineering 2011. DOI: 10.1002/cnm.1443.
13. Houzeaux G, Aubry R, Vzquez M. Extension of fractional step techniques for incompressible ows: The
preconditioned orthomin(1) for the pressure schur complement. Computers and Fluids 2011; 44:297313.
14. Available from: http://www.bsc.es..
15. FitzHugh RA. Impulses and physiological states in theoretical models of nerve membrane. Journal of Biophysics
1961; 1:445466.
16. Fenton FH, Cherry EM, Karma A, Rappel W. Modeling wave propagation in realistic heart geometries using the
phase-eld method. Chaos 2005; 15(013502).
17. Belytschko T, Liu WK, Moran B. Nonlinear Finite Elements for Continua and Structures. John Wiley and Sons:
Chichester, 2000.
18. Yin FC, Chan CC, Judd RM. Compressibility of perfused passive myocardium. American Journal of Physiology -
Heart and Circulatory Physiology 1996; 271(H1864-H1870).
19. Moore CC, McVeigh ER, Elias AZ. Noninvasive measurement of three-dimensional myocardial deformation with
tagged magnetic resonance imaging during graded local ischemia. Journal of Cardiovascular Magnetic Resonance
1999; 1(3):207222.
20. Legrice I, Hunter P, Young A, Smaill B. The architecture of the heart: A databased model. Philosophical
Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences 2001;
359(1783):12171232.
21. Streeter DD, Powers WE, Ross MA, Torrent-Guasp F. Three-dimensional ber orientation in the mammalian left
ventricular wall. In Cardiovascular System Dynamics, Barn J, Noordegraf A, Raines J (eds). MIT Press: Cambridge,
MA, 1978; 73.
22. Dokos S, Smaill BH, Young AA, LeGrice IJ. Shear properties of passive ventricular myocardium. American Journal
of Physiology - Heart and Circulatory Physiology 2002; 283(6):H26502659.
23. Watanabe H, Sugiura S, Kafuku H, Hisada T. Multiphysics simulation of left ventricular lling dynamics using
uid-structure interaction nite element method. Biophysical Journal 2004; 87(3):20742085.
24. Veress AI, Segars WP, Weiss JA, Tsui BMW, Gullberg GT. Normal and pathological ncat image and phantom data
based on physiologically realistic left ventricle nite-element models. IEEE Transactions on Medical Imaging 2006;
25(12):16041616.
25. Dorri F, Niederer PF, Lunkenheimer PP. A nite element model of the human left ventricular systole. Computer
Methods in Biomechanics and Biomedical Engineering 2006; 9(5):319341.
26. Kerckhoffs RCP, Healy SH, Usyk TP, McCulloch AD. Computational methods for cardiac electromechanics.
Proceedings of the IEEE April 2006; 94(4):769783.
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)
COUPLED PARALLEL ELECTROMECHANICAL MODEL OF THE HEART
27. Costa KD, Holmes JW, McCulloch AD. Modelling cardiac mechanical properties in three dimensions. Philosophi-
cal transactions of the Royal Society of London. Series A: Mathematical, physical and engineering sciences 2001;
359(1783):12331250.
28. Holzapfel GA. Nonlinear Solid Mechanics: A Continuum Approach for Engineering. John Wiley and Sons:
Chichester, 2000.
29. Holzapfel GA, Ogden RW. Constitutive modelling of passive myocardium: A structurally based framework for mate-
rial characterization. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering
Sciences 2009; 367(1902):34453475.
30. Usyk TP, Mazhari R, McCulloch AD. Effect of laminar orthotropic myober architecture on regional stress and
strain in the canine left ventricle. Journal of Elasticity 2000; 61:143164.
31. Yin FC, Strumpf RK, Chew PH, Zeger SL. Quantication of the mechanical properties of noncontracting canine
myocardium under simultaneous biaxial loading. Journal of Biomechanics 1987; 20(6):577589.
32. Hunter PJ, Nash MP, Sands GB. Computational Electromechanics of the Heart, Computational Biology of the Heart.
John Wiley and Sons: London, 1997. 346407.
33. Rice JJ, de Tombe PP. Approaches to modeling crossbridges and calcium-dependent activation in cardiac muscle.
Progress in Biophysics and Molecular Biology 2004; 85(2-3):179195.
34. Humphrey JD. Cardiovascular Solid Mechanics. Cells, Tissues, and Organs. Springer: New York, NY, 2001.
35. Holzapfel GA. Computational Biomechanics of Soft Biological Tissue, Volume 2 of Encyclopedia Of Computational
Mechanics. John Wiley & Sons: Chichester, 2004. Chapter 18, 604635.
36. Hunter PJ, McCulloch AD, ter Keurs HEDJ. Modelling the mechanical properties of cardiac muscle. Progress in
Biophysics and Molecular Biology 1998; 69:289331.
37. Computing Laboratory. Oxford University. Available from: http://www.cs.ox.ac.uk/chaste.
38. Bishop MJ, Plank G, Burton RAB, Schneider JE, Gavaghan DJ, Grau V, Kohl P. Development of an anatomically
detailed MRI-derived rabbit ventricular model and assessment of its impact on simulations of electrophysiological
function. American Journal of Physiology - Heart and Circulatory Physiology 2010; 298(2):H699H718.
39. Moore CC, Lugo-Olivieri CH, McVeigh ER, Zerhouni EA. Three-dimensional systolic strain patterns in the normal
human left ventricle: Characterization with tagged mr imaging. Radiology 2000; 214(2):453466.
40. Rice JJ, Wang F, Bers DM, de Tombe PP. Approximate model of cooperative activation and crossbridge cycling in
cardiac muscle using ordinary differential equations. Biophysical journal 2008; 95(5):23682390.
41. Trayanova N, Eason J, Aguel F. Computer simulations of cardiac debrillation: A look inside the heart. Computing
and Visualization in Science 2002; 4:259270.
42. Trayanova N, Rice JJ. Cardiac electromechanical models: From cell to organ. Frontiers in Physiology 2011; 2(0).
DOI: 10.1002/cnm
Copyright 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Biomed. Engng. (2012)

You might also like