Information States in Cardiac Rhythm

Bryan Kerman(!) and Michael Meyer(2)

(I)Infodynamics, Ancaster, Ontario, Canada L9G 3M8 bryan.kennan@cciw.ca

(2)Department of Physiology, Max Planck Institute for Experimental Medicine,
Goettingen, Germany D37075

Summary. The cardiac rhythm of a healthy person is shown to be characterized by anegative

exponential probability distribution in heart rate fluctuations for sufficient time delay between
beats. This distribution, the hallmark of a statistical mechanical formulation, allows for the
characterization of the textural and structural information. Properties of the joint variation of these
forms of information show evidence of the existence of two states separated by a critical value of
heart rate variability. For low variability the heart beats regularly, but as the variability increases
the heart rate becomes increasingly intermittent. The implication ofthe information states is that
the heart is a self adjusting mechanism over the span of metabolic demands. The method allows
for a comparison of how some hearts (congestive heart failure and children) differ from adult
normal hearts, and others (normal mice) resemble those ofhumans. it is shown thatthe information
treatment of non-linear cardiodynamics described here incorporates the multi fractal property of
chaos in the cardiac rhythm.

1 Introduction

The interval between heart beats for a healthy person varies both
systematically with daily activities depending on the metabolic demands as well
as stochastically within such activities. The stochastic aspect is itself a function
of the heart's state rendering the problem inherently nonlinear [1]. A number of
recent studies have sought to describe cardiodynamics in terms of the scaling
properties of the cardiac rhythm [2-6], the existence of a stable attractor [7,8],
approximate entropy [9] and Lyapunov exponents [10].
In this paper we wish to consider an alternative formulation for non-linear
cardiodynamics in terms of statistical mechanics. The heart is a pump whose pulse
rate is linearly related to the energy'demand, i.e. oxygen consumption rate of the
body [11] up to the point of the anaerobic threshold. Accordingly, variation in
heart rate is an analogue for energy changes. Our view of the heart is as a system
varying between different energy levels. Such a view is fundamental to Gibbs'
classical consideration of a thermodynamical system with a probabilistically
defined energy distribution and is basic to any statistical mechanics formulation.
Motivation for this approach comes from some recent analyses of

G. A. Losa et al. (edT.), Fractals in Biology and Medicine

© Birkhäuser Verlag 2002
196 B. Kerman and M Meyer

geophysical data [12] where the measure of energy was intensity difference in
radar imagery of sea ice. It was shown that the naturally occurring negative
exponential distribution of such differences (called the 'Gibbs' property) was
equivalent to a statement of how the entropy, or in the sense of communication
theory, information, could be decomposed into textural and structural information.
In other areas of study, such as mechanics, a comparison of two thermodynamical
variables is referred to as a 'constitutive relationship' or 'equation of state'. The
correlation ofthe two types of information (entropy) within the imagery depended
on the intensity and described such a constitutive relationship. It was further
shown that individuallinear sub-sets of the correlation which indicated multiple
states within the ice imagery related to distinct ice types within a mixture of ice
types. it is suggested the reader consult [12] for essential concepts and operational
details.
It is our goal to develop an energetic measure of the cardiac rhythm and
evaluate its properties for potential characterization of states of the heart's
activity.

1.2 -,----

-
- - - -- --r-------,

-
( .)
Q)
(/)

1a.0
~
Q)

e::::I
0::::
0.6 ' --+-
-1--'- : --+---+-- --f----'---+-------i
o 100 200 300 400 500 600 700
Time (sec)

Figure 1 Ten minute samples of the time interval between

the R pulse of the heart beat of a healthy person sleeping
and in normal daytime activity.

2 Evidence of Statistical Mechanics in Cardiac Rhythm

We consider the heart as a pump which essentially conserves the volume

it displaces in a stroke [11]. While there is some elasticity to the heart, as a first
approximation, the heart regulates its volume flow by altering the rate, h, at
which it operates. Accordingly it is reasonable to characterize the 'energy', e(t),
with time, t, used by the heart as
Information States in Cardiac Rhythm 197

. 1
e(t)=h=- (1)
!1t
where & is the time interval between successive (R-R) beats. Fig. 1 provides
an example of 10 minutes of a time series of heartbeat interval taken from a
healthy person both in sleep and in normal daily activities. As is well-known,
during sleep the average heart beat interval is longer than during the day. In terms
of our model CEq. 1), at night-time the heart is in a lower energy state. Also note
the more regular day-time pattern compared to the more variable night-time
pattern. Accordingly, the texture of the cardiac rhythm is a function of the energy
level. A scaling factor of 0.8 sec was arbitrarily chosen to render a mean energy
level of approximately unity for normal daytime activities.
Consider initially the variability in the cardiac rhythm, as displayed in Fig.
1, over a longer period of 6 hours. In particular, we wish to examine the interval
to interval difference in energy level with time. The difference can be either
positive or negative. it is worthwhile to separate the two in case they have
different statistical properties associated with different factors including the
sympathetic and parasympathetic branches of the autonomic nervous system,
arterial pressure changes, vascular resistance and thermo-regulation.

1 -r-------------------------~

~
:0
~
o 0.1 ..
.....
a..
coc:
o
3),01 -.--
"0
c:
o
()

o. 001 ·i~-- _ _ _+_-----+-----_+_"--_+----i-___i

-0.3 -0.2 -0.1 0 0.1 0.2 0.3

Energy (Heart Rate) Change

Figure 2 Probability distribution of energy (heart rate)

fluctuations for a healthy, sleeping person averaged over
6 hours. Note the transition to a negative exponential
distribution for a fluctuation of about M~.06.

Motivated by basic statistical mechanics and previous results [12], we have

computed the probability of such energy changes. The results for a 6 hour sleep
session are provided in Fig. 2. Three features are apparent. Firstly, there seems to
be no preference in the current sampled sleep sequence for positive over negative
changes in energy level. Further there is a distinct change in the nature of the
198 B. Kerman and M Meyer

fluctuations if the energy increment exceeds approximately 0.06. Thirdly, the

probability distribution for i1e~.06 is log-linear indicating that the probability
of sufficiently large energy differences follows a negative exponential form, i.e.
heart variability conditional on the local heart rate has the Gibbs property.
Such a result has also been noted for variations in processes as diverse as
the stock market [13], company sales [14,15], changes in the GNP of countries
[16] and in sea-ice imagery [12] where a common theme ofan energy measure is
involved.

~0.1~~_!
:E
ro
..0
e
Q.().01
de=.10

1 10 100
Tau

Figure 3 Probability of encountering an absolute

energy change, i1e , at r heart beats later. Note the
apparent scaling behaviour at an interval ..$10 and the
asymptotic time interval invariant condition for ..>! 0
approximately.

It is useful to extend the analysis of energy differences to time intervals

beyond nearest neighbours. We have therefore examined the probability of a given
energy increment at a given number of successive beats, r , i.e. p( i1e,r) . Results
for the sleep session are provided in Fig.3. It can be seen that the nature of
fluctuations is both a function of the energy change and short scaling behaviour
at an interval time up to about 10-15 beats. The probability of comparatively small
energy changes (i1e=O.O 1) decreases with time up to about r=10 while the
probability of large excursions (i1e=O.l) increases with time. For an energy
change of about 0.06 as identified earlier in Fig. 2, a change occurs in Fig. 3 as to
whether the probability is increasing or decreasing in the adjustment interval up
to about r=10.
Both branches reach a time interval invariant state but with probabilities
differing by about an order of magnitude. The convergence to the time interval
Information States in Cardiac Rhythm 199

invariant state follows a scaling relationship as suggested by the outlying straight

lines. it is well-known [17] that the long-range R-R time differences are fractal
(lif noise).It is also reported [8] that there is a small but significant order in short-
range fluctuations within about 8 beats. Fig. 3 reflects that order in as much as it
predicts that a current small heart rate variation will be exceeded within about 10
beats and that the probability of a repeat of a current large variation decreases with
time.
Fig. 3 also indicates that if the time delay between estimates of energy is
chosen to exceed r=10 there is no time dependence associated with selection of
the delay time and no role for the 'critical' energy increment of about 0.06.
Accordingly, we have calculated the probability of a given energy excursion for
r=15 , chosen to clearly approximate the long-range, asymptotic region. The
results are presented in Fig. 4.

0.1 -'.-- - - - - - - - , - - - - - - - - - ,

~
:0
co
...
.0
o
c..
(ij0.01 :
c
o
E
\J
C
o
()

°
0.001
-0.3 -0.2 -0.1 0.1 0.2 0.3
d_e/<d_e>

Figure 4 Probability of an energy fluctuation at time

intervals of 1"=15 beats during sleep.

As anticipated, there is little evidence of any deviation from a negative

exponential distribution except at very small energy changes near the error limit
of the analysis. it is concluded that the sleep-time cardiac rhythm for sufficient
delays is distributed in a scale-invariant, Gibbs manner. Hereafter this limit will
be referred to as the asymptotic condition .. It is also concluded that there exists a
scaling behavior during an adjustment phase of about 10-15 beats which warrants
more investigation.
200 B. Kerman and M Meyer

3 Information States

3.1 Metabolic Variation

During the non-sleeping phase of a day, there can exist periods in which
the heart functions at a higher energy level in the sense of higher average heart
rate. To examine such periods a conditional sampling procedure was established
for a sample 24-hour EeG supplied by A. Goldberger. The statistics of energy
change, both those associated with incremental energy increases and decreases,
were evaluated. The deviation at a given time was that from the average energy
level determined over a period of 100 heart beats symmetrically positioned about
that instant. Only differences for a sufficiently long delay ('r= 15) were
computed.

7,----------------,
6

3
2 +-+-+-+-af-0.....n...;._.+-.0-t.99_~o-+1.-+2_-t-+-
1•4...;1i-+-I-1-t.6-t-+-I---I
3

-10 -8 -6 -4 -2 0 2 4 6 8 10
de/<de>

Figure 5 Probability of an energy fluctuation for a time

interval separation of T=15 for some average energy states
(0.77-\.63) observed on a 24-hour ECG. Note the
synchronized variation of the intercept and slope of the
family of lines.

The results for 5 energy levels (0.77,0.99,1.20,1.41 and 1.63) are shown
in inverted fonn in Fig. 5 for -In p versus t:.e/</:1e>. Thenonna1izingvariable <I:1e>
is the one-sided average difference of energy in each category. Fig. 5 has a
number offeatures. For the most densely populated energy levels (0.77 to 1.20)
there is sufficient data to make a good fit of the linear relationships for both
positive and negative fluctuations. The distributions are very close to being
symmetric for either positive or negative excursions. The best fit lines show little
difference in the intercept point for l1e =0 even when there is more experimental
scatter as with the higher energy states.
Information States in Cardiac Rhythm 201

3.2 Constitutive Relationship

Two properties of Fig. 5 are useful in characterizing the range of energy
states - the slope, [3, and intercept, In Q, of each branch of each energy level
(see Fig. 4). While both vary with the mean energy level they are neither
monotonic nor similar for the same energy level between patients. However the
variable [3 -Iln Q , which in statistical mechanics has properties of external work,
varies systematically with <!J.h> which, as we have postulated above, has the
properties of energy. The results from 24 hour ECGs of normal humans (Fig 6)
show that in fact two linear correlations exist between these energy-like variables
separated at a critical value of <!J.h>=(L065 Hz. Empirically, these relationships
can be represented by
[3-1 In Q= 3.7(<!J.h> -0.65) + 0.19 (2)

for <!J.h>~0.065 Hz and

[3-1 In Q =11.6(<!J.h> -0.65) + 0.19 (3)

for <!J.h>~.065 Hz.

0.4,---------------.

0.3 -- •
ro
+-'
Q)
~.2··
ac
0.1 ...

0.0 f-+---t---i-----+--f---+-+--;.----i
0.00 0.02 0.04 0.06 0.08 0.10
<dh_dot> (Hz)

Figure 6 Variation of the statistical equivalent of external

work /3-1 In Q "in terms of the average variation in heart
rate, <iJJ> , for a range of average heart rates in the
ensemble ofECG records.

Typically, information states [12] are displayed in a graph of structural

information (In Q) versus textural information ([3<!J.h> ). Analysis of these
202 B. Kerman and M Meyer

parameters based on 10 24-hour ECGs obtained in the University Hospital in

Goettingen - all of humans with normal hearts - are presented in Figure 7. Also
shown in Figure 7 are 10 24-ECGs of patients diagnosed with congestive heart
failure. In each case 5 equally probable intervals in the circadian range of
variability were analyzed. Also shown are analyses of 20 multi-hour ECGs of
children and mice.

c:
0
4 ~r

!
I . ".
nOffT',.aj hiJm.a~ chf . children . normal mice

~+
'-I
~~
.81 T•",.t.~T
~T
ro ...:'
'II: 'II

J . ¥;T
.....
......
::l ~~
g!
.' , .,T,
... "," T
~
I
....,
'-
(f) ~. (

o~ T

0 0.5 1 1.5 2
Textural Information

Figure 7 Constitutive relationships for structural

information, In Q, in terms of textural
information, f3<ilh> . Data from 24-hour ECGs of
normal persons and patients with congestive heart
failure, and shorter periods for children and normal
mice.
A number of important results can be seen in Figure 7. Clearly there is a
well normal human range of correlation of structural (S) and textural (T)
information. This correlation appears to have two limbs - an upper one for S>2
which is linear for S>2.5, and a lower limb where T=l. The congestive heart
failure patients define the lower limb and are distinctly different in their behavior
than normal hearts - a fact well known in clinical studies. The second aspect
involving normal human hearts is that the range of children's S and T extends
well beyond the range of normal adult hearts along the upper limb. It can be
shown that the range gradually decreases as the child approaches maturation. Also
the results show conclusively that the behaviour of mice hearts very closely
approaches the behaviour of chaotic variability in human hearts - a fact assumed
but never proven when using mice as a surrogate for humans.
The general form of the relationship - increasing structural information
with decreasing textural information - implies a control over the heart rhythm in
which the narrowing of an exponentially-distributed probability distribution
Information States in Cardiac Rhythm 203

({3<l:!h> increasing) implies that the occurrence rate (Q-I) of any such
fluctuation also increases, and vice versa. In terms of the regulation of the heart
rhythm, the constitutive relationship states that less variable fluctuations in heart
rate are observed more frequently as the heart is stressed. In the limit, no variation
occurs for every beat. Empirically, the constitutive relationship for the upper limb
is given asymptotically by
In Q = 4.1- 1.7 {3<l:!h> (4)

As mentioned above, the relationship for the congestive heart failure

patients occupies a region distinctly removed from that for normal hearts. The
well-known regularity of congestive heart failure is expressed by the limited range
of variation in {3<l:!h> and the lack of structural information (complexity) found
in normal hearts. There is some evidence that the congestive heart patients
represent a bifurcation of the normal heart information state near {3<l:!h>=o.9 .
We note that it is possible to describe the entire probability distribution of
heart rate by
. 1 .
p(l1h) = --exp(-{3<M>l1h) (5)
Q<M>
for a given average variation in heart rate <l:!h>. In Q and {3 can be computed
by solving Eqs. 2 or 3 and 4 iteratively. This may be useful in simulating heart rate
variability.
It remains to understand the non-monotonic and non-unique relationship
. .
of the variations in heart rate, <l:!h>, to the average heart rate , <11>, to
adequately parameterize the normal cardiac rhythm. It also remains to evaluate
these relationships in terms of the actual energy utilized as measured by oxygen
consumption rate. It is also useful to extend the analysis to average heart rates
beyond those encountered by a healthy person during normal daily activities to
verify the constitutive behaviour of In Q for {3<A.h>"? 0.9. Further it is clearly
useful to compare as wide a variety of pathological cardiac rhythms as possible
to determine their discrimination by the statistical mechanics technique.

3.3 Multifractal Property

It has been established recently that the cardiac rhythm [6] is multifractal.
It is also known [19] that a Gibbs measure is often multifractal. A Gibbs measure
[21], p=exp( - {3ilh) , can be constructed from the empirical results for the
conditional probability distributions as a function of h and the local heart rate
variation, ilh, over a sufficient time delay, r . It can be shown that the resulting
time series is multifractal - in fact it has the so-called universal multifractal
structure associated with Levy flight processes [20]. The Levy parameters, a and
204 B. Kerman and M Meyer

C1 ' defined there, are approximately 1.4 and 0.023 for the 24 hour time series of
the Gibbs measure derived from heart rate.

4 Conclusions

In conclusion, cardiac rhythm appears to be succinctly represented in terms

of statistical mechanics, and in particular in terms of states characterized by a
decomposition of information (entropy) within the record. This result is useful in
itself in characterizing chaos in a healthy heart compared to one with congestive
heart failure. The statistical mechanics formulation is compatible with other recent
investigations which have established the multifractal nature of cardiac rhythm.
The Gibbs properties appear to be transitive with multifractality under only weak
constraints.

Acknowledgements

The authors would like to thank Dr A. L. Goldberger of the Harvard

Medical School and Dr H.E. Stanley of Boston University for their support in the
initial stages of this work.

References

[1] A.L. Goldberger, BJ. West, Applications of nonlinear dynamics to

clinical cardiology, Ann New York Acad Sci. 504 (1987) 195-212.
[2] C-K. Peng, 1 Mietus, J. Hausdorif, S. Havlin, H.E. Stanley, AL.
Goldberger, Long-range anticorre1ations and non-Gaussian behaviour of
the heart beat, Phys. Rev. Left. 70 (1993) 1343-1346.
[3] C-K. Peng, S. Havlin, H.E. Stanley, AL. Goldberger, Quantification of
scaling exponents and crossover phenomenon in nonstationary heartbeat
time series, Chaos 5 (1995) 82-87.
[4] P. Ch. Ivanov, M.G. Rosenblum, C.-K. Peng, J. Mietus, S. Havlin, H. E.
Stanley, Scaling behaviour of heartbeat intervals obtained by wavelet-
based time series analysis, Nature 383 (1996) 323-327.
[5] P. Ch. Ivanov, M.G. Rosenblum, C.-K. Peng, lE. Mietus, S. Havlin, H.E.
Stanley, AL. Goldberger, Scaling and universality in heart rate variability
distributions, Physica A 249 (1998) 587-593.
[6] P.Ch. Ivanov, L.A.N. Amaral, AL. Goldberger, S. Havlin, M.G.
Rosenbium, Z.R. Struzik, H.E. Stanley, Multifractality in human heartbeat
dynamics, Nature 399 (1999) 461-465.
[7] AL. Goldberger, D.R. Rigney, BJ. West, Chaos and fractals in human
physiology, Sci. Am. 262 (1990) 42-49.
[8] J.H. Lefebvre, D.A Goodings, M.V. Kamath, E.L. Fallen, Predictability
of normal heart rhythms and deterministic chaos, Chaos 3 (1993) 267-276.
[9] S.M. Pincus, AL. Goldberger, Physiologic time-series analysis: what does
Information States in Cardiac Rhythm 205

regularity quantify?, Am. 1. Physiol. 226 (1994) HI643-1656.

[10] A Wolf, 1.B. Swift, H.L. Swinney, 1.A Vastano, Determining Lyapunov
exponents from time series, Physica D 85 (1985) 285-3 17.
[11] K. T. Weber, Gas transport and the cardiopulmonary unit, in K.T. Weber,
1.S. Janicki (Eds.), Cardiopulmonary Exercise Testing, Saunders, (1986)
30.
[12] B.R. Kerman, Information states in sea ice imagery, IEEE Trans. Geophys.
Remote Sens. 37 (1999) 1435-1446.
[13] R. N. Mantegna, H.E. Stanley, Scaling behavior in the dynamics of an
economic index, Nature 376 (1995) 46-49.
[14] M.H.R. Stanley, L.A.N. Amaral, S. V. Buldyrev, S. Havlin, H. Leschhom,
P. Maass, M. A Salinger, H.E. Stanley, Nature 379 (1996) 804-806.
[15] L.AN. Amaral, S.V. Buldyrev, S. Havlin, H .. Leschhorn, P. Maass, M. A.
Salinger, H.E. Stanley, M.H.R. Stanley, Scaling behavior in economics:
1. Empirical results for company growth, 1. Phys I France 7 (1997) 621-
633.
[16] Y. Lee, L.AN. Amaral, D. Canning, M. Meyer, H.E. Stanley, Universal
features in the growth dynamics of complex organizations, Phys. Rev.
Left. 81(1998) 3275-3278.
[17] M. Meyer, Scaling properties of heartbeat interval fluctuation in health
and disease, in M.N. Novak (Ed.), Fractals and Beyond, World Scientific,
(1998) 33-42.
[18] L. Lipsitz, AL. Goldberger, Loss of 'complexity' and aging, 1. Am. Med.
Assoc. 267 (1992) 1806-1809.
[19] Y. Pesin, H. Weiss, The multifractal analysis of Gibbs measures:
Motivation, mathematical foundation and examples, Chaos 7 (1997) 89-
106.
[20] T. Falco, F. Francis, S. Lovejoy, D. Schertzer, B. Kerman, M. Drinkwater,
Universal multifractal scaling of synthetic aperture radar images ofsea ice,
IEEE Trans. Geophys. Remote Sens. 34 (1996) 906-9 14.
[21] B.R. Kerman, K. Johnson, Properties of a probability measure for sea ice
imagery, GlobalAtmos. Ocean Sys. 6 (1998) 35-92.