IX CONGRESO COLOMBIANO DE METODOS NUMERICOS: Simulación en Ciencias y Aplicaciones Industriales IX

CCMN 2013, Agosto. 21-23, 2013, UAO Cali, Colombia

Continuous-time Feedback Control of Alternans in a Model of

Purkinje Fibers
Alejandro Garzón1 , Roman O. Grigoriev2 , and Flavio H. Fenton2
1
School of Mathematics,Universidad Sergio Arboleda, Bogota, Colombia
2
School of Physics, Georgia Institute of Technology, Atlanta, Georgia, USA
e-mail: [email protected], [email protected], [email protected]

Abstract Alternans – an arrhythmic response of cardiac tissue to periodic pacing by pulses of electric current – often
serves as a precursor to a more dangerous, and potentially lethal, state of fibrillation. Therefore, suppression of alternans may
be seen as a plausible way to prevent fibrillation. We use a novel partial differential equation model of Purkinje fibers (one-
dimensional cardiac tissue) as a test-bed for comparing impulsive and continuous-time feedback control for suppression of
alternans. In the impulsive approach, control current is applied for a brief time interval following each pacing stimulus. In
the continuous-time approach, a piece-wise constant control current is applied at all times. We show that the continuous-time
approach suppresses alternans for longer fibers and higher pacing frequencies compared to the impulsive approach.
Key words: Cardiac dynamics, partial differential equations, periodic orbits, linear control.

1 INTRODUCTION
The contraction of the heart is controlled by an electrical wave (electrical activity) originating at the sinoatrial node. In the
normal rhythm this wave propagates over the heart in a highly coordinated way. Occasionally, the cardiac rhythm can make
a transition to a lethal arrhythmic behavior known as ventricular fibrillation (VF) characterized by turbulent dynamics lacking
spatial coordination which renders the heart incapable of pumping blood. The first step in the transition from the normal rhythm
to VF is an alternation of the duration of the electrical excitation from beat to beat, known as electrical alternans [1]. It makes
sense then to assume that suppression of alternans which leads to stabilization of the normal rhythm might serve as a method for
prevention of VF.
In this work we study control methods for the suppression of alternans in a model of Purkinje fibers. These fibers, whose
physiological function is to conduct electrical excitation from the atria to the ventricles, have been extensively used in simulations
and experiments as models of one-dimensional cardiac tissue [2].

2 METHODS
2.1 Model Equations
The electrical activity in a single cardiac cell can be described by the following system of ordinary differential equations
1
V̇ = [Iion (V, y) + Iext (t)] , (1a)
Cm
ẏ = h(V, y), (1b)

where V is the transmembrane voltage, Cm is the cell membrane capacitance per unit area, Iion is the ionic current through the
cell membrane, y is a vector of variables associated with the ion channel conductance and ionic concentrations (gating variables)
and Iext is the external current applied by electrodes. The ionic models which define the functional forms of Iion (V, y) and
h(V, y) vary according to the specific type of cardiac tissue under study. To simulate Purkinje fibers we used an ionic model
with three gating variables, y = [y1 , y2 , y3 ], recently developed by one of us [8] which accurately reproduces experimental
measurements, as explained below. It is convenient to use a scaled voltage variable
V − Voff
u= , (2)
Vsc
where Voff and Vsc are chosen so that u takes values mostly in the interval [0, 1]. Equations (1a) and (1b) can be gathered into a
single equation as

ż = F (z) + Iext (t)û, (3)

where the state variable now is the column vector z = [u, y], Iext = Iext /(Cm Vsc ) and û = [1, 0, 0, 0]. To describe the electrical
activity on a fiber of length L, an independent variable x indicating the position of a cardiac cell needs to be introduced and Eq.

96
(3) generalizes to the partial differential equation (PDE)

∂t z = D∂x2 z + F (z) + jext (x, t)û, (4)

where D = D11 ûû† , with D11 = 10−4 cm2 /ms a scalar diffusion constant, which describes electric conductivity between
neighboring cells. Vanishing-electric-current boundary conditions are imposed at both ends: ∂x u|x=0,L = 0. The external
current density jext (x, t) is a sum of two contributions: the pacing and control current densities, jp (x, t) and jc (x, t), respectively.
The pacing current is applied by an electrode located at x = xp near the left end (x = 0) of the fiber,

jp (x, t) = Ip (t)g(x − xp ), (5)

where g(x) is a narrow spatial profile and Ip (t) is the pacing current consisting of a periodic train of pulses
∞
X
Ip (t) = Ip0 r(t − tn ), (6)
n=0

where tn = nT with T the pacing period, and r(t) is a square pulse of duration ∆T = 5 ms,
(
1, 0 ≤ t ≤ ∆T
r(t) = (7)
0, otherwise.

Each pacing pulse produces an excitation wave that travels down the fiber. The dynamics of cardiac tissue is conveniently
described in terms of the action potential duration (APD), defined, for a particular location x, as the time elapsed from the
moment the voltage surpasses some threshold value, uth , up until the moment if falls below that threshold. The APD is therefore
a function of x and the pacing interval n, denoted as APDn (x). Depending on the pacing period T , Purkinje fibers present
two asymptotic regimes: when T is smaller that some critical value Tc , the APD is constant from beat to beat, APDn (x) =
APDn+1 (x) for all x. This regime corresponds to the normal rhythm; when T < Tc the APD alternates between long and short
values from beat to beat, this regime corresponds to the state of alternans. Unlike previous models of Purkinje fibers, the ionic
model [8] accurately reproduces the experimentally measured bifurcation diagram of asymptotic APD as a function of T and the
shape of the voltage signal at a particular location.
Our goal is to apply a control current, Ic (t), so that the state of alternans is suppressed and the normal rhythm restored even
when T < Tc . The control current is applied by an electrode located at x = xc ,

jc (x, t) = Ic (t)g(x − xc ). (8)

The control problem consists of determining Ic (t) so as to satisfy the control goal. The solution to this problem is presented
below.

2.2 Linear stability analysis

As linear stability analysis shows, when T is decreased below Tc , the normal rhythm does not disappear, but becomes unstable
and is replaced by the state of alternans. To determine stability of the normal rhythm, we calculate the corresponding time-T
periodic orbit z0 (x, t) by solving the equation

z0 (x, 0) = G [z0 (x, 0); T, 0] , (9)

where G[ · ; T, 0] is the time evolution operator which advances the solution of (4) from t = 0 to t = T . Eq. (9) is solved by the
Newton-Krylov method [3, 4]. The operator G is implemented numerically by advancing in time the solution of (4) using the
fourth order Runge-Kutta method with a time step of 2.5 × 10−3 ms. A second order central finite difference approximation is
used for the diffusion term with a mesh size ∆x = 0.01 cm. To study the stability of z0 (x, t), we consider the evolution of the
deviation δz = z(x, t) − z0 (x, t), given in the linear approximation by

∂t δz = JN (t)δz − jc û, (10)

where JN (t) = ∂x2 + JF (t) and JF (t) = DF/Dz|z0 (x,t) is a time-periodic differential operator. Let U (t, 0) denote the time
evolution operator that advances the solution of (10) from t = 0 to an arbitrary t,

δz(x, t) = U (t, 0)δz(x, 0). (11)

The stability of the normal rhythm is determined by the eigenvalues λi (Floquet multipliers) of the monodromy operator U (T, 0),

U (T, 0)ei (x, 0) = λi ei (x, 0). (12)

97
The eigenvalues and right eigenfunctions, ei (x, 0), are calculated numerically using the Arnoldi method [5] implemented by the
MATLAB (Mathworks, Inc.) function eigs. The normal rhythm is unstable when at least the leading eigenvalue lies outside the
unit circle, |λ1 | > 1 (the eigenvalues are sorted in order of decreasing absolute value, |λ1 | ≥ |λ2 | ≥ |λ3 | ≥ ...). Fig 2 shows
the number of unstable eigenvalues (|λi | > 1) for different fiber lengths L and pacing periods T . The normal rhythm becomes
unstable (and small perturbations evolve toward the state of alternans) at a critical value Tc ≈ 205 ms for the range of lengths
shown. A region of stability is also found for some lengths for 152 ms < T < 162 ms.

2.3 Galerkin projection

The evolution equation (10) can be simplified by projecting it onto a basis of time-periodic functions ei (x, t) related to the right
eigenfunctions ei (x, 0) by

ei (x, t) = e−ωi t U (t, 0)ei (x, 0), (13)

where ωi are the Floquet exponents, ωi = log λi /T . To perform the projection we need the eigenfunctions of the adjoint operator
U † (T, 0) (left eigenfunctions of U (T, 0)),

U † (T, 0)fi (x, 0) = λ∗i fi (x, 0). (14)

An adjoint time-periodic basis fi (x, t) is obtained from the left eigenfunctions fi (x, 0) using the relation,
∗
eωi t †
fi (x, t) = U (T, t)fi (x, 0). (15)
λ∗i
The left and right basis satisfy the orthogonality condition
Z L
hfi (x, t), ej (x, t)i ≡ fi (x, t) · ej (x, t) dx = δij . (16)
0

Expressing the deviation δz(x, t) as a linear combination of the right basis functions
∞
X
δz(x, t) = ξi (t)ei (x, t), (17)
i=1

substituting (17) into (10), applying the operation hfi (x, t), ·i to both sides of the resulting equation and using (16) we obtain
(for a more detailed derivation see Ref. [9]) an evolution equation for the modal amplitudes ξi (t),

ξ˙i = ωi ξi + bi (t)Ic (t), (18)

where bi (t) = hfi (x, t), g(x − xc )ûi. We consider two control methods, in both of which the feedback is computed as a function
of the state at the discrete times tn . Hence, it is sufficient to consider the dynamics on the stroboscopic section ξin = ξi (tn ) of
the flow (18) from tn to tn+1 ,
Z T
n+1 n
ξi = λi ξi + λi e−ωi t bi (t)Icn (t) dt, (19)
0

where Icn (t)

= Ic (nT + t), t in [0, T ]. The map (19) shows that highly stable modes (characterized by |λi |  1) decay quickly
and need not be considered in the control problem.

2.4 Impulsive control

The first control method studied uses a control current consisting of brief pulses applied at times t0n = nT + τ , 0 ≤ τ < T ,

Icn (t) = qn δa (t − τ ), (20)

where δa (t) is a discrete version of the Dirac delta function

(
1/a, 0 ≤ t ≤ a, a  T,
δa (t) = (21)
0, otherwise.

Substituting (20) into (19) and writing the dynamics of modes i = 1, · · · , m with |λi | > 0.1 in vector form, we arrive at the
linear time-invariant map

ξ n+1
i = Aξ n + B d qn , (22)

98
where ξ n = [ξ1n , ξ2n , ..., ξm
n
], A is an m × m diagonal matrix with elements Aii = λi and B d is an m × 1 matrix with elements
d
Bi1 = λi exp(−ωi τ )bi (τ ). The original control goal of stabilizing the normal rhythm reduced to finding a sequence of control
perturbations {qn } such that ξ n approaches zero asymptotically. This goal can be accomplished by using linear feedback

qn = K d ξ n , (23)

where ξ n is obtained from δz(x, tn ) using

ξin = hfi (x, 0), δz(x, tn )i . (24)

The gain matrix K d in the present work is calculated using the linear quadratic regulator (LQR) method [6] implemented by
the MATLAB function dlqr. In a previous study [4] we showed that the control method based on (22) and (23) is capable of
suppressing alternans in a wider range of parameters (L, T ) compared to previously proposed empirical methods based on the
idea of time-delay autosynchronization [7]. In the present work we improve that study in two ways: we use a more realistic
model of Purkinje fibers; and we allow the control current to be piece-wise constant, so that feedback control current is applied
for the entire duration between pacing impulses, as explained below.

2.5 Continuous-time control

The second control method investigated here uses feedback current which is constant on subintervals of duration δT = T /N ,

Icn (t) = I nk , (25)

for t in [(k − 1)δT, kδT ], k = 1, 2, ..., N . Using a procedure similar to that described in the previous section, we obtain the
following map for the evolution of the modal amplitudes under the effect of this current:

ξ n+1 = Aξ n + B c I n , (26)

where B c is an m × N matrix with elements

Z kδT
c
Bik = λi e−ωi t bi (t)dt (27)
(k−1)δT

and I n = [I n1 , I n2 , ..., I nN ]. Similarly, the vector I n of control currents is given by the feedback law

I n = K c ξn , (28)

where K c is calculated using the LQR method. Both impulsive and continuous-time control were tested applying the control
currents (20) and (25), respectively, in the original PDE model (4). For a given period T 0 , we used as initial condition the normal
rhythm for the period T 00 = T 0 + 2 ms, simulating a gradual reduction of the pacing interval.

3 RESULTS
Figure 1 shows the norm of the deviation kξ n k from the normal rhythm for impulsive and continuous-time control for a fiber
of length L = 1 cm and a pacing period T = 188 ms. The number of sub-intervals used for continuous-time control was
N = 20. For impulsive control, after an initial increase (transient amplification due to time-delayed action of feedback), kξ n k
demonstrates exponential decay, indicating that the method is successful. For longer fibers, transient amplification of kδz(x, t)k
drives the dynamics into the nonlinear regime where (10) is no longer a valid approximation, leading to the failure of impulsive
control. In contrast, for continuous-time control, kξ n k decreases monotonically, and at a higher rate compared to impulsive
control. This absence of transient amplification related to a much shorter time-delay of the feedback action likely explains why
continuous-time control is successful where impulsive control fails.
The regions in the (L, T ) parameter space where feedback control succeeds (or fails) are shown in Fig. 2 for both impulsive
and continuous time control. We find that continuous-time control succeeds in the entire investigated range of L and T where
the normal rhythm is unstable, which includes fibers up to 3 cm long, while impulsive control is only effective for fibers with
length of 1 cm and below. The ability to control large tissue sizes with a single electrode might have potential advantages in
clinical applications where only a few electrodes may be used.

4 CONCLUSIONS AND REMARKS

We showed that, using a continuous-time control current, successful suppression of alternans is attained for longer fibers and
higher pacing frequencies compared to a method which uses brief pulses of control current. Furthermore, in the cases where both
methods succeed, continuous-time control suppresses alternans faster compared to impulsive control. The methods discussed

99
 0
10

−1
10

−2
10

kξ n k −3
10

−4
10

−5
10
0 20 40 60 80 100
n

Figure 1: Norm of the vector of modal amplitudes kξ n k as a function of time for impulsive (solid line) and continuous
time control (dashed line). L = 1 cm, T = 188 ms and N = 20.

3 1000000113333444444444444331000
2.5 1 0 0 0 0 0 0 1 1 3 3 3 3 3 3 4 4 4 4 4 4 4 4 3 3 3 3 1 0 0 0
L (cm)

2 1000000002233333333333333321000
1.5 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 0 0 0
1 1111111111112222222222222111000
150 160 170 180 190 200 210
T (ms)

Figure 2: Stability diagram showing the number of unstable eigenvalues for the normal rhythm. White indicates the
region in parameter space where the normal rhythm is stable. Impulsive control successfully suppresses alternans in
the gray region. Continuous-time control is successful in both the gray and black regions.

100
in this paper use feedback based on knowledge of the state at a discrete set of times tn . We have found [9] that even faster
suppression of alternans is possible using feedback which takes into account the system state at intermediate times between tn
and tn+1 , in which case, (26) generalizes to a linear time-varying map. Presently both methods assume complete knowledge of
the system state [as implied by Eq. (24)], which is not possible in an experimental setting. We are in the process of integrating
these methods with an algorithm (state observer) which reconstructs the system state from voltage measurements.

REFERENCES
[1] J. M. Pastore, S. D. Girouard, K. R. Laurita, F. G. Akar, and D. S. Rosenbaum, “Mechanism linking T-wave alternans to the
genesis of cardiac fibrillation,” Circulation, vol. 99, pp. 1385, Mar. 1999.
[2] D. J. Christini, M. L. Riccio, C. A. Culianu, J. J. Fox, A. Karma, and R. F. Gilmour, Jr., “Control of electrical alternans in
canine cardiac Purkinje fibers,” Phys. Rev. Lett. vol. 96, pp. 104101, 2006.
[3] D. A. Knoll and D. E. Keyes, “Jacobian-free Newton-Krylov methods: a survey of approaches and applications,” J. Comput.
Phys., vol. 193, pp. 357, 2004.
[4] A. Garzón, R. O. Grigoriev and F. H. Fenton, “Model-based control of cardiac alternans in Purkinje fibers,” Phys. Rev. E,
vol. 84, pp. 041927, Oct. 2011.
[5] R. B. Lehoucq, D. C. Sorensen, and C. Yang, “ARPACK Users’ Guide: Solution of Large-Scale Eigenvalue Problems with
Implicitly Restarted Arnoldi Methods” (SIAM, Philadelphia, 1998).
[6] P. Dorato, C. T. Abdallah, and V. Cerone, “Linear Quadratic Control” (Krieger, Melbourne, FL, 2000).
[7] B. Echebarria and A. Karma, “Spatiotemporal control of cardiac alternans,” Chaos, vol. 12, pp. 923, Aug. 2002.
[8] F. H. Fenton and E. M. Cherry, “Minimal model for cardiac Purkinje fibers derived from experiments,” in preparation.
[9] A. Garzón, R. O. Grigoriev and F. H. Fenton, “Continuous-time feedback control of alternans in Purkinje fibers”, in prepa-
ration.

101