ARTICLE IN PRESS

Physica A 355 (2005) 439–460

www.elsevier.com/locate/physa

Model for the respiratory modulation of the heart

beat-to-beat time interval series
Alberto Capurroa,, Luis Diambrab, C.P. Maltaa
a
Instituto de Fı´sica, Universidade de São Paulo, Rua do Matão, Travessa R 187,
CEP 05508-900 São Paulo, SP, Brazil
b
Instituto de Ciências Biomédicas, Universidade de São Paulo, Av. Lineu Prestes 1524,
CEP 05508-900 São Paulo, SP, Brazil

Received 16 January 2004

Available online 18 April 2005

Abstract

In this study we present a model for the respiratory modulation of the heart beat-to-beat
interval series. The model consists of a set of differential equations used to simulate the
membrane potential of a single rabbit sinoatrial node cell, excited with a periodic input signal
with added correlated noise. This signal, which simulates the input from the autonomous
nervous system to the sinoatrial node, was included in the pacemaker equations as a
modulation of the iNaK current pump and the potassium current iK . We focus at modeling the
heart beat-to-beat time interval series from normal subjects during meditation of the
Kundalini Yoga and Chi techniques. The analysis of the experimental data indicates that while
the embedding of pre-meditation and control cases have a roughly circular shape, it acquires a
polygonal shape during meditation, triangular for the Kundalini Yoga data and quadrangular
in the case of Chi data. The model was used to assess the waveshape of the respiratory signals
needed to reproduce the trajectory of the experimental data in the phase space. The embedding
of the Chi data could be reproduced using a periodic signal obtained by smoothing a square
wave. In the case of Kundalini Yoga data, the embedding was reproduced with a periodic
signal obtained by smoothing a triangular wave having a rising branch of longer duration than

Corresponding author. Tel.: +55 11 3091 6976; fax: +55 11 3091 6833.
E-mail address: alberto@fma.if.usp.br (A. Capurro).

0378-4371/$ - see front matter r 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2004.04.135
 ARTICLE IN PRESS
440 A. Capurro et al. / Physica A 355 (2005) 439–460

the decreasing branch. Our study provides an estimation of the respiratory signal using only
the heart beat-to-beat time interval series.
r 2005 Elsevier B.V. All rights reserved.

Keywords: Respiratory modulation; Pacemaker; Heart rate variability; Rabbit SA node; Meditation

1. Introduction

Cells within the sinoatrial (SA) node comprise the primary pacemaker site within
the heart. These cells are characterized as having no true resting potential, but
instead generate regular spontaneous action potentials. The membrane slowly
depolarizes during the diastole of the cardiac contraction cycle leading to the onset
of an action potential. The repolarization, that occurs after the spike, activates the so
called slow diastolic depolarization again and initiates a new cycle. For a
comprehensive review of the ionic currents underlying this membrane voltage
oscillation see [1].
The intrinsic automaticity (spontaneous pacemaker activity) of the SA node
determines a spiking rate of 100–110 action potentials per minute. This intrinsic
rhythm is modulated by autonomic nerves, with parasympathetic (vagal) influence
being dominant over sympathetic influence at rest. The ‘‘vagal tone’’ brings the
resting heart rate down to 60–80 beats/min. Sympathetic activation increases the
pacemaker rate with concomitant inhibition of vagal tone [2].
The acetylcholine (ACh) released by the vagal nerve binds to the muscarinic
receptors at the post-synaptic membrane, and induces three main actions [1,3,4]:
(i) through a reduction in the cyclic adenosine monophosphate (AMPc) pro-
duction it inhibits the cationic current iNaK , thus decreasing the slope of the
slow diastolic depolarization and slowing down the pacemaker rate without
changing the membrane potential, (ii) it opens a G-protein coupled potassium
channel that hyperpolarizes the cell and slows down the pacemaker rate, and
(iii) it inhibits the long lasting component of the calcium current iCa . The
first mechanism is opposite to the b adrenergic action of the sympathetic
system that increases the AMPc production [5] and the slope of the slow
diastolic depolarization. The sympathetic action also activates the iCa , but
this current has little or no effect on the heart rate. Experimental evidence
indicates that the main process that mediates the effect of low-concentration
ACh actions on the heart rate is the inhibition of the cationic current
iNaK [6].
The normal respiratory cycle is accompanied by changes in autonomic tone
that modulates the heart rate. The activity of the vagal nerve endings increases
during exhalation, and the activity of sympathetic fibers increases during inhalation,
causing the ‘‘respiratory modulation’’ (RM) or ‘‘sinus arrhythmia’’, i.e.
during inhalation the heartbeat intervals shorten and during exhalation they stretch.
The oscillation in vagal action is responsible for most of the RM, because it is faster
than the sympathetic action [2]. At typical respiratory frequencies (greater than
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 441

0.15 Hz), increase of the heart rate occurs simultaneously with the onset of
inspiratory activity, while at frequencies less than 0.15 Hz, increase of the heart
rate precedes inspiration [7]. The RM is reflected in the high-frequency power
band of the heart rate variability (HRV), which is abolished by atropine infusion,
implying that it is parasympathetically mediated [8]. It has been recently reported
that the RM may improve the efficiency of pulmonary gas exchange in healthy
humans [9].
The RM has larger amplitude in children and young subjects than in elders. In
various pathological conditions the HRV is decreased, and the RM is partially or
completely absent [8,10,11]. The chaotic structure found in normal subjects [12]
tends to disappear in pathological states [13]. The heart rate dynamics was also
studied during a variety of physiological conditions [14–17]. During Kundalini Yoga
and Chi meditation practices in normal subjects, there have been reported extremely
prominent heart rate oscillations correlated with slow breathing [18], with an
amplitude significantly larger than the RM measured in the same individuals before
meditation.
In this study we present a model for the influence of respiration on the heart beat-
to-beat interval series, focusing on heart beat data from healthy subjects during
meditation sessions of Kundalini Yoga and Chi techniques. We first show that the
embedding of the experimental beat-to-beat interval data acquires a polygonal shape
during meditation, and that the time course of respiration depends on the meditation
technique. Then, we use the model to assess the waveshape of the RM needed
to reproduce the trajectory of the experimental beat-to-beat interval series in the
phase space.

2. Methods

2.1. Meditation protocols

Two specific meditation techniques were studied: (i) Chinese Chi (or Qigong)
meditation (as taught by Xin Yan) and (ii) Kundalini Yoga meditation (as taught by
Yogi Bhajan). In this subsection we reproduce a brief description of the meditation
protocols followed [18].
The eight Chi meditators, 5 women and 3 men (age range 26–35 years), wore a
Holter recorder for approximately 10 h during which time they went about their
ordinary daily activities. After approximately 5 h of recording, each one of them
practiced 1 h of meditation. Meditation beginning and ending time were delineated
with event marks. During these sessions, the Chi meditators sat quietly, listening to
the taped guidance of the master. The meditators were instructed to breath
spontaneously while visualizing the opening and closing of a perfect lotus on the
abdomen.
The four Kundalini Yoga meditators, 2 women and 2 men (age range 20–52
years), wore a Holter monitor for approximately one and half hours, 15 min of
baseline quiet breathing were recorded before the meditation period of 1 h. The
 ARTICLE IN PRESS
442 A. Capurro et al. / Physica A 355 (2005) 439–460

meditation protocol consisted of a sequence of breathing and chanting exercises,

performed while seated in a cross-legged posture. The beginning and ending of the
various meditation sub-phases were delineated with event marks.
The Kundalini Yoga subjects were considered to be at an advanced level
of meditation training. In the case of Chi meditation, the subjects were relative
novices, having begun their meditation practice about 1–3 months before the
recordings used for this study. The persons from both groups were in good general
health conditions.

2.2. Data files

We used time series, from physionet [19], of 4 subjects prior and during Kundalini
Yoga meditation, 8 subjects prior and during Chi meditation, 4 healthy control
subjects spontaneously breathing during sleeping hours, and 4 healthy subjects
during supine metronomic breathing at 0.25 Hz. The heart beat-to-beat intervals
were obtained from electrocardiogram (ECG) data by Peng et al. [18] and made
available in physionet.
The data files (in text format) consisted of two columns, the first containing the
heart beat instant of time, and the second containing the instantaneous heartbeat
rate. The sequence of beat-to-beat time intervals was generated using the data in the
first column. From this sequence we selected segments of 300 intervals that had no
trends or intervals of double length due to undetected heartbeats. From the selected
segments of intervals we determined the autocorrelation function and a bidimen-
sional embedding [20]. The value of the embedding delay was determined by the first
zero of the autocorrelation function.

2.3. Model for the RM of the heart pacemaker

2.3.1. Pacemaker model

Several models of the SA node pacemaking in the rabbit heart have been
developed in the last years [21–25]. We selected the model in Endresen et al. [23]
because it provides a realistic and simple representation of the membrane potential
oscillations. This model, based on experimental in vitro studies of isolated rabbit SA
node cells [1,26–29], was developed to fit experimental observations of (i) the action
potential shape and (ii) the intracellular ionic concentrations of potassium, calcium
and sodium.
The model takes into account the ionic currents iK ; iCa ; iNa ; iNaK , and iNaCa . The
membrane voltage v is given by the net charge difference between the intracellular
and extracellular compartments as
FV
v¼ f½Ki þ 2½Cai þ ½Nai  ½Ke  2½Cae  ½Nae g , (1)
C
where ½ i ð½ e ) denotes the intracellular (extracellular) ionic concentration, C is the
membrane capacitance, F is Faraday’s constant and V is the cell volume.
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 443

The intracellular ionic concentrations ½Ki , ½Cai and ½Nai evolve according to
d 2iNaK  iK
½Ki ¼ , (2)
dt FV
d 2iNaCa  iCa
½Cai ¼ , (3)
dt 2FV
d iNa  3iNaK  3iNaCa
½Nai ¼ . (4)
dt FV
The currents are given by
 
ev 1 ½Ke
iK ¼ kK x sinh  ln , (5)
2kT 2 ½Ki
 
ev 1 ½Cae
iCa ¼ ðkCa fd 1 þ kb;Ca Þ sinh  ln , (6)
kT 2 ½Cai
 
ev 1 ½Nae
iNa ¼ kNa hm1 sinh  ln , (7)
2kT 2 ½Nai
 
eðv  vATP Þ ½Ke 3 ½Nae
iNaK ¼ kNaK tanh þ ln  ln , (8)
2kT ½Ki 2 ½Nai
 
ev 3 ½Nae ½Cae
iNaCa ¼ kNaCa sinh  ln þ ln , (9)
2kT 2 ½Nai ½Cai
where kK , kCa , kNa , kNaK and kNaCa are the conductance parameters corresponding
to each current (kb;Ca is the conductance of the background calcium current). vATP is
the half activation potential of the NaK pump (ATP denotes cyclic adenosine
triphosphate), k is the Bolztmann constant, e is the elementary charge, and T is the
absolute temperature.
The activation mechanism of calcium and sodium channels is very fast, so we use
the steady state fraction of open channels (denoted by d 1 for the calcium channels,
and m1 for the sodium channels) given by
  
1 2eðv  vd Þ
d 1 ¼ 1 þ tanh , (10)
2 kT
  
1 2eðv  vm Þ
m1 ¼ 1 þ tanh , (11)
2 kT
where vd and vm are the half activation potentials of the calcium and sodium
channels, respectively.
The fraction of open potassium channels, x, evolves as
      
dx 1 2eðv  vx Þ 1 2eðv  vx Þ
¼ cosh  1 þ tanh x , (12)
dt tK kT 2 kT
 ARTICLE IN PRESS
444 A. Capurro et al. / Physica A 355 (2005) 439–460

where vx is the half activation potential of the potassium channel and tK is the
relaxation time of the potassium channel.
In addition to the activation mechanism, the calcium and sodium channels have an
inactivation mechanism. The inactivation rates of the calcium and sodium channels
are given by
      
df 1 2eðv  vf Þ 1 2eðv  vf Þ
¼ cosh  1  tanh f , (13)
dt tCa kT 2 kT
      
dh 1 2eðv  vh Þ 1 2eðv  vh Þ
¼ cosh  1  tanh h , (14)
dt tNa kT 2 kT
where vf and vh are the half inactivation potentials of the calcium and sodium
channels, respectively. tCa and tNa are the relaxation time of the inactivation process
of the calcium and sodium channels, respectively.
As the values of tK and vx are very close to the analogous magnitudes of tCa and
vf , respectively, it is possible to relate the inactivation gating of calcium to the
activation gating of potassium as x þ f ¼ 1. This allows to reduce the number of
differential equations in the model [23]. Eqs. (12) and (13) are identical but have
opposite dynamics. So, we have suppressed Eq. (13), and replaced f by 1  x in
Eq. (6).

2.3.2. Pacemaker modulation

The model for the rabbit SA node presented above [23] constitutes a plausible
approximation to the human SA node, which is the source of the beat-to-beat
interval data used in the present study. Our model for the RM of the human SA node
was obtained by introducing a modulation in the sodium–potassium conductance
parameter kNaK , and in the relaxation time of the potassium current tK in Eqs. (8)
and (12) of the model for the rabbit SA node. This was accomplished by setting
kNaK ðtÞ ¼ kNaK IðtÞ , (15)
and
tK
tK ðtÞ ¼ , (16)
5:0IðtÞ
with IðtÞ being a time dependent input signal. By replacing kNaK by kNaK ðtÞ in
Eq. (8), the current iNaK is modulated by IðtÞ. The current iK is modulated by IðtÞ in
Eq. (5) by replacing tK by tK ðtÞ in Eq. (12). The reason for using the input signal
multiplied by 5.0 in (16) is explained later in this section.
The currents iNaK and iK where selected to be the target of regulation in the model
because they mediate the main action of the autonomous nervous system on the
heart rate [1]. As the inactivation gating of calcium channels was related to the
activation gating of potassium channels as x þ f ¼ 1, we also included in the model a
regulation of iCa , i.e. the relaxation time of the calcium inactivation is modulated in
the same way as the relaxation time of the potassium activation. This last
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 445

modulation was included in the model because it is also described in the literature [1],
although it may be not relevant for the heart rate.
The input IðtÞ was formed by the sum of two time dependent signals (SðtÞ and
NðtÞ) and the constant term A. SðtÞ is a periodic wave representing the influence of
respiration. NðtÞ is a realization of correlated noise addressed to the variability of the
breathing rhythm around a periodic pattern and to other influences coming through
the autonomic innervation of the SA node. The constant term A is inversely related
to the tonic activity of the colinergic input to the SA node (i.e. the vagal tone is
represented by A). The sum SðtÞ þ NðtÞ evolves as a mirror image of the oscillatory
component of the vagal action and parallel to the respiratory cycle.
In this way, the inputs to the pacemaker model affect the main currents that are
regulated by the action of the autonomous nervous system [1]. The manner in which
the influences are introduced (Eqs. (15) and (16)) is arbitrary but physilologically
plausible. In a previous study [30], we simulated the RM by injecting an external
current in the membrane of a simpler model (Hodgkin–Huxley equations), without
including direct modulations of the ionic currents. Indeed, the trajectory of the beat-
to-beat intervals in the phase space can be qualitatively reproduced by a general type
of oscillator, provided the representation of the input signal in the output spike train
is close to linear (see Section 3.1). With the present model, in which the input signals
modulate specific ionic currents of the pacemaker, we could reproduce the results
also quantitatively. The modulation of specific ionic currents is a necessary condition
for this success, since we checked that the embedding shape of the data could not be
reproduced when the input signals to the rabbit pacemaker model consisted of
external current.
In Section 3.1 it will be shown that, in order to reproduce the experimental results,
the input signal has to be quasi-linearly represented in the output spike train. In
order to achieve this condition kNaK and kK were set to 15.0 and 120.0 pA,
respectively. The other parameter values were set as provided in Endresen et al. [23].
The values used for the parameters, initial conditions, observed constants, and
fundamental physical constants are given in the Table 1.
We integrated the system using a forward Euler algorithm (using C codes). At first
the model was run excited only by the constant term A (with SðtÞ and NðtÞ set to
zero) until a constant beat rate was reached. Then, the time dependent modulations
were activated and the actual simulation started. Each realization lasted 106 sample
steps, that corresponds to 200 s (sample step ¼ 0:2 ms). This was enough to obtain
200–300 simulated heart beat-to-beat intervals, using a spike detection algorithm
explained at the end of this section.
We checked that the time course followed by the membrane voltage and the ionic
currents of our simulation are similar to those shown in Endresen et al. [23]. The
Fig. 1A and B shows an example of the membrane potential oscillation and the
currents iK ; iCa ; iNa ; iNaK , and iNaCa (the time dependent modulations were set to
zero). The sense of the current is always positive (i.e. outward sense) for iNaK and iK .
When the time dependent modulation SðtÞ þ NðtÞ is activated, we observe changes in
these currents, which are related to the cycle of SðtÞ. The oscillation of iNaK in each
action potential is shifted in the inward sense (i.e. to lower values) when the intervals
 ARTICLE IN PRESS
446 A. Capurro et al. / Physica A 355 (2005) 439–460

Table 1
Values used for the parameters, initial conditions, observed constants, and fundamental physical constants

Adjustable parameters Physical constants

Name Value Name Value

kCa 26.2 pA k 1:38  1020 mJ K1

kb;Ca 0.01645 pA e 1:6  1019 C
kNa 112.7 pA F 96485:3 C mol1
kNaCa 1400.0 pA R 8314:5 J kmol1 K1
kK 120:0 pA
kNaK 15:0 pA Observed constants

Name Value

T 310.15 K
Initial conditions V 10  103 mm3

Name Value ½Ke 5.4 mM

x0 0.1 ½Cae 2 mM
h0 0.008 ½Nae 140 mM
½Ki0 130.66 mM C 47 pF
½Cai0 0.0006 mM vx ¼ vf 25.1 mV
½Nai0 18.73 mM vd 6.6 mV
vm 41.4 mV
vh 91 mV
vATP 450 mV
tK ¼ tCa ¼ tNa 200 ms

The asterisks mark the parameters that are different from the value provided in Endresen et al. (2000).

shorten to its minimum value (Fig. 1C and D), corresponding to the peak of SðtÞ. In
the case of iK , the maximum value does not change with SðtÞ, but its duration
increases at the valley of SðtÞ when the intervals reach their larger duration (not
shown).
The difference in the membrane potential from the peak to the valley of SðtÞ is
only 2.5 mV (not shown) suggesting that the dominant mechanism of interval
modulation is provided by the regulation of iNaK , as seems to be the case in the SA
node cells [6,1]. In order to estimate the relative contribution of iNaK and iK to the
interval modulation of our model, we focused at the maximum interval modulation
amplitude induced by a typical RM oscillation of the Yoga data simulation. The
oscillation amplitude reached 0.28 s when the two currents received the modulation,
and was reduced to 0.24 s when we selectively eliminated the effect on iK . When we
eliminated selectively the effect on iNaK , the amplitude was reduced to only 0.04 s
(not shown). Similar results were obtained using a RM oscillation of the Chi data
simulation. Thus, as occurs in the rabbit SA node [1], the main effect of interval
modulation in our model is mediated by the iNaK , while the modulation of iK is
needed to reproduce the heart rate of the different data sets. To set our model in this
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 447

20 0.8

Beat-to-beat interval (s)

0

0.7
-20
v (mV)

-40
0.6

-60

-80 0.5
0 0.4 0.8 1.2 1.6 0 4 8 12 16 20
(A) Time (s) (C) Time (s)
200

iK 15
100
iNaK (pA)

i NaK 13
i (pA)

i NaCa iNa
11
-100

i Ca
-200 9
0 0.4 0.8 1.2 1.6 0 4 8 12 16 20
(B) Time (s) Time (s)
(D)

Fig. 1. (A) Membrane potential v vs. time (A ¼ 0:01, SðtÞ ¼ 0:0 and NðtÞ ¼ 0:0). (B) Ionic currents vs.
time (corresponding to the membrane potential shown in the panel A): iNaK (full thick line), iK (full thin
line), iCa (dashed line), iNaCa (dot-dashed line) and iNa (dotted line). (C) Beat-to-beat interval vs. time. The
input modulation was the last cycle of IðtÞ, with the parameters and waveshape used for the simulation of
the Kundalini Yoga data (Table 2), with the exceptions of SðtÞ amplitude and period, here set to 0.033 and
16.7 s, respectively. (D) Ionic current iNaK vs. time (corresponding to the beat-to-beat interval plot shown
in the panel C). Note that the oscillation of iNaK is shifted in the direction of the zero value when the beat-
to-beat intervals reach their lower duration (higher beat rate), and vise-versa.

regime, the input to kNaK was IðtÞ, while the input to tK was 5:0IðtÞ (Eqs. (15)
and (16)).
The analysis of Eqs. (15) and (16) shows that the input signals are dimensionless.
The methods to obtain these signals are presented in the following.
In order to obtain NðtÞ, realizations lasting 106 sample steps of the following
random differential equation (x is a white Gaussian noise),
dN
¼ l1 N þ l2 x , (17)
dt
were low pass filtered with a Hanning window (half width ¼ 0:06 s). The equation
was numerically solved using a forward Euler algorithm. The correlation level of
 ARTICLE IN PRESS
448 A. Capurro et al. / Physica A 355 (2005) 439–460

NðtÞ, controlled through the parameter l1 [31], was set to 0.04 (l2 was set to 1.0). The
mean value of NðtÞ tends to zero, but in order to ensure an exact zero mean in
realizations with a limited number of points, we added or subtracted an appropriate
value so as to drive the signal mean value to zero. The standard deviation (SD) of
NðtÞ was used as a measurement of the signal amplitude. It was initially set to 1.0, by
multiplying each value by the appropriate factor. In this way, the final SD of the
signal actually input into the model could be easily controlled by multiplying the
time series by the SD value desired in each simulation. NðtÞ was produced with a
code in C and fed into the pacemaker code.
For the simulation of the Kundalini Yoga data, the periodic signal SðtÞ was
obtained by smoothing a triangular wave of zero mean using a Hanning window
(half width ¼ 6:0 s). For the simulation of the Chi data, SðtÞ was obtained by
smoothing a square wave of zero mean, using a Hanning window (half
width ¼ 3:0 s). For the simulation of pre-meditation and control data, we used a
sine wave of zero mean. The period of SðtÞ was adjusted to each of the data sets to be
modeled (Table 2).
In the case of the triangular wave used to simulate the Kundalini Yoga data, the
percentage of the period occupied by the rising branch of the wave was
systematically changed in steps of 10%, to finally set it at 80%. After smoothing
the triangular wave, this percentage changed to 60%.
The amplitude of the three periodic waves used for SðtÞ was initially set to 1.0. In
this way, the final amplitude of the signal injected into the model could be easily
reached by multiplying SðtÞ by the value of amplitude desired in each simulation.
The amplitudes of SðtÞ and NðtÞ (Table 2) were set, in successive tests, to reproduce
(i) the shape of the interval plot and embedding of the beat-to-beat interval data, and
(ii) the first autocorrelation coefficient of the beat-to-beat interval data equal or less
than zero.
Segments of the time dependent modulation SðtÞ þ NðtÞ used for the simulation of
meditation are depicted in Fig. 7. SðtÞ was generated using a code in C and then fed
into the code for the pacemaker.
The model described above simulates the membrane potential vðtÞ of the
pacemaker that triggers the heartbeat. The timing of the spikes was determined
with an algorithm that detects the crossing of a threshold with positive derivative.
The detecting level (threshold) was set at 20:0 mV. The absence of errors in the
spike detection was checked graphically.

Table 2
Parameter sets used in the numerical simulation of Kundalini Yoga data (row 2), Chi data (row 4)

Case SðtÞ period (s) SðtÞ amplitude NðtÞ amplitude SðtÞ=NðtÞ A

Pre-Yoga 3.3 0.006 0.01 0.6 0.1

Yoga 15.0 0.03 0.018 1.67 0.01
Pre-Chi 4.0 0.008 0.018 0.44 0.09
Chi 18.2 0.028 0.018 1.56 0.06

The parameters used for the same individuals before meditation are given in rows 1 and 3, respectively.
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 449

3. Results

In this section, we describe the experimental data, and make comparison with the
model simulations.

3.1. Experimental data

In Fig. 2 we display examples of the Kundalini Yoga and Chi data. As reported by
Peng et al. [18], extremely prominent heart rate oscillations correlated with breathing
appear during both forms of meditation. These oscillations have significantly larger
amplitude and lower frequency during meditation than before meditation or in
healthy control subjects [18].
The beat-to-beat intervals recorded before and during Kundalini Yoga meditation
are shown in Fig. 2A and B, while the beat-to-beat intervals recorded before and
during Chi meditation are shown in Fig. 2E and F. We see clearly that the beat-to-
beat interval exhibits low-frequency oscillations with very large amplitude during
both forms of meditation. The recordings performed in the same subjects before
meditation do not display such large oscillations, although the RM can be
recognized as low-amplitude oscillations of faster rate (Fig. 2A and E).
In the 4 cases of Kundalini Yoga meditation analyzed, we observed (in addition to
the oscillations) a pronounced decrease of the mean beat-to-beat interval (e.g.,
compare Fig. 2A with B). The comparison of the mean interval values before and
during meditation with the U-test gives significant differences with a level of risk
lower than 1%. The decrease of the mean beat-to-beat interval was small or did not
occur in the 8 cases of Chi meditation (e.g., compare Fig. 2E with F) (there are no
significant differences in the U-test).
The oscillation of the heart beat interval duration is noteworthy in the
autocorrelation function of the beat-to-beat series. The full lines in Fig. 3 show
the autocorrelation of the cases depicted in Fig. 2. During meditation, the
autocorrelation presents high amplitude slowly damped oscillations (full lines in
Fig. 3B and D) due to the considerable regularity of the interval oscillations. Before
meditation the RM can be recognized in the autocorrelation as low-amplitude
oscillations of faster rate (full lines in Fig. 3A and C).
The amplitude of the autocorrelation oscillations is larger in Yoga meditation
than in Chi, indicating that the beat-to-beat interval oscillations are more regular in
these last data sets (e.g., compare full lines in Fig. 3B and D). The larger regularity of
the beat-to-beat interval oscillations during Yoga meditation holds in general for all
data sets: the amplitude of the first autocorrelation oscillation is significantly larger
for Yoga (U-test at 5% of risk). The mean values SD are 0:85 0:22 for Yoga and
0:33 0:24 for Chi. The amplitude of the beat-to-beat interval oscillations, estimated
through the SD of the complete beat-to-beat interval series, is not significantly
different in both types of meditation.
The shape of the embedding changes significantly during meditation in the
cases depicted in Fig. 2. The points lie in a roughly circular shape before meditation
(Fig. 2C and G). During Kundalini Yoga meditation the points of the embedding lie
 ARTICLE IN PRESS
450 A. Capurro et al. / Physica A 355 (2005) 439–460

1.2 1.2

Beat-to-beat interval (s)

Beat-to-beat interval (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0 40 80 120 0 40 80 120
(A) Beat (B) Beat

1.2 1.2
Beat-to-beat interval [i+5] (s)

Beat-to-beat interval [i+7] (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(C) Beat-to-beat interval [i] (s)
(D) Beat-to-beat interval [i] (s)

1.2 1.2
Beat-to-beat interval (s)
Beat-to-beat interval (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0 40 80 120 0 40 80 120
(E) Beat (F) Beat

1.2 1.2
Beat-to-beat interval [i+6] (s)

Beat-to-beat interval [i+6] (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(G) Beat-to-beat interval [i] (s) (H) Beat-to-beat interval [i] (s)
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 451

1 1

0.75

0.5
0.5
Autocorrelation

Autocorrelation
0.25
0
0

-0.25
-0.5

-0.5

-0.75 -1
0 1 2 3 4 5 6 0 5 10 15 20 25 30
(A) Interval order (B) Interval order

1 1

0.75

0.5
0.5
Autocorrelation

Autocorrelation

0.25
0
0

-0.25
-0.5

-0.5

-0.75 1-
0 1 2 3 4 5 6 0 5 10 15 20 25 30
(C) Interval order (D) Interval order

Fig. 3. (A) Full lines: autocorrelation functions of the beat-to-beat interval series shown in Fig. 2. Dashed
lines: 99% confidence intervals of the autocorrelation functions corresponding to the simulations depicted
in each panel (N ¼ 50 model realizations). The beat-to-beat intervals of single realizations of these four
simulations are shown in Fig. 5. (A) Before Yoga meditation. (B) During Yoga meditation. (C) Before Chi
meditation. (D) During Chi meditation.

on a triangular shape (Fig. 2D), and during Chi meditation the shape is
quadrangular (Fig. 2H).
The shape of the embedding is related to the shape of the beat-to-beat interval
plot. The triangular embedding is associated with the beat-to-beat interval
oscillation having sharp peaks (Fig. 2B), while the quadrangular embedding is
related to beat-to-beat interval oscillation having peaks and valleys of similar width
(Fig. 2F).
A remarkable property of the beat-to-beat interval data during meditation is the
presence of noise induced indentations both in the peaks and valleys of the interval

Fig. 2. Beat-to-beat interval (s) vs. beat, (A) before and (B) during Kundalini Yoga meditation. (C)
Embedding of the time series shown in the panel A (300 intervals with delay ¼ 5). (D) Embedding of the
time series shown in the panel B (300 intervals with delay ¼ 7). Beat-to-beat interval (s) vs. beat, (E) before
and (F) during Chi meditation. (G) Embedding of the time series shown in the panel E (300 intervals with
delay ¼ 6). (H) Embedding of the time series shown in the panel F (300 intervals with delay ¼ 6).
 ARTICLE IN PRESS
452 A. Capurro et al. / Physica A 355 (2005) 439–460

oscillations (Fig. 2B and F). In other words, the edges of the embeddings (triangular
or quadrangular) have similar width, i.e. they are similarly noisy (Fig. 2D and H).
This indicates that the SA node operates in a regime in which the relationship
between the input modulations from the autonomous nervous system and the output
beat-to-beat interval series is approximately linear, i.e. the saturation level described
in Goldberger et al. [32] is not reached. This property has to be taken into account by
the model in order to properly reproduce the embedding shape of the data.
In the next subsection, we shall assess through numerical simulations the shape of
the input signals needed to reproduce the results presented above. The simulations
are based in the cases shown in the Fig. 2, because these segments have a very regular
waveshape in the beat-to-beat interval plot. Regarding the recordings from the other
individuals, in 3 out of 4 cases of Kundalini Yoga meditation analyzed, the
embedding and the beat-to-beat interval plots are similar to the case depicted in the
Fig. 2, for example Fig. 4A and B. In the remaining case the embedding is more
irregular although it also evokes a triangular shape (not shown). In the case of Chi
meditation we found larger differences between the different subjects. In 3 out of 8

1.2 1.2
Beat-to-beat interval [i+8] (s)
Beat-to-beat interval [i+9] (s

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(A) Beat-to-beat interval [i] (s) (B) Beat-to-beat interval [i] (s)

1.2 1.2
Beat-to-beat interval [i+6] (s)

Beat-to-beat interval [i+4] (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(C) Beat-to-beat interval [i] (s) (D) Beat-to-beat interval [i] (s)

Fig. 4. Embedding of the beat-to-beat intervals of different individuals during Yoga meditation, (A) 1121
intervals, (B) 906 intervals. Embedding of the beat-to-beat intervals of different individuals during Chi
meditation, (C) 4027 intervals, (D) 3759 intervals.
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 453

cases the embedding shape and the beat-to-beat interval plot are similar to the case
shown in the Fig. 2H and F, for example Fig. 4C. In the remaining 5 subjects, the
embedding is oval or has irregular shape (for example Fig. 4D), and the oscillation of
the beat-to-beat interval plot is less regular than in Fig. 2F. In some of these cases the
waveshape of the oscillation resembles that before meditation although the
amplitude is higher and the frequency lower. In general the pre-meditation
recordings present roughly circular shapes as shown in Fig. 2C and G, but in many
cases they are more irregular as also happens with the control subjects during
sleeping hours, and with the subjects during metronomic breathing (not shown).
The larger variability within different subjects observed in the case of Chi
meditation may be related to the shorter training period of the practitioners used for
recording the data. The reason being that a larger training period enables the
practitioner to obtain a more regular breathing rhythm during meditation. In Fig. 4
we show examples of embeddings calculated from the complete recording time in
different individuals during meditation.

3.2. Simulation of the beat-to-beat interval series

Using the model for the RM of the SA node (Eqs. (1)–(17)), we generated
numerically the beat-to-beat interval series during and before both types of
meditation. In successive tests, we selected the appropriate parameter sets for
obtaining good agreement with the experimental data. These parameters are
displayed in the Table 2. From this table we can draw the following conclusions. The
period and the amplitude of SðtÞ are larger for meditation data. The ratio SðtÞ=NðtÞ
(given in the 5th column of the Table 2) is larger during meditation. The parameter
A, set so as to fit the mean rate of each recording, resulted larger during meditation,
and resulted larger for Yoga than for Chi meditation. The parameters obtained for
the control subjects are similar to those for the pre-meditation situation (not shown).

3.2.1. Simulation of Kundalini Yoga data

The Kundalini Yoga data could be reproduced using a signal SðtÞ obtained by
smoothing a triangular wave. The rising branch of this signal lasted 60% of the
period (Fig. 7A). In a simulation using the parameter set given in the second row
of the Table 2, we obtained a beat-to-beat interval oscillation with sharp peaks
(Fig. 5B), and a triangular embedding (Fig. 5D), just like the experimental data (Fig.
2B and D).
The pre-meditation data could be reproduced using a sinusoidal SðtÞ, the points of
the embedding falling roughly inside a circle (Fig. 5C). The waveform of the periodic
signal was not crucial in this case, because the amplitude of SðtÞ for pre-meditation is
much smaller than in the case of meditation (see parameters in the Table 2). In these
conditions, SðtÞ cannot be distinguished from the ‘‘background noise’’ represented
by NðtÞ.
The key factor to obtain a polygonal embedding in the simulations is the
amplitude ratio of the periodic and aperiodic modulations, SðtÞ=NðtÞ (5th column of
the Table 2). The embedding of the simulation before meditation becomes triangular
 ARTICLE IN PRESS
454 A. Capurro et al. / Physica A 355 (2005) 439–460

1.2 1.2

Beat-to-beat interval (s)

Beat-to-beat interval (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0 40 80 120 0 20 40 60 80 100 120
(A) Beat (B) Beat

1.2 1.2
Beat-to-beat interval [i+5] (s)

Beat-to-beat interval [i+7] (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(C) Beat-to-beat interval [i] (s) (D) Beat-to-beat interval [i] (s)

1.2 1.2
Beat-to-beat interval (s)
Beat-to-beat interval (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0 40 80 120 0 40 80 120
(E) Beat (F) Beat

1.2 1.2
Beat-to-beat-interval [i+6] (s)
Beat-to-beat interval [i+6] (s)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(G) Beat-to-beat interval [i] (s) (H) Beat-to-beat interval [i] (s)
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 455

1.2 1.2
Beat-to-beat interval [i+7] (s)

Beat-to-beat interval [i+7] (s)

1 1

0.8 0.8

0.6 0.6

20% 50%
0.4 0.4
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2
(A) Beat-to-beat interval [i] (s) (B) Beat-to-beat interval [i] (s)

1.2
Beat-to-beat interval [i+7] (s)

0.8

0.6
80%
0.4
0.4 0.6 0.8 1 1.2
(C) Beat-to-beat interval [i] (s)

Fig. 6. Embedding of the beat-to-beat interval series simulated using periodic signals with different
percentages of the period occupied by the rising branch of the wave. The general parameters were as in the
second row of Table 2 with the exceptions of SðtÞ amplitude and period, here set to 0.033 and 16.7 s,
respectively. The rising branch of SðtÞ occupied 20% of the period in (A), 50% of the period in (B), and
80% of the period in (C). These percentages are printed in the respective panels.

if the amplitude of the periodic signal SðtÞ is increased, the other parameters being
kept fixed. Similarly, the polygonal embedding obtained in the simulations of
meditation progressively looses definition as the ratio SðtÞ=NðtÞ decreases.
The spatial orientation of the embedding varied with the percentage of the period
occupied by the rising branch of SðtÞ. This feature can be better appreciated in
simulations in which the triangular wave SðtÞ was not smoothed with the Hanning
window (Fig. 6). Using rising branches ranging from 10% to 30% we obtained a

Fig. 5. Numerical simulation of the beat-to-beat interval (s) vs. beat, (A) before and (B) during Kundalini
Yoga meditation. (C) Embedding of the time series shown in the panel A (240 intervals with delay ¼ 5).
(D) Embedding of the time series shown in the panel B (240 intervals with delay ¼ 7). Numerical
simulation of the beat-to-beat interval (s) vs. beat, (E) before and (F) during Chi meditation. (G)
Embedding of the time series shown in the panel E (240 intervals with delay ¼ 6). (H) Embedding of the
time series shown in the panel F (240 intervals with delay ¼ 6).
 ARTICLE IN PRESS
456 A. Capurro et al. / Physica A 355 (2005) 439–460

triangular embedding (Fig. 6A), but its orientation is different from the Yoga
data (Fig. 2D). For rising branches of 50% the embedding tends to a quadrangular
shape (Fig. 6B) because the left edge of the triangle shown in the Fig. 6A
bends forming a new angle. The orientation of this quadrangular shape (Fig. 6B)
is different from the embedding of the Chi data (Fig. 2H). For rising branches
ranging from 70% to 90% of the period, the embedding becomes triangular again,
and this time the orientation is as observed in the Yoga data (Fig. 6C). Thus, the
orientation of the embedding is related to the percentage of the period occupied by
the rising branch of SðtÞ. We chose the percentage that better reproduced the
embedding shape of the data. This percentage was 80% without smoothing the
triangular wave (Fig. 6C), and changed to 60% after the smoothing (Fig. 5D). The
shape of the simulated embedding matched the data embedding only when we used
the smoothed SðtÞ. To check the criterion of the embedding orientation, we counted
the number of beats in the increasing and decreasing branches of the beat-to-beat
interval oscillations of the experimental and numerical data. In both cases, the
percentage of the decreasing branch of the interval oscillation (that corresponds to
the rising branch of SðtÞ) has a good agreement with the value selected according to
the embedding orientation.
The results described above suggest that the respiratory signal of the Kundalini
Yoga meditation may look like the noisy periodic wave shown in Fig. 7A, in the
sense that the wave has sharp peaks, and that the inhalation lasts longer than the
exhalation.

3.2.2. Simulation of Chi data

The Chi data could be reproduced using a signal SðtÞ obtained by smoothing a
square wave (Fig. 7B). In the simulation using the parameter set given in the fourth

0.1 0.1

0.05 0.05
S(t)+N(t)

0 0

-0.05 -0.05

-0.1 -0.1
0 20 40 60 80 100 20 40 60 80 100
(A) t (s) (B) t (s)

Fig. 7. Single realization of the time dependent input modulation SðtÞ þ NðtÞ obtained with the parameter
set provided in the Table 2 used for the numerical simulation of Kundalini Yoga (A) and Chi (B)
meditation data. The ordinate axis shows the amplitude of the modulation on iNaK (dimensionless). The
same modulations were used for iK with an amplitude 5.0 times larger.
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 457

row of the Table 2, we obtained a beat-to-beat interval oscillation with wide peaks
and valleys (Fig. 5F), and a quadrangular embedding (Fig. 5H), just like the
experimental data (Fig. 2F and H).
As the embedding orientation of the model coincided with that of the experimental
data, we concluded that in this case the rising and decreasing branches of the signal
SðtÞ should have the same duration. To confirm this, we checked that both branches
of the beat-to-beat interval oscillation had similar number of intervals, both in the
experimental and numerical data.
These results suggest that the respiratory signal of Chi meditation may look like
the noisy periodic wave shown in Fig. 7B, in the sense that it has peaks and valleys of
approximatively same width.
The pre-meditation data could be reproduced using a sine, as for the pre-Yoga
data. The points of the embedding fall in a roughly circular shape (Fig. 5G). The
data sets recorded during Chi meditation in which the embedding shape was not
square but oval (e.g., Fig. 4D) could be reproduced using a sine of larger amplitude
and period than for the case before meditation (not shown).

3.2.3. Statistical validation of the numerical simulations

The simulations presented in Fig. 5 correspond to single realizations of the model.
As each realization is performed with a different seed of noise (Eq. (1)), we calculated
the averaged autocorrelation function of 50 model realizations with its 99%
confidence intervals, in order to check if the autocorrelation of the beat-to-beat
interval data is contained within these limits.
In Fig. 3 the first oscillation of the autocorrelation functions estimated from the four
data segments that provided the parameter sets for each simulation (i.e. before and
during Yoga meditation, and before and during Chi meditation) are plotted with full
line. Note that in the four cases the full lines are within the dashed lines that represent
the confidence intervals of the simulations. The only exception is the case of Chi
meditation for autocorrelation coefficients of order larger than 16. In this case the
autocorrelation oscillation of the simulation is damped more slowly than in the data.

4. Discussion

The present work is based on the paper by Peng et al. [18] that first reported the
occurrence of high-amplitude heart rate oscillations during Chi and Kundalini Yoga
meditation practices. These authors applied spectral analysis and a technique based
on Hilbert transform to quantify the heart rate oscillations. They found that the
amplitude of the oscillations and its Fourier and Hilbert powers are significantly
greater during meditation than before meditation or in control subjects. Measuring
fluctuations in the mean cardiac electrical axis that accompany breathing [18,33,34]
derived respiratory signals from body surface ECGs and showed that the heart rate
oscillations are correlated with respiration at the same rate.
Our contribution starts with the analysis of data segments using a bidimensional
embedding [30], and continues with the development of a biophysical model to
 ARTICLE IN PRESS
458 A. Capurro et al. / Physica A 355 (2005) 439–460

reproduce the RM of the beat-to-beat interval data. In this model, we assumed

that the respiratory input from the autonomous nervous system to the SA node
mediates the fluctuations between heart beat intervals in short time scales [8], as the
duration of the selected data segments in which our simulations are based
(200–300 s). The origin of the rhythmic activity of the colinergic input to the SA
node is a very complex issue, but it seems clear that the respiratory movements cause
parallel variations in the vagal tone that affect the discharging rate of the SA node
[2]. This is supported by early works showing that the RM is suppressed by cervical
vagotomy [35]. In addition, the high-frequency power band of the HRV
(0.15–0.4 Hz), which contains the RM, is suppressed by parasympathetical blocking
with atropine infusion [36]. Thus, in our model we represent the high-frequency
power band of the HRV with a periodic wave with added correlated noise. The low-
frequency band (0.04–0.15 Hz), due to baroreceptor and thermoregulatory
influences, is represented only generically by the added correlated noise. As the
power spectrum of NðtÞ covers both the high- and low-frequency bands of the HRV,
this signal is addressed to aperiodic influences affecting the RM and the low-
frequency power band. We ignored other components of heart rate regulation,
contained in the very low power (0.0033–0.04 Hz) and ultra low power (less than
0.0033 Hz) frequency bands of the HRV, which are excluded due to the use of short
lasting data segments without trends. The very low-frequency band is due to the
renin-angiotensin and peripheral vasomotor systems, while the ultra low-frequency
band is due to neuroendocrine and circadian rhythms [8]. In the following we discuss
some specific points of our results.

4.1. The RM during meditation

The first claim that we can make from our analysis of the data confirms the results
shown in Peng et al. [18]: during both types of meditation the RM on HRV becomes
very large, and predominates over aperiodic components. Before meditation,
although noteworthy in the autocorrelation function, the RM is hidden in a
background of aperiodic variability.
The large difference in the value of the parameter SðtÞ and in the ratio SðtÞ=NðtÞ
used in the simulation of meditation and pre-meditation data (Table 2) is a
consequence of the very high amplitude of the oscillations that appear during
meditation. It is important to stress the similarity of the parameters for both
meditation techniques. This is in accordance with the suggestion that meditation
protocols from different cultures may share certain types of autonomic exercises
mediated in part by specialized breathing maneuvers [18].

4.2. Sustained rate changes and resting vagal tone

The parameter A allowed us to reproduce in the model the sustained changes

observed in the heart rate. This is remarkable because, using simpler oscillators we
could reproduce qualitatively the waveshape of the meditation data [30], but not the
sustained changes in rate. With the present model, we are able to reproduce
 ARTICLE IN PRESS
A. Capurro et al. / Physica A 355 (2005) 439–460 459

quantitatively the heart rate as well as the amplitude and waveshape of the
variability found in different data sets. The parameter A represents the existence
of a resting vagal tone that continuously slows down the frequency of the SA
node, more precisely the vagal tone is represented by A. The pronounced
decrease in the mean value of the beat-to-beat interval observed during Kundalini
Yoga meditation may be related to a decrease in the tonic vagal action, i.e. this
meditation type has the larger value of A (Table 2). The concomitant large increase
of the RM amplitude suggests that the oscillatory component of the vagal action (i.e.
the mirror image of SðtÞ þ NðtÞ) may be greatly enhanced. Thus, there may exist a
dissociation in the regulation of the oscillatory and tonic components of the vagal
action.

4.3. Differences in the time course of respiration

The use of the embedding allowed us to focus on certain differences between

the time course of respiration followed in both forms of meditation. The shape of
the embedding is triangular for Kundalini Yoga data, and quadrangular for
Chi data. These shapes are related to the beat-to-beat interval plots, but in such plots
the differences are not easily detected due to the noisy components of the variability.
In the embedding, although the trajectories vary, they are limited to geometric
shapes easy to recognize. These shapes can be considered as hallmarks of each
meditation type.
If we knew the relationship between SðtÞ þ NðtÞ and the respiratory movements,
the time course of respiration could be inferred from the variability of the heart beat-
to-beat interval series. We assumed that this relationship is linear, because this seems
to be the case for frequencies in the range of breathing [7]. With this condition, the
respiration during Kundalini Yoga and Chi meditations should have the waveshapes
shown in Fig. 7. Respiratory signals may be derived from body surface ECGs by
measuring fluctuations in the mean cardiac electrical axis that accompany breathing
[33,34]. These ECG derived estimations can be produced using the same
ECG voltage files from which the heart beat interval series where obtained. There
exists the possibility of contrasting the respiratory signals shown in Fig. 7 with these
ECG derived estimations of the respiratory movements. These comparisons will be
carried out in a future study, when these ECG voltage file are made available in
physionet [19].

Acknowledgements

AC acknowledges financial support from FAPESP (Fundac- ão de Amparo à

Pesquisa do Estado de São Paulo), process number 00/14357-8. CPM acknowledges
partial financial support from CNPq and FAPESP. AC is researcher of PEDECIBA
(Uruguay) and thanks to the EU Advanced Course in Computational Neuroscience
2001 (Trieste/Italy).
 ARTICLE IN PRESS
460 A. Capurro et al. / Physica A 355 (2005) 439–460

References

[1] D. DiFrancesco, Ann. Rev. Physiol. 55 (1993) 455.

[2] R.M. Berne, M.N. Levy (Eds.), Physiology, Mosby-Year Book, St. Louis, MO, 1993 (Chapter 25).
[3] D. DiFrancesco, C. Tromba, J. Physiol. 405 (1988) 477.
[4] D. DiFrancesco, C. Tromba, J. Physiol. 405 (1988) 493.
[5] D. DiFrancesco, P. Tortora, Nature 351 (1991) 145.
[6] D. DiFrancesco, P. Ducouret, R.B. Robinson, Science 243 (1989) 669.
[7] J.P. Saul, R.D. Berger, M.H. Chen, R.J. Cohen, Am. J. Physiol. Heart Circ. Physiol. 256 (1989)
H153.
[8] P.K. Stein, R.E. Kleiger, Annu. Rev. Med. 50 (1999) 249.
[9] N.D. Giardino, R.W. Glenny, S. Borson, L. Chan, Am. J. Physiol. Heart Circ. Physiol. 284 (2003)
H1585.
[10] A.L. Goldberger, L.A.N. Amaral, J.M. Hausdorff, P.Ch. Ivanov, C.K. Peng, H.E. Stanley, Proc.
Natl. Acad. Sci. USA 99 (2002) 2466.
[11] M. Costa, A.L. Goldberger, C.K. Peng, Phys. Rev. Lett. 89 (2002) 068102.
[12] A. Babloyantz, A. Destexhe, Biol. Cybern. 58 (1988) 203.
[13] C.S. Poon, C.K. Merrill, Nature 389 (1997) 492.
[14] Y. Ichimaru, K.P. Clark, J. Ringler, W.J. Weiss, Comput. Cardiol. 17 (1990) 657.
[15] M. Sakakibara, S. Takeuchi, J. Hayano, Psychophysiology 31 (1994) 223.
[16] L.A. Lipsitz, F. Hashimoto, L.P. Lubowsky, J.E. Mietus, G.B. Moody, O. Appenzeller, A.L.
Goldberger, Br. Heart J. 74 (1995) 390.
[17] P.F. Migeotte, G.K. Prisk, M. Paiva, Am. J. Physiol. Heart Circ. Physiol. 284 (2003) H1995.
[18] C.K. Peng, J.E. Mietus, Yanhui Liu, K. Gurucharan, P.S. Douglas, H. Benson, A.L. Goldberger, Int.
J. Cardiol. 70 (1999) 101.
[19] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.Ch. Ivanov, R.G. Mark, J.E. Mietus,
G.B. Moody, C.K. Peng, H.E. Stanley, PhysioBank, physioToolkit and physionet: Components of a
new research resource for complex physiologic signals, Circulation 101 (2000) 23 http://
circ.ahajournals.org/cgi/content/full/101/23/e215.
[20] M. Brennan, M. Palaniswami, P. Kamen, Am. J. Physiol. Heart Circ. Physiol. 283 (2002) H1873.
[21] H.F. Brown, J. Kimura, D. Noble, S.J. Noble, A. Taupignon, Proc. R. Soc. London Ser. B 222 (1984)
329.
[22] S.S. Demir, J.W. Clark, C.R. Murphey, W.R. Giles, J. Physiol. Cell physiol. 266 (1994) C832.
[23] L.P. Endresen, K. Hall, J.S. Hoye, J. Myrheim, Eur. Biophys. J. 29 (2000) 90.
[24] H. Zhang, A.V. Holden, I. Kodama, H. Honjo, M. Lei, T. Vagues, M.R. Boyett, Am. J. Physiol.
Heart Circ. Physiol. 279 (2000) H397.
[25] Y. Kurata, I. Hisatome, S. Imanishi, T. Shibamoto, Am. J. Physiol. Heart Circ. Physiol. 283 (2002)
H2074.
[26] T. Shibasaki, J. Physiol. (London) 387 (1987) 227.
[27] J. Guo, K. Ono, A. Noma, J. Physiol. (London) 483 (1995) 1.
[28] A. Noma, Jpn Heart J. 37 (1996) 673.
[29] A. Zaza, M. Micheletti, A. Brioschi, M. Rocchetti, J. Physiol. (London) 505 (1997) 677.
[30] A. Capurro, L. Diambra, C.P. Malta, Physica A 327 (2003) 168.
[31] A. Capurro, K. Pakdaman, T. Nomura, S. Sato, Phys. Rev. E 58 (1998) 4820.
[32] J.J. Goldberger, S. Challapalli, R. Tung, M.A. Parker, A.H. Kadish, Circulation 103 (2001) 1977.
[33] G.B. Moody, R.G. Mark, A. Zoccola, S. Mantero, Comput. Cardiol. 12 (1985) 113.
[34] G.B. Moody, R.G. Mark, M.A. Bump, J.S. Weinstein, A.D. Berman, J.E. Mietus, A.L. Goldberger,
Comput. Cardiol. 13 (1986) 507.
[35] M. De Burgh Daly, in: R.M. Berne, N. Sperelakis (Eds.), Handbook of Physiology, 1979, p. 529
(Chapter 16).
[36] R. Pomeranz, J. Macaulay, M.A. Caudill, I. Kutz, D. Adam, D. Gordon, K.M. Kilborn, A.C.
Barger, D.C. Shannon, et al., Am. J. Physiol. Heart Circ. Physiol. 284 (1985) H151.