P1: SGR

Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Journal of Computational Neuroscience 5, 35–51 (1998)

°
c 1998 Kluwer Academic Publishers. Manufactured in The Netherlands.

Entrainment, Instability, Quasi-periodicity, and Chaos

in a Compound Neural Oscillator

MASAKAZU MATSUGU
Imaging Research Center, Canon Inc., Ohta-ku, Tokyo, 146, Japan

JAMES DUFFIN
Department of Physiology, University of Toronto, Toronto, Ontario, Canada M5S 1A8

CHI-SANG POON∗
Harvard-MIT Division of Health Sciences and Technology,
Massachusetts Institute of Technology, Cambridge, MA 02139, USA
[email protected]

Received June 6, 1996; Revised May 22, 1997; Accepted June 4, 1997

Action Editor: John Rinzel

Abstract. We studied the dynamical behavior of a class of compound central pattern generator (CPG) models
consisting of a simple neural network oscillator driven by both constant and periodic inputs of varying amplitudes,
frequencies, and phases. We focused on a specific oscillator composed of two mutually inhibiting types of neuron
(inspiratory and expiratory neurons) that may be considered as a minimal model of the mammalian respiratory rhythm
generator. The simulation results demonstrated how a simple CPG model—with a minimum number of neurons
and mild nonlinearities—may reproduce a host of complex dynamical behaviors under various periodic inputs.
In particular, the network oscillated spontaneously only when both neurons received adequate and proportionate
constant excitations. In the presence of a periodic source, the spontaneous rhythm was overriden by an entrained
oscillation of varying forms depending on the nature of the source. Stable entrained oscillations were inducible
by two types of inputs: (1) anti-phase periodic inputs with alternating agonist-antagonist drives to both neurons
and (2) a single periodic drive to only one of the neurons. In-phase inputs, which exert periodic drives of similar
magnitude and phase relationships to both neurons, resulted in varying disruptions of the entrained oscillations
including magnitude attenuation, harmonic and phase distortions, and quasi-periodic interference. In the absence
of significant phasic feedback, chaotic motion developed only when the CPG was driven by multiple periodic inputs.
Apneic episodes with repetitive alternation of active (intrinsic oscillation) and inactive (cessation of oscillation)
states developed when the network was driven by a moderate periodic input of low frequency. Similar results were
demonstrated in other, more complex oscillator models (that is, half-center oscillator and three-phase respiratory
network model). These theoretical results may have important implications in elucidating the mechanisms of
rhythmogenesis in the mature and developing respiratory CPG as well as other compound CPGs in mammalian and
invertebrate nervous systems.

Keywords: nonlinear dynamics, respiratory rhythmogenesis, apnea, pacemaker, half-center oscillator, periodic
forcing, in-phase inputs, out-phase inputs

∗ To whom all correspondence should be addressed.

P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

36 Matsugu, Duffin and Poon

1. Introduction cortical (Haas et al., 1986), chemical (Takahashi et al.,

1992), or mechanical (Graves et al., 1986) afferent in-
In animal nervous systems, CPGs are composed of puts, or by locomotory activity (Bechbache and Duffin,
specialized groups of neurons that produce rhythmic 1977; Bramble and Carrier, 1983; Corio et al., 1993).
activities (Kopell, 1988). In general there are two ba- Little is known about how these inputs influence the
sic mechanisms for ryhthmogenesis: the spontaneous respiratory CPG when their strengths are below the
bursting activity of pacemaker cells or oscillations pro- threshold for entrainment.
duced by the collective interaction within a network It has also been suggested that a periodically driven
of nonpacemaker neurons (Friesen and Stent, 1977; neural oscillator is functionally equivalent to a het-
Delcomyn, 1980; Pinsky and Rinzel, 1994; Skinner eroassociative memory in that the unique spatiotem-
et al., 1994). Many CPGs are built on either of these poral activity pattern induced by the periodic source
mechanisms, but there are exceptions where both co- represents a form of information storage (Matsugu
exist. For example, some invertebrate CPGs, such as and Poon, 1993). Thus, compound CPGs with ei-
the leech cardiac and the lobster gastric mill rhythms, ther endogenous or exogenous periodic inputs may be
are composed of a mixture of endogenous pacemak- important for many physiological functions that can-
ers and network oscillators (Selverston and Mazzoni, not be subserved by pacemakers or network oscilla-
1989). tors alone. In this study, we examined the behaviors
Respiratory rhythm in vertebrates may be another of several CPG models under both constant and peri-
such example of a compound CPG (Duffin et al., 1995). odic inputs with varying amplitudes, frequencies, and
Neurons with conditional bursting properties have re- phases. These inputs correspond, respectively, to the
cently been identified in the brainstem of the neonatal constant excitations from chemoreceptors and the pe-
rat (Arata et al., 1990; Smith et al., 1991) among other riodic drives from pacemakers or rhythmic afferents.
neuronal groups traditionally thought to underlie res- Although the general theory of entrainment of non-
piratory rhythmicity (Richter et al., 1986). In the adult linear oscillators has been extensively investigated
rat respiratory rhythm is likely to be generated by a (Ermentrout, 1982; also see review by Glass and
network oscillator because the rhythm is interrupted Mackey, 1988) with specific application to the respi-
by blocking inhibitory synapses (Hayashi and Lipski, ratory system (Petrillo and Glass, 1984), these stud-
1992). It has therefore been suggested that either the ies are not specifically relevant to the analysis of
medullary respiratory rhythm generator undergoes a rhythmogenesis in a neural network oscillator un-
developmental change from a perinatal pacemaker to der conditions of periodic perturbations below the
an adult network oscillator (Hayashi and Lipski, 1992) entrainment threshold. Similarly, previous studies
or that both mechanisms operate together as a hybrid of synchronization among coupled phase oscillators
oscillator with each dominating at different stages of (Kopell and Ermentrout, 1988; Collins and Stewart,
development (Richter et al., 1992; Smith, 1996). 1994; Kazanovich and Borisyuk, 1994) and the mod-
It is not clear how the behavior of such compound eling of coupled pacemakers (Cohen et al., 1992; Cym-
neural oscillators differs from those of isolated pace- balyuk et al., 1994) are also not specific to the presently
makers or network oscillators, although it is reasonable proposed model of respiratory CPG under the influence
to expect that the contribution of each component to of a periodic source.
the resultant oscillatory pattern would be dependent on Although several neural network models for respira-
their relative strengths. At one extreme, the activity of tory rhythmogenesis have been proposed (Richter et al.,
a pacemaker component might have little effect if it is 1986; Botros and Bruce, 1990; Sammon and Bruce,
too weak or the network is strongly excited, while at 1991; Ogilvie et al., 1992; Gottschalk et al., 1992,
the other extreme a vigorous pacemaker activity could 1994; Smith, 1996; Balis et al., 1994), few investiga-
entrain the network oscillator. Less certain is the be- tors have explored the properties of these models when
havior of a compound oscillator when the pacemaker driven by external inputs. How the respiratory CPG
and network components are of similar potency. processes neural information from exogenous and en-
Network CPGs may also be influenced by periodic dogenous inputs is a question of fundamental impor-
inputs from exogenous sources other than pacemakers. tance in understanding the mechanism of homeostatic
For example, the respiratory CPG in humans and other regulation of vital physiological variables in health and
mammals has been shown to be entrained by rhythmic disease states (Poon, 1995). Sammon and Bruce (1991)
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 37

and Gottschalk et al. (1992) investigated the behavior its decrementing activity during a burst (Matsuoka,
of various respiratory CPG models with synchronous 1985). It has been shown that the early-burst inspi-
phasic feedbacks from pulmonary stretch receptors and ratory neurons have inhibitory projections to the expi-
demonstrated the induction of chaos under this con- ratory Bötzinger complex neurons respectively (Duffin
dition. To investigate whether asynchronous periodic and Douse, 1993) and vice versa (Ezure and Manabe,
inputs may induce complex nonlinear dynamics inde- 1986). Alternatively, respiratory oscillation may arise
pendently of phasic feedbacks, we therefore examined from reciprocal inhibition between early-burst inspi-
the behaviors of several respiratory CPG models with ratory neurons and decrementing expiratory neurons.
periodic excitations but devoid of phasic feedbacks, However, the connections between these neuronal
thereby simulating the respiratory rhythms in humans groups have not been totally verified. Furthermore,
in the awake resting state (Widdicombe, 1961; Guz and such a network configuration would not represent a
Trenchard, 1971) or in REM sleep (Phillipson, 1986), minimal model since it requires adaptations in both re-
where vagal stretch receptors are weak, or in vago- ciprocally inhibiting neurons.
tomized animals (Lahiri et al., 1975; Takahashi et al., The inputs driving the I and E neurons originate
1992). from a source D and represent steady drives from
Our simulation results revealed several conditions chemoreceptors or other excitatory afferents, and peri-
that may be crucial for the maintenance of stable spon- odic drives from either endogenous pacemaker cells or
taneous or entrained oscillations in a network CPG periodically active afferents such as those associated
driven by constant and/or periodic inputs. These find- with locomotory activity.
ings may have important implications in understanding The model neurons may be thought of as either rep-
the mechanisms of rhythmogenesis in the mature or de- resentative single cells of a population, where the ac-
veloping respiratory CPG in mammals or other com- tivity represents membrane potential, or populations of
pound CPGs in many vertebrates and invertebrates. similar cells, where activity represents the net activity
of the whole population. For the purposes of numer-
ical simulation, the activity of the neurons and their
2. Central Pattern Generator Models connection strengths (see Fig. 1) were expressed in
arbitrary units. Thus, the amplitude and frequency of
2.1. Minimal Respiratory CPG Model the network oscillation may be calibrated for specific

Two types of respiratory CPG models have been pro-

posed: two-phase models that are characterized by dis-
tinct inspiratory and expiratory phases (Duffin, 1991)
and three-phase models that are also characterized by a
third, intermediate phase called the postinspiratory (or
stage I expiratory) phase (Richter et al., 1986). Neither
model has been conclusively proven. However, for the
present initial analysis a simple two-phase model was
chosen as a minimal model. Applicability of the results
to more complex models was tested by simulations with
a more complex two-phase model (Skinner et al., 1994;
Appendix A) and a previously published three-phase
model (Gottschalk et al., 1994; Appendix B).
Figure 1. Minimal network oscillator model consisting of mutually
A minimal two-phase model of respiratory rhythmo-
inhibiting neurons I and E, which are excited by driving source D.
genesis (Fig. 1(a)) is composed of two mutually inhibit- Virtual interneuron F provides an intrinsic adaptation function for
ing neurons: I and E, corresponding to the early-burst neuron I . Excitatory and inhibitory synapses are shown as open and
inspiratory neurons and expiratory Bötzinger complex filled circles, respectively. Numbers denote the connection strengths,
neurons respectively (Duffin, 1991). A virtual interneu- which produce a stable spontaneous oscillation with suitable DC
inputs from driving source D. Other parameter values are as follows:
ron F, which provides negative recurrent feedback for
Hi = 0 (i = I, E, F), R I = R E = −10, R F = 0, TI = TE = 0.08,
the I neuron, accounts for the adaptation properties TF = 1.5. All simulations were performed using MATLAB, and
of the I neuron (probably due to a calcium-activated solutions of differential equations were obtained by using the third-
potassium current; Mifflin et al., 1985) that lead to order Runge-Kutta method with adaptive step size.
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

38 Matsugu, Duffin and Poon

dimensional units by suitable scaling. The feedback phase relationships of the periodic inputs to the I and
neuron F can simulate the adaptation effect in neuron I E neurons. As summarized below, the bifurcation pat-
without causing self-induced oscillations provided the terns exhibited several interesting characteristics of in-
following inequality is met (Matsuoka, 1985): creasing complexity.

(TF − TI,E )2 ≥ 4TF TI,E C F ,

where TF is the time constant for neuron F (1.5 s), TI,E 3.1. Spontaneous Oscillations Under Steady
is the time constant for the I and E neurons (0.08 s), (DC) Inputs
and C F is the connection strength from neuron F to
neuron I. Figure 2 shows typical waveforms in the activities
The equation describing the activity of any neuron i of the mutually inhibiting neurons in three types of
in the model is (Duffin, 1991): models (that is, Duffin model, Gottschalk model, and
a half-center oscillator model) under DC inputs. In
d xi X the Gottschalk model and half-center oscillator model,
Ti + x i = Ri + Ci j g(x j − H j ) + Si (1)
dt both the time constant and amplitude of neural activities
j
were rescaled for convenience of graphic inspection,
where xi is the current activity of the ith neuron; Ri while the ratios of DC inputs to the network neurons
is its resting activity (−10 for the I and E neurons, 0 were kept the same as the original models. The wave-
for neuron F); Hi is its threshold (0 for all neurons); forms of the Duffin model and Gottschalk model are
Ci j is the strength of the connection from the jth to similar to those originally reported by Duffin (1991)
the ith neuron; and g(z) is a nonlinear function (e.g., and Gottschalk et al. (1994). In the half-center oscil-
g(z) = max(0, z), or g(z) = tanh(az)) used to model lator, the waveforms of both neurons are identical (ex-
the activation threshold for the model neurons. In keep- cept for a 180 deg phase shift), in conformance with
ing with parsimony, only the former expression was the symmetric network configuration.
used which is the simplest form of nonlinearity nec- Using a steady driving source D (A I = A E = 0),
essary for rhythmogenesis (Botros and Bruce, 1990; we examined the oscillation patterns of the model at
Duffin, 1991; Gottschalk et al., 1992, 1994; Ogilvie varying DC levels. It was found that the ratio of the
et al., 1992). DC inputs to the I and E neurons, DC E : DC I , was
For simplicity, the inputs to oscillator neurons I and critical for spontaneous oscillation to occur. Specif-
E from driving source D (denoted S I and S E , respec- ically, limit-cycle oscillation developed only when
tively) are assumed to be of the form the ratio was within the range DC I > DC E > 0.7DC I
(Fig. 3(a)). Over this range, the frequency of oscilla-
S I = A I sin(2π f I t + φ I ) + DC I (2) tion ( f spontaneous ) varied only slightly between 0.3 and
0.4 Hz (18 to 24 breaths per minute) except toward the
S E = A E sin(2π f E t + φ E ) + DC E , (3)
bifurcation points (DC E = 0.7DC I , DC I ), where the
oscillation frequency decreased rapidly.
where DC is the steady component and A, f, φ are
At a chosen DC E /DC I ratio of 0.8 f spontaneous was
respectively the amplitude, frequency, and phase angle
unaffected by the magnitude of DC providing a min-
of the sinusoidal component of the signal to the I or
imum level (DC E = 16.8) was exceeded (Fig. 3(b)).
E neurons (with corresponding subscripts). Sinusoidal
Both the required minimum DC levels and their critical
inputs to the I and E neurons are said to be in-phase
ratio for spontaneous oscillation of the model were de-
or anti-phase if f I = f E and φ I − φ E = 0 or 180◦ ,
pendent on the connection strengths chosen. The gen-
respectively (Cohen et al., 1992; Cymbalyuk et al.,
eral finding was that, for a network oscillator consisting
1994).
of mutually inhibiting neurons, the initiation of oscil-
lation is critically dependent on the strengths of the
3. Results driving inputs to the neurons relative to the strengths
of the mutual inhibitions, but the frequency of sponta-
The behavior of the network CPG model was criti- neous oscillation is mainly determined by the connec-
cally dependent on the strengths of the DC and peri- tion strengths of the network, its adaptive properties
odic input components as well as the frequency and and, to a lesser extent, the DC E /DC I ratio.
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 39

Figure 3. (a) Spontaneous frequency of the network oscillator as

a function of the ratio of the DC input to each neuron (DC E /DC I )
in the absence of oscillatory inputs (that is, A I = A E = 0). The
absolute value of the frequency is calibrated for a particular ex-
perimental preparation by suitable scaling of the model parame-
ters. Note that the spontaneous frequency varies only slightly with
the DC ratio except near the bifurcation points. (b) Limit cy-
Figure 2. Waveforms of spontaneous oscillation. (a) Duffin model. cles of spontaneous oscillations under purely DC inputs with con-
Solid line represents activity of I (early-inspiratory) neuron, dashed stant DC ratio (DC E /DC I = 0.8) and increasing strength. The
line is the activity of E (expiratory) neuron. DC input strengths to limit-cycle trajectories in this and subsequent figures are syn-
I and E neurons are 50 and 40, respectively. (b) Gottschalk model. thesized from the activity of neuron E (ordinate) and I (ab-
Solid and dashed lines show early-inspiratory (DC level: 1.0) and scissa). Increasing the DC level of the driving input D resulted
expiratory (DC level: 1.1) neurons, respectively. (c) Half-center in an augmentation of the magnitude of response but had lit-
oscillator. The DC input strengths are 0.35 for the two neurons, and tle effect on either the waveform or frequency of the induced
other parameter values are shown in Appendix A. oscillation.
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

40 Matsugu, Duffin and Poon

3.2. Entrainment to Strong Periodic Inputs (1 : 1) entrained to the driving frequency but the limit-
cycle trajectories exhibited a progressive transforma-
3.2.1. In-phase Inputs. Using a sinusoidally varying tion with drastic changes in character as the driving
driving source D (A I = A E = 50) with the drives to frequency was varied (Fig. 4(a)). At driving frequen-
the I and E neurons in-phase (φ I − φ E = 0; f I = f E ), cies between 1.5 and 2.0 Hz (that is, approximately
and DC E /DC I = 0.8(DC I = 50), we examined the five times the spontaneous frequency), the limit-cycle
oscillation patterns of the model at varying sinusoidal trajectory was a simple loop with the major and mi-
driving frequencies. The patterns of oscillation are dis- nor axes pointing in the horizontal and vertical direc-
played as limit-cycle trajectories by plotting E neu- tions, respectively, indicating that the I and E neu-
ron activity versus I neuron activity in the phase plane rons were nearly 90◦ out of phase (that is, somewhat
(Fig. 4(a)) and as continuous time functions (Fig. 4(b)). between in-phase and anti-phase). Thus, in the in-
For driving frequencies from 0.15 to 2.0 Hz (9 to phase input configuration the I and E neurons were
120 breaths/min), each neuron of the model was totally entrained to the periodic input with differing phase
shifts. As the driving frequency was decreased to-
ward the spontaneous frequency, the limit-cycle tra-
jectory deformed continuously until, at a frequency of
∼1.2 Hz, a cusp was formed that gradually turned into
a twist when the driving frequency was decreased to
0.8 Hz (that is, approximately two times the sponta-
neous frequency). At lower driving frequencies, the
limit-cycle trajectory again deformed continually until
at a frequency of ∼0.43 Hz (that is, near the spon-
taneous frequency) the twist disappeared and a sim-
ple loop returned. As shown in Fig. 4(a) at f ≤ 0.5
Hz the major axis of the loop was tilting to the left
indicating that the oscillations of the I and E neu-
rons were anti-phase, unlike the in-phase input. As
the driving frequency was decreased still further, a
cusp reappeared at ∼0.33 Hz and was fully developed
as the driving frequency fell below the spontaneous
frequency.
As Fig. 4(b) illustrates, the deformation of the limit-
cycle trajectory was the result of changes not only in
the waveforms of the I and E neurons activity but also
in their relative phases of entrained oscillation. For
example, the twist in the phase trajectory was due to
the presence of a prominent second harmonic in the
activity of the E neuron, with a phase relationship such
that the two peaks in the E neuron activity in each cycle
Figure 4. Entrained oscillations with in-phase sinusoidal inputs fell on either side of the trough in the I neuron activity
(A I = 50; A E = 40 units) and constant DC inputs (DC E /DC I = in the same cycle. Similarly, the cusp in the trajectory
0.8; DC I = 50) from driving source D. (a) Limit-cycle trajectories
showing marked changes in behavior of entrained oscillation includ-
resulted from a less prominent second harmonic, which
ing the appearance and disappearance of a twist and cusp in the loop induced an inflection point in the E neuron activity
and its transformation from a circular (90◦ phase difference between coincident with the trough of the I neuron activity.
the I and E neurons) to a leftward (180◦ ) orientation as the driving The amplitude of the entrained oscillation remained
frequency decreased toward the spontaneous oscillation frequency. relatively stable at low driving frequencies. How-
(b) Activities of the E (solid lines) and I (dashed lines) neurons ver-
sus time in seconds illustrating the changes in waveforms and phase
ever, at driving frequencies above ∼1.0 Hz the am-
relationships of neural activity that produced the twist and cusp in plitude decreased progressively with increasing drive
the phase plane. frequency (Fig. 5).
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 41

Figure 5. Amplitude of entrained oscillations as a function of driv-

ing frequency. Driving input conditions are as in Fig. 4. Note that
Figure 6. Entrained oscillations under similar conditions as in
entrained oscillation is much stronger under anti-phase input than in-
Fig. 3 but with anti-phase inputs. Note that in this case the limit-cycle
phase input. In both cases, the amplitude of entrained oscillation is
trajectories have similar shapes and phase relationships (∼180◦ ) over
maximum over a range of driving frequencies near the spontaneous
a wide range of driving frequencies.
frequency of the network, and the response decreases with increasing
departure of driving frequency from the optimal frequencies.
fact that in both cases only one neuron is excited at
any given time. The stability of the oscillation pat-
3.2.2. Anti-phase Inputs. Using the same sinu- tern in response to anti-phase inputs was also observed
soidally varying driving source D as described above, (unpublished data) for more complex models such as
except that the inputs to the I and E neurons were anti- the three-phase respiratory CPG model (Gottschalk et
phase (φ I − φ E = 180◦ ), we examined the oscillation al., 1994) with oscillatory inputs to early-inspiratory
patterns of the model at varying sinusoidal driving fre- and/or expiratory neurons, and a half-center, bursting
quencies. Over the same range of driving frequencies oscillator (Skinner et al., 1994) with similar oscillatory
(0.15 to 2.0 Hz or 9 to 120 breaths/min) the amplitude of drives.
the entrained oscillation was significantly greater (more
than double) than that resulting from the in-phase in- 3.2.3. Critical Amplitudes for Entrainment. For
puts (Fig. 5). The maximum response occurred at a both in-phase and anti-phase periodic inputs (with
driving frequency of ∼0.6 Hz (that is, slightly higher f I = f E ) the forced oscillation of the network was
than the spontaneous frequency of 0.4 Hz), and the found to be totally (1 : 1) entrained to the driving fre-
response decreased rapidly with increasing departure quency when the amplitude of the periodic input was
of the driving frequency from the optimal frequency. sufficiently large. The critical amplitude, Ac , for total
At all driving frequencies the limit-cycle trajectory re- entrainment to occur was found to be related to its DC
mained relatively stable without appreciable changes in level: the higher the DC level, the greater the amplitude
character (that is, appearance of twists and cusps and required to induce total entrainment. Not surprisingly,
changes in phase angles between the I and E neuron for DC levels well below the threshold for spontaneous
activities) as observed for the in-phase inputs (Fig. 6). oscillation, the model was always totally entrained re-
Similar stability of the oscillatory pattern was ob- gardless of the input amplitude. Figure 7(a) shows the
served when the sinusoidal input was applied only critical amplitude Ac versus the DC level of input D for
to one of the mutually inhibiting neurons (for exam- in-phase inputs with frequencies of 0.5 and 1 Hz (30
ple, I or E), in conformance with cross-correlational and 60 breaths/min).
data (Balis et al., 1994). The similarity of the response In addition, it was found that if the periodic driv-
behaviors corresponding to the anti-phase and single- ing input was very strong, it could induce oscillation
ended inputs should not be surprising considering the in the model even under conditions that precluded
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

42 Matsugu, Duffin and Poon

oscillated at its spontaneous frequency, while above

Ac the model was entrained 1 : 1 to the driving fre-
quency. The values of Amin and Ac increased with
increases in the driving frequency. For example, with
DC E /DC I = 0.8(DC I = 50) and a driving frequency
of 1.2 Hz the critical amplitudes for in-phase inputs
were found to be Amin = 2.5 and Ac = 25, whereas at
a driving frequency of 5 Hz the critical values were 15
and 45 respectively. The value of Amin was generally
lower for in-phase than anti-phase inputs; for Ac the
converse was true.
The resonance structure (Alexander et al., 1990) of
the network CPG model was investigated by observing
the rotation numbers resulting from various driving in-
put amplitudes and frequencies for both in-phase and
anti-phase inputs to neurons I and E . Figure 8 shows
the Arnold tongues resulting from driving inputs with
DC E /DC I = 0.77(DC I = 49.5). It should be noted
that changes in the DC level may alter the resonance
structure since Ac is dependent on DC. As Fig. 8 il-
lustrates, the critical amplitude for entrainment was
generally lower for anti-phase inputs than for in-phase
inputs.

3.3. Instability Under Low-Frequency

Periodic Inputs

3.3.1. Modes of Instability. A remarkable behavior

emerged when the network of Fig. 1 was fed by periodic
inputs with intermediate amplitude and sufficiently low
frequency (for example, less than one-tenth of the in-
Figure 7. Entrainment of the model oscillation by in-phase sinu- trinsic frequency). As is evident from Fig. 9, the osc-
soidal inputs. (a) Critical amplitude for 1 : 1 entrainment plotted
illation pattern in this case is repetitively interrupted
against DC I (DC E /DC I = 0.8) for two sinusoidal input frequen-
cies. (b) Activities of E (solid lines) and I (dashed lines) neurons by periods of silence (that is, cessation of oscillation),
versus time in seconds during entrainment when the DC ratio vio- resulting in periodic oscillation. The frequency of oc-
lated the conditions for spontaneous oscillation. Driving frequency currence of the apneic episodes is determined by the
was 1 Hz, and DC levels were 50 units for both neurons. The model frequency of the periodic input, whereas the frequency
oscillated at 1 Hz and 0.125 Hz for driving input amplitudes of 100
of the oscillatory pattern is identical to the spontaneous
and 10 units respectively.
frequency (about 0.4 Hz), which is determined by the
ratio of the DC components (see Fig. 3).
the development of spontaneous oscillation (for ex- Generally, periodic breathing is induced only when
ample, DC E > DC I , or DC E < 0.7DC I ), particularly the amplitude of the periodic input falls within a certain
when the driving input was anti-phase. It was also intermediate range (Amin max
apnea , Aapnea ), providing the input
found that even if the driving input D did not at- frequency is sufficiently low. Below Amin apnea the network
tain the critical amplitude for total entrainment, partial oscillation is not affected whereas above Amax apnea the net-
entrainment could still occur provided a minimum in- work is totally entrained by the periodic input. The val-
put amplitude, Amin , was exceeded. In this event the ro- ues of the lower and upper limits are dependent on the
tation number (ratio of driving frequency to entrained values of DC I and the DC ratio, DC E /DC I . This be-
frequency) of the resulting oscillation was a rational havior may be readily understood if one treats the slow
number of simple ratio. Below Amin the network model input as a changing parameter that takes the system
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 43

Figure 9. Apneic oscillation pattern of (a) I neuron and (b) E

neuron. The periodic input (frequency, 0.02 Hz; amplitude, 0.5) is
anti-phase with DC components DC I = 50.0 and DC E = 49.9.
Both neurons exhibit alternating transitions between spontaneous
oscillation periods and silent periods.

Case II. When the DC ratio is close to the limiting

values (that is, 0.7 or 1) for spontaneous oscillation,
periodic breathing may be elicited even by periodic
inputs with relatively small magnitudes. This is the
marginally stable configuration. For example, when
DC I is 50 and DC E is 49.9 (that is, DC E /DC I ' 1),
an input amplitude in the range (0.1, 10) is sufficient to
induce the apneic behavior. Furthermore, in this case
Amin max
apnea and Aapnea are directly proportional to the DC
strength. For example, with DC I = 50, DC E = 49.9
we have Amin apnea ' 0.1 and Aapnea ' 10, and with each
max

of the DC components halved we have Amin apnea ' 0.05

Figure 8. Resonance structure of the model. Arnold tongues ob- and Amax
apnea ' 5.
tained by plotting the critical amplitude for total (1 : 1) and partial en-
trainment against driving frequencies of in-phase (a) and anti-phase Case III. When the DC ratio is beyond the range
(b) sinusoidal inputs. Note that the critical amplitudes for entrain-
ment are generally greater for in-phase inputs than for anti-phase
required for spontaneous oscillation, the network is
inputs. Both figures show the phase locking zones of 1 : 1, 2 : 1, and inherently apneic and periodic breathing may be in-
3 : 1, and there are other zones with more complex rotation numbers duced with sufficiently large input amplitudes. For ex-
between those with simple rotation numbers. ample, for DC I = 50 and DC E /DC I = 1, we have
apnea ' 0.005 and Aapnea ' 50.
Amin max

The duration of apnea was also found to be depen-

in and out of the oscillatory regime of the bifurcation dent on the amplitude of the periodic input. Generally,
diagram in Fig. 3. Several cases may be distinguished. the apneic duration may either increase (Case I) or
decrease with the input amplitude (Case II or III), de-
pending on the DC ratio. For example, in Case I with
Case I. When the DC ratio is well within the range the same DC components as those in Fig. 10, the silent
(0.7, 1) for spontaneous oscillation to occur (Fig. 3), period is much shorter than the breathing (that is, ac-
induction of periodic breathing requires relatively tive) period for amplitudes less than 5, whereas with in-
large input amplitudes. This is the most stable net- creased amplitudes (for example, A > 25), the oscilla-
work configuration. For example, for DC I = 50 and tion becomes less prominent (Fig. 10(d)). Conversely,
DC E /DC I = 0.8, Amin
apnea ' 20 (see Fig. 10). in Case III, the duration of apnea Tap and that of bursting
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

44 Matsugu, Duffin and Poon

Figure 10. Transition of apneic episode in I neuron. The frequency Figure 11. Apneic episode of early inspiratory neuron in the forced
of the anti-phase input is 0.02 Hz with input amplitudes: 5 (a), 15 (b), Gottschalk model. Anti-phase oscillatory inputs were given to early-
25 (c), and 50 (d). The DC components of the input are DC I = 50 inspiratory and expiiratory neurons. Input frequency (0.067 Hz ) is
and DC E = 40. Changes in the input amplitude result in different much smaller than the intrinsic frequency of the original Gottschalk
apneic patterns with varying duration of apnea phase. With stronger model (approximately 0.6 Hz). We observe distinct phases of inac-
input, the network tends to be entrained to the input, and with weaker tive state that are alternately interrputed by active, osciilatory phase.
input the network is inclined to self-oscillate.

when the amplitude of the anti-phase input (frequency

Tb were Tap = 25, Tb = 25 for an input amplitude A
0.1 Hz) was increased. The behaviors in Fig. 12(a) and
of 0.5, while we obtained Tap = 35, Tb = 15, for
(c) are somewhat transitory between the spontaneous
A = 0.005 (in both cases, DC I = DC E = 50).
oscillation mode and apnea, and between the apneic
episode and entrainment, respectively. These results
3.3.2. Instability in Other CPG Models. Next, we ex-
suggest the generality of the conditions for the induc-
plored possible apneic behaviors in a three-phase CPG
tion of apneic episodes by low-frequency periodic in-
model (Gottschalk et al., 1994) and a half-center CPG
puts.
model with bursting neurons (Skinner et al., 1994) un-
der similar driving conditions.
Figure 11 shows the activity of the early-inspiratory 3.4. Quasi-periodicity and Chaos Under
neuron in the forced Gottschalk model (Appendix B) Intermediate Periodic Inputs
with anti-phase inputs to the early-inspiratory and ex-
piratory neurons (frequency = 0.067 Hz, amplitude = Under certain conditions, chaotic (strongly turbulent)
1.0). The simulation result clearly indicates repetitive and quasi-periodic (weakly turbulent) oscillations were
alternations of active and inactive (apneic) states. The provoked when the network oscillator model was
switching from active to inactive (or vice versa) state driven by either in-phase or anti-phase periodic inputs.
is 1 : 1 phase-locked to the external drive, and the os-
cillation frequency in the active period is exactly the 3.4.1. Quasi-periodicity. Quasi-periodicity was typi-
intrinsic frequency (0.6 Hz) of the original (unforced) cally provoked when (1) the driving input amplitude
Gottschalk model. was intermediate between Amin and Ac and (2) the
In the forced Skinner model (Appendix A), the two driving frequency ( f I = f E ) was between the dom-
mutually inhibitory neurons were fed by similar sinu- inant (total or partial) entrainment modes as defined
soidal inputs. Again, apneic episodes were induced by the Arnold tongues, particularly when the driv-
under similar driving conditions of low frequencies ing frequency was much higher or lower than the
and intermediate amplitudes. Figure 12 shows a tran- spontaneous oscillation frequency. For example, with
sition of oscillation modes in the forced Skinner DC E /DC I = 0.8(DC I = 50) and A I = A E = 50,
model from recurrent apnea (Fig. 12(a), amplitude in-phase inputs to the I and E neurons produced a quasi-
10; Fig. 12(b), amplitude 15) to total entrainment periodic oscillation primarily when the input frequency
(Fig. 12(c), amplitude 20; Fig. 12(d), amplitude 50) exceeded 2.15 Hz (129 breaths per minute), much
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 45

Figure 13. Model behavior for in-phase sinusoidal inputs of 3 Hz

with varying amplitudes and constant DC excitation DC E /DC I =
0.8; DC E = 50), showing transitions in model oscillation modes.
Spontaneous oscillation results for an input amplitude of 0.5 units
(a); quasi-periodic oscillation results for an amplitude of 1.5 units (b)
and 21 units (c); and entrainment results for an amplitude of 50 units
(d). Note that the phase loop tilts to the left at low input amplitudes
indicating an anti-phase relationship between the I and E neurons,
and to the right at high amplitudes indicating in-phase relationship.

a driving frequency well below the spontaneous fre-

quency of the network. These behaviors could also
be observed for sigmoidal nonlinearity (for example,
g(z) = z/(1 + exp(−bz))) in the activation Eq. (1).
In contrast, under similar conditions but with anti-
phase inputs, the network oscillator model became
less prone to quasi-periodicity (and more easily en-
Figure 12. Transition of oscillatory mode from apneic episode (a, b) trained; see Section 3.3). When quasi-periodicity was
to total entrainment (d) in a forced half-center model (forced Skinner provoked, it often assumed a much less complex form
model). than those found for in-phase inputs. Similar results
were obtained with the half-center model and three-
higher than the spontaneous oscillation frequency of phase respiratory CPG model.
the network.
Figure 13 shows the transitions in oscillatory be- 3.4.2. Chaos. Despite the proneness to quasi-periodic
havior of the model under in-phase inputs at a driving oscillation in a compound CPG, in none of the models
frequency of 3 Hz, as the amplitude of the driving in- being studied was there evidence of any chaotic mo-
put was increased. With a weak periodic input the tion induced by a periodic input. The only exception
network model oscillated at its spontaneous frequency was when the I and E neurons of the model were ex-
with the I and E neurons activity out of phase to cited by periodic inputs of differing frequencies (for
each other (Fig. 13(a)). As the amplitude of the input example, f I = 3 Hz; f E = 4.5 Hz). Under this condi-
was increased to an intermediate range (Fig. 13(b) and tion, a chaotic state in the network oscillation ensued
(c)), the oscillation became quasi-periodic. With even (Matsugu and Poon, 1997). To characterize the chaotic
stronger inputs (that is, A ≥ 38) the network was to- behavior, the Lyapunov exponents of the trajectories
tally entrained to the driving frequency with both I and were calculated using Wolf’s algorithm (Parker and
E neurons oscillating nearly in-phase to the periodic Chua, 1989) with a data length of approximately 7,600.
input (Fig. 13(d)). A similar pattern was observed at The largest exponent found was +0.06, suggesting
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

46 Matsugu, Duffin and Poon

weak chaos. This interpretation was supported by the tractus solitarius in the brain stem may elicit a dose-
estimated power spectrum of the induced oscillations dependent apnea, mimicking the responses to activa-
(in either I or E neuron), which exhibited a broad- tion of pulmonary stretch receptors (Bonham et al.,
band behavior without recognizable repetitive peaks at 1993) and cardiopulmonary C fibers (Vardhan et al.,
discrete frequencies. This is in contrast to the power 1993). Similarly, a low-frequency periodic input of
spectrum of a quasi-periodic oscillation resulting from intermediate strength may induce apneic episodes by
an in-phase input at a frequency of 3 Hz. altering the balance of the excitations to the I and E
The bifurcation to chaos was dependent on both the neurons (Figs. 9 and 10). For example, during certain
amplitudes and frequencies of the inputs. For example, rhythmic movement activities (such as weight training)
when the amplitudes of the inputs were increased from the respiratory rhythm may be periodically interrupted
21 to 50, the chaotic trajectories reduced to a simple in synchrony with the rhythm of muscular exertion.
periodic oscillation (Matsugu and Poon, 1997). On the
other hand, when the amplitudes were kept constant
(A I = A E = 21) while the frequencies were varied 4.2. Entrainment Modes Induced by Periodic Inputs
(for example, f I = 3 Hz; f E = 3.75 Hz), the chaotic
trajectories reduced to a more complex form of periodic 4.2.1. Entrained Oscillations More Stable with Anti-
oscillation. phase than In-phase Inputs. A second major finding
was that the entrained oscillations resulting from anti-
phase inputs (or equivalently, single oscillatory inputs)
4. Discussion were both stronger and more stable than those result-
ing from in-phase inputs under similar driving input
4.1. Rhythmicity is Contingent on Balanced conditions. This behavior results because the critical
Excitatory Drives amplitude for entrainment is lower for anti-phase inputs
than for in-phase inputs (Fig. 8), and can be intuitively
One of the first and more elementary features of the deduced by considering that the agonist-antagonist pat-
CPG model was that spontaneous rhythmicity was tern of anti-phase inputs is intrinsically more compat-
contingent on balanced excitations from an external ible with network oscillation in reciprocally inhibiting
drive (such as the central and peripheral chemorecep- neurons.
tors inputs or metabolic inputs), a behavior mimick- For suprathreshold (that is, A ≥ Ac ) periodic inputs
ing that observed in mammals (Cunningham et al., the entrained rhythm was less susceptible to magnitude
1986; Phillipson et al., 1981). Sustained rhythm was attenuation or harmonic and phase distortions (phase
produced only when both the I and E neurons were angle between I and E neuron activity <180◦ ) when the
suitably excited with the relative magnitudes of exci- inputs were anti-phase than when they were in-phase
tation lying within a certain range (Fig. 3(a)). Once (Figs. 4 to 7). This behavior is important because an
spontaneous oscillation was initiated, it assumed a entrained respiratory rhythm with decreased amplitude
characteristic frequency that was determined by net- and with <180◦ phase difference between the I and
work configuration and was independent of the mag- E neuron activity would be deleterious to the genera-
nitudes of excitatory afferent inputs (Fig. 3(b))—a tion of respiratory movement and adequate pulmonary
behavior again mimicking that found in vagotomized ventilation. Likewise, for subthreshold (intermediate)
mammals (Lahiri et al., 1975). inputs the resultant oscillation was also less prone to
An interesting implication of this finding is that ces- quasi-periodicity (Fig. 13). This behavior is advanta-
sation of oscillation may result from the lack of gen- geous for the respiratory CPG because some of the
eral excitation as well as preferential excitation of the strongest external inputs are anti-phase (for example,
I or E neurons. Thus, afferent inputs whose central the inspiratory shortening and expiratory prolonging
transmission is mediated by excitatory synapses could Hering-Breuer reflex via phasic vagal feedback in many
unbalance the CPG if they excite only the I or E neu- animals). Indeed, the fact that the respiratory CPG has
ron (but not both), thereby suppressing (rather than been found to be readily entrained to a variety of exoge-
augmenting) network oscillation without the need for nous periodic inputs in a highly robust and harmonious
inhibitory interneurons. Indeed, recent studies have manner (Bechbache and Duffin, 1977; Bramble and
demonstrated that preferential activation of excitatory Carrier, 1983; Haas et al., 1986; Graves et al., 1986;
amino acid receptors in certain areas of the nucleus Takahashi et al., 1992; Sammon and Darnall, 1994)
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 47

suggests that many central and peripheral inputs to the entrainment modes. It should be noted, however, that
respiratory CPG may be anti-phase. such quasi-periodic oscillations are seldom observed in
neonates and thus may not represent a significant risk
to respiratory instability. Furthermore, unlike periodic
4.2.2. Possible Implications in Developing Respira- breathing (which is highly hazardous), the induced
tory CPG. Our finding that a compound CPG is in- quasi-periodic oscillations are highly rhythmic in na-
trinsically stronger, more harmonious and more sta- ture with no signs of apnea. This is particularly true
ble in the anti-phase than in-phase configuration may if the pacemaker cell projects antagonistically to the
have important implications in the neural development I and E neurons in an anti-phase (or single-ended)
of some mammalian species that evidence a transfor- manner. The most profound disruptions of the respi-
mation from a pacemaker driven respiratory rhythm ratory pattern occur in the third critical stage when
during the neonatal period (Arata et al., 1990; Smith the pacemaker frequency becomes much lower than
et al., 1991) to a network-based rhythm in adulthood the spontaneous frequency ( f ¿ f spontaneous ). Respira-
(Hayashi and Lipski, 1992; Richter et al., 1992; Smith, tory instability in the form of periodic breathing may
1996). A strong and harmonious entrained oscillation ensue if the amplitude of the pacemaker drive falls
is crucial for the neonate, who may rely primarily on within the intermediate range (Amin apnea < A < Aapnea )
max
pacemaker input for the maintenance of the respiratory (Figs. 9 to 11). Particularly at risk for such instabil-
rhythm. Furthermore, a stable transition from pace- ity are infants with decreased chemical drive relative
maker to network-based oscillation is of paramount to the pacemaker input (such that Amin apnea < A is satis-
importance in minimizing the risks of respiratory insta- fied) and/or unbalanced drive to the I and E neurons
bility and sudden infant death during the developmental (i.e., the DC E /DC I ratio borders on or falls outside the
period when network oscillation becomes increasingly range for spontaneous oscillation). It has been recently
prominent with progressive maturation of the chemore- shown (Kinney et al., 1995) that victims of sudden
ceptor excitatory system (Henderson-Smart and Co- infant death syndrome are associated with decreased
hen, 1988) as well as the GABAergic inhibitory system muscarinic cholinergic receptor activity in the arcu-
(Hayashi and Lipski, 1992; Schweitzer et al., 1992) and ate center of the ventromedullary surface, a putative
adaptation system via potassium ion channels (Kalia et chemosensitive region.
al., 1993). Our simulation results suggest the possibil- It is generally believed that a primary cause of pe-
ity of at least three critical stages in which the infant riodic breathing is increased gain or phase delay of
is most susceptible to respiratory instability in a devel- the chemoreflex loop via the carotid chemoreceptors
oping compound CPG. (Khoo et al., 1982). Although this model seems to fit
The first critical stage occurs when neural inhibi- the behaviors in adult subjects under various physio-
tion/adaptation lags behind excitation in maturation logical states, it fails to explain the periodic breath-
(for example, due to pathologic neural development) ing frequently observed among infants during the first
so that chemoafferent excitation is strong enough to few weeks of life (Barrington et al., 1987) when the
initiate spontaneous network oscillation but with a fre- peripheral chemoreceptors are not yet well developed
quency much lower than the drive frequency of the (Kholwadwala and Donnelly, 1992; Carroll et al.,
pacemaker (that is, f spontaneous ¿ f ) due to weak recip- 1993). Our simulation results (Figs. 8 to 11) suggest
rocal inhibition and adaptation. Quasi-periodic oscilla- the possibility that respiratory instability during this
tion may ensue if the relative amplitudes of the network period may derive centrally from inharmonious inter-
oscillation and drive input encroach on the interme- actions of network oscillation and pacemaker drive in-
diate range (Ac > A > Amin ), with a drive frequency stead of peripherally from destabilizing chemoreceptor
intermediate between dominant entrainment modes as feedback.
defined by the Arnold tongues (Fig. 8). The second
critical stage of unstable transition of the respiratory
CPG occurs when the network oscillator is fully de- 4.3. Absence of Chaos in Compound CPG
veloped and pacemaker influence begins to decline.
Quasi-periodicity may again ensue if the amplitude of The third major finding of this study was that chaotic
the pacemaker oscillation decreases to the intermedi- motion is unlikely to develop in a network CPG in the
ate range ( Ac > A > Amin ) while the bursting frequency absence of significant phasic feedback. Thus, the spon-
of the pacemaker is intermediate between dominant taneous oscillation of the network was highly regular
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

48 Matsugu, Duffin and Poon

with simple Hopf bifurcations (Glass and Mackey, (Aihara and Matsumoto, 1986), some bursting neurons
1988) only at the critical points for the initiation and (Canavier et al., 1990), and cardiac rhythm (Guevara
termination of the rhythm (Fig. 3). This finding is and Glass, 1982). Generally, a periodic input may in-
in agreement with previous experimental observations duce chaos in a nonlinear oscillator only if the latter
that phasic feedback from pulmonary stretch receptors has undergone at least two Hopf bifurcations resulting
is a strong determinant of the respiratory rhythm in in two or more fundamental frequencies (Ostlund et
many animals (Poon, 1995) and that vagotomy greatly al., 1983; Bergé et al., 1984; Tomita, 1986; Glass and
reduces the nonlinear dimension of respiratory activity Mackey, 1988).
(Sammon and Bruce, 1991). Indeed, the oscillation of The lack of evidence of chaotic motion in our com-
the network model remained periodic or quasi-periodic pound CPG model contradicts the hypothesis that chaos
even when the network was driven periodically. This may play a role in the pathogenesis of SIDS (Sheridan
was particular true when the periodic input was anti- and Kostlany, 1994), although this does not rule out
phase. Weakly chaotic behavior was observed only the possibility of other complex mechanisms that may
when the network CPG was driven by multiple peri- induce such chaotic motion in the neonate.
odic sources, and even so the bifurcation to chaos was
highly nonrobust and was readily abolished by changes
in the amplitude and frequency of the periodic drives. 4.4. Summary and Conclusions
Such weak chaotic motion, even if present, is unlikely
to be detectable in experimental measurements that are In summary, the present study demonstrated that the
inevitably corrupted by random noise (Barahona and behavior of a periodically perturbed neural network
Poon, 1996). oscillator without phasic feedback is characterized by
The absence of chaotic regime in a compound at least six oscillation modes (Table 1). In all cases,
CPG is in contrast to other periodically driven oscil- the neural network oscillator interacts more harmo-
lators such as the forced Duffing oscillator and the niously (with less distortions in waveshape) with anti-
forced Brusselator (Tomita, 1986), squid giant axons phase inputs having relative antagonistic effects on the

Table 1. Oscillation modes of a compound CPG.

Mode Input conditions Characteristicsa

Spontaneous oscillation Amin > A ≥ 0 Spontaneous frequency set by

DC I > DC E > 0.7DC I reciprocal inhibition and
adaptation
DC E ≥ 16.8
Total entrainmentb A ≥ Ac 1 : 1 entrainment to external
periodic drive
(rotation number = 1)
Partial entrainmentb Ac > A > Amin Rotation number is a rational
f within Arnold tongues number 6= 1

Quasi-periodicityb Ac > A > Amin Erratic phase trajectory

f between Arnold tongues Power spectrum shows multiple
peaks at discrete frequencies

Chaos Multiple periodic inputs with Erratic phase trajectory

frequency and amplitude similar Broad-band power spectrum
to quasi-periodic mode

apnea < A < Aapnea

Amin
Intermittent apnea max Transition between apnea and
f ¿ f spontaneous periodic oscillation at intrinsic
frequency ( f spontaneous )
with Amin max
apnea , Aapnea dependent on DC I , DC E

a Seetext for explanation of characteristics and their physiologic correlates.

b Network oscillator is more readily entrained by anti-phase periodic inputs than in-phase inputs as well as less
prone to magnitude attenuation, harmonic and phase distortions, and quasi-periodic interference.
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 49

reciprocally inhibiting neurons than otherwise. These where, V 1 , V 2 is the voltage of each neuron, other
simulation results are in excellent agreement with a parameter values are set as follows (refer to Skinner
wide variety of experimental findings concerning the et al., 1994, for the meaning of each parameter): C =
normal behavior of the respiratory CPG and may shed 1.0, g K = 2.0, gCa = 1.1, g L = 0.5, VCa = 1.0,
important light on its abnormal behavior in certain VK = −0.7, VL = −0.5, Vsyn = −0.7, V1 = −0.01,
pathologic states. Because the proposed model rep- V2 = 0.15, V3 = 0.0, V4 = 0.3, gsyn = 0.6, φ N = 0.2,
resents a minimum kernel for rhythmogenesis in any Vthresh = 0.0, Vslope = 0.000001, I0 = 0.3, A = 0.5.
neural network oscillator, the present results may pro- In the above equation, Iext is diferent from Skinner
vide a useful basis for the elucidation of complex os- et al. (1994) in that it gives oscillatory input. We fol-
cillatory behaviors in other compound CPGs found in lowed Rinzel and Ermentrout (1989) to specify the
many invertebrate and vertebrate nervous systems. activation parameters for each identical neuron, and
Finally, nonlinear interactions of respiratory CPG the DC component I0 , of periodic external input was
and pacemaker inputs could result in respiratory insta- chosen so that each neuron is self-oscillatory (that is,
bility and apnea, but this would be most likely caused 0.3 < I0 < 0.45).
by unbalanced excitations of the inspiratory and ex-
piratory related neurons rather than the induction of
chaotic motion in the respiratory CPG. Appendix B: Forced Gottschalk Model

The model is composed of five distinct respira-

tory neurons; early-inspiratory (first neuron), ramp-
Appendix A: Forced Half-Center Oscillator inspiratory (second neuron), late-inspiratory (third
neuron), postinspiratory (fourth neuron), and expira-
The following gives the details of the forced half-center tory (fifth neuron) neurons.
oscillator model used in our reference model composed Among mutually inhibiting neurons, we chose the
of two mutually inhibiting Morris-Lecar type neurons. early-inspiratory and expiratory neurons to be fed by
In this model each neuron is tuned to be self-oscillatory. anti-phase oscillatory inputs, and other neurons were
This mode of neuronal activation is chosen because made to receive only DC excitations.
otherwise (that is, nonoscillatory mode) we could not The activation equation:
get the network parameters which realize network os-
illation. d xi X
The activation equation (Skinner et al., 1994) of τ + xi = ai j max(xi , 0) + (si − ci w) − bi z i
dt j
neuron one (designated by superscript 1) is
dz i
T + z i = max(0, xi )
dV 1 dt
C = Iext − (g L (V 1 − VL ) + gCa M∞
1
dt dw
γ + w = max(0, x4 )
×(V 1 − VCa ) + g K N 1 (V 1 − VK ) dt
+ gsyn S∞1
(V 1 − Vsyn )) si = Ai sin(2π f t + φi ) + DCi ,

dN1 ¡ 1 ¢ where xi is the activity of the ith neuron, si denotes an

= λ1N N∞ − N1
dt input for the ith neuron, A1 = A5 = 1.0, Ak = 0 (for
Iext = I0 + A sin(2π f t + θ0 ) 2 ≤ k ≤ 4), DCk = 1.0 (for k ≤ 4), DC5 = 1.1,
µ µ 1 ¶¶ φ5 = 0.5π , φk = 0.0 (for k ≤ 4).
1 V − V1
1
M∞ (V 1 ) = 1 + tanh See Gottshalk et al. (1994, p. 60, Fig. 2) for other
2 V2
µ µ 1 ¶¶ parameter values.
1 V − V3
N∞ (V ) =
1 1
1 + tanh
2 V4 Acknowledgments
µ 1 ¶
V − V3
λ N (V ) = φ N cosh
1 1
This work was supported in part by Office of Naval
2V4
µ µ 2 ¶¶ Research grant N00014-95-1-0414, National Sci-
1 V − Vthresh
1
S∞ (V 2 ) = 1 + tanh , ence Foundation grant BCS-9216419, National Insti-
2 Vslope tutes of Health grants HL-50614 and HL-52925 (to
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

50 Matsugu, Duffin and Poon

CSP), and Canadian Medical Research Council grant Duffin J (1991) A model of respiratory rhythm generation. Neuro-
MA7842 (to JD). report 2:623–626.
Duffin J, Douse MA (1993) Botzinger expiratory neurons inhibit
propriobulbar decrementing inspiratory neurons. NeuroReport
References 4:1215–1218.
Duffin J, Ezure K, Lipski J (1995) Breathing rhythm generation:
Aihara K, Matsumoto G (1986) Chaotic oscillations and bifurca- Focus on the rostral ventrolateral medulla. News in Physiol. Sci.
tions in squid giant axons. In: AV Holden, ed. Chaos. Princeton 10:133–140.
University Press, Princeton, NJ. pp. 257–269. Ermentrout GB (1982) n:m Phase-locking of weakly coupled oscil-
Alexander JC, Doedel EJ, Othmer HG (1990) On the resonance lators. J. Math. Biology 12:327–342.
structure in a forced excitable system. SIAM J. Appl. Math. Ezure K, Manabe M (1986) An origin of inspiratory inhibition of ex-
50:1373–1418. piratory neurons in the Botzinger complex. J. Physiol. Soc. Japan
Arata A, Onimaru H, Homma I (1990) Respiration-related neurons 48:384.
in the ventral medulla of newborn rats in vitro. Brain Res. Bull. Friesen WO, Stent GS (1977) Generation of a locomotory rhythm
24:599–604. by a neural network with recurrent cyclic inhibition. Biol. Cybern.
Balis UJ, Morris KF, Koleski J, Lindsey BG (1994) Simulations of 28:27–40.
a ventrolateral medullary neural network for respiratory rhythmo- Glass L, Mackey MC (1988). From Clocks to Chaos. Princeton Uni-
genesis inferred from spike train cross-correlation. Biol. Cybern. versity Press, Princeton, NJ.
70:311–327. Gottscchalk A, Geitz KA, Richter DW, Ogilvie MD, Pack AI (1992)
Barahona M, Poon C-S (1996) Detection of nonlinear dynamics in Nonlinear Dynamics of a Model of the Central Respiratory Pat-
short, noisy time series. Nature 381:215–217. tern Generator. In: Y Honda et al., eds. Control of Breathing
Barrington KJ, Finer NN, Wilkinson MH (1987) Progressive short- and Its Modeling Perspective. Plenum Press, New York. pp. 51–
ening of the periodic breathing cycle duration in normal infants. 55.
Pediatr. Res. 21:247–251. Gottschalk A, Ogilvie MD, Richter DW, Pack A (1994) Computa-
Bechbache RR, Duffin J (1977) The entrainment of breathing fre- tional aspects of the respiratory pattern generator. Neural Compu-
quency by exercise rhythm. J. Physiol. 272:553–561. tation 6:56–68.
Bergé P, Pomeau Y, Vidal C (1984) Order Within Chaos. Wiley, New Graves C, Glass L, Laporta D, Meloche R, Grassino A. (1986)
York. Respiratory phase locking during mechanical ventilation in anes-
Bonham AC, Coles SK, McCrimmon DR (1993) Pulmonary stretch thetized human subjects. Am. J. Physiol. 250 (Regulatory, Integra-
receptor afferents activate excitatory amino acid receptors in the tive Comp. Physiol. 19):R902—R909.
nucleus tractus solitarii in rats. J. Physiol. 464:725–745. Guevara MR, Glass L (1982). Phase locking, period doubling bi-
Botros SM, Bruce EN (1990) Neural network implementation of a furcations and chaos in a mathematical model of a periodically
three-phase model of respiratory rhythm generation. Biol. Cybern. driven oscillator: A theory for the entrainment of biological oscil-
63:143–153. lators and the generation of cardiac dysrhythmias. J. Math. Biology
Bramble DM, Carrier DR (1983) Running and breathing in mam- 14:1–23.
mals. Science 219:251–256. Guz A, Trenchard DW (1971) Pulmonary stretch receptor activity in
Canavier CC, Clark JW, Byrne JH (1990) Routes to chaos in a model man: A comparison with dog and cat. J. Physiol. London 213:329–
of bursting neuron. Biophys. J. 57:1245–1251. 343.
Carroll JL, Bamford OS, Fitzgerald RS (1993) Postnatal maturation Haas F, Distenfeld S, Axen K (1986) Effects of perceived musical
of carotid chemoreceptor responses to O2 and CO2 in the cat. J. rhythm on respiratory pattern. J. Appl. Physiol. 61:1185–1191.
Appl. Physiol. 75:2383–2391. Hayashi F, Lipski J (1992) The role of inhibitory amino acids in
Cohen M, Grossberg S, Pribe C (1992) A neural pattern generator control of respiratory motor output in an arterially perfused rat.
that exhibits frequency-dependent in-phase and anti-phase oscil- Respir. Physiol. 89:47–63.
lations. In: Conference on Neural Networks for Learning, Recog- Henderson-Smart DJ, Cohen GL (1988) Chemical control of breath-
nition and Control. Boston University, May 14–16, p. 32. ing in early life. In: PJ Schwartz, DP Southall, M Valdes-Dapena,
Collins JJ, Stewart I (1994) A group-theoretic approach to rings of eds. The Sudden Infant Death Syndrome. Annals of the New York
coupled biological oscillators. Biol. Cybern. 71:95–103. Academy of Sciences, New York Academy of Sciences, New York,
Corio M, Palisses R, Viala D (1993) Origin of the central entrain- vol. 533, pp. 276–288.
ment of respiration by locomotion facilitated by MK 801 in the Kalia M, Schweitzer P, Champagnat J, Denavit-Saubie M (1993) Two
decerebrate rabbit. Exp. Brain Res. 95:84–90. distinct phases characterize maturation of neurons in the nucleus
Cunningham DJC, Robbins PA, Wolff CB (1986) Integration of res- of the tractus solitarius during early development: Morphological
piratory responses to changes in alveolar partial pressures of CO2 and electrophysiological evidence. J. Comp. Neurol. 327:18–36.
and O2 and in arterial pH. In: NS Cherniack, JG Widdicombe, Kazanovich YB, Borisyuk RM (1994) Synchronization in a neural
eds. Handbook of Physiology, Section 3: The Respiratory Sys- network of phase oscillators with central element. Biol. Cybern.
tem, Control of Breathing, Part 2, American Physiological Society, 71:177–185.
Bethesda, MD., vol. II, pp. 475–528. Kholwadwala D, Donnelly DF (1992) Maturation of carotid chemo-
Cymbalyuk GS, Nikolaev RM, Borisyuk RM (1994) In-phase and receptor sensitivity to hypoxia: In vitro studies in the newborn rat.
anti-phase self-oscillation in a model of two electrically coupled J. Physiol. (London) 453:461–473.
pacemakers. Biol. Cybern. 71:153–160. Khoo MCK, Kronauer RE, Strohl KP, Slutsky AS (1982) Factors
Delcomyn F (1980) Neural basis of rhythmic behavior in animals. inducing periodic breathing in humans: A general model. J. Appl.
Science 210:492–498. Physiol. 53:644–659.
P1: SGR
Journal of Computational Neuroscience KL523-03-Matsugu December 13, 1997 10:21

Nonlinear Dynamics in Compound CPG 51

Kinney HC, Filiano JJ, Sleeper LA, Mandell F, Valdes-Dapena M, Modeling, Simulation and Control. CRC Press, Boca Raton, FL.
White WF (1995) Decreased muscarinic receptor binding in the ar- pp. 2404–2421.
cuate nucleus in sudden infant death syndrome. Science 269:1446– Richter DW, Ballantyne D, Remmers JE (1986) How is the respira-
1450. tory rhythm generated? A model. News Physiol. Sci. 1:109–112.
Kopell N (1988) Toward a theroy of modelling central pattern gener- Richter DW, Ballanyi K, Schwarzacher S (1992) Mechanisms of
ators. In: AH Cohen, S Rossignol, S Grillner, eds. Neural Control respiratory rhythm generation. Curr. Opin. Neurobiol. 2:788–
of Rhythmic Movements. Wiley, New York. pp. 369–413. 793.
Kopell N, Ermentrout GB (1988) Coupled oscillators and the design Rinzel J, Ermentrout GB (1989) Analysis of neural excitability and
of central pattern generators. Math. Biosci. 90:87–109. oscillations. In: C. Koch, ed. Methods in Neuronal Modeling. MIT
Lahiri S, Mei SS, Kao FF (1975) Vagal modulation of respiratory Press, Cambridge, MA. pp. 135–169.
control during exercise. Respirat. Physiol. 23:133–146. Sammon MP, Bruce EN (1991) Vagal afferent activity increases
Matsugu M, Poon CS (1993) Estimation of information capacity in dynamical dimension of respiration in rats. J. Appl. Physiol.
oscillatory neural networks. In: Proceedings of the International 70:1748–1762.
Joint Conference on Neural Networks, Nagoya, Japan. pp. 425– Sammon MP, Darnall RA (1994) Entrainment of respiration to rock-
428. ing in premature infants: coherence analysis. J. Appl. Physiol.
Matsugu M, Poon CS (1997) Nonlinear Dynamics in a Compound 77:1548–1554.
Central Pattern Generator. In: Proceedings of the Fifth Annual Schweitzer P, Fortin G, Beloeil JC, Champagnat J (1992) In vitro
Computational Neuroscience Meeting. Plenum Press, Boston. study of newborn rat brain maturation: Implications for sudden
Matsuoka K (1985) Sustained oscillations generated by mutually infant death syndrome. Neurochem. International 20:109–112.
inhibiting neurons with adaptation. Biol. Cybern. 52:367–376. Selverston A, Mazzoni P (1989) Flexibility of computational units in
Mifflin S, Ballantyne D, Backman S, Richter DW (1985) Evidence invertebrate CPGs. In: R Durbin, C Miall, G Mitchison, eds. The
for a calcium activated potassium conductance in medullary respi- Computing Neuron. Addison-Wesley, Reading, MA. pp. 205–228.
ratory neurons. In: Neurogenesis of Central Respiratory Rhythm. Sheridan MS, Kostlany E (1994) SIDS and chaos. Medical Hypoth-
MTP Press, Lancaster. pp. 179–182. esis 42:11–12.
Ogilvie MD, Gottschalk A, Anders K, Richter DW, Pack AI (1992) Skinner FK, Kopell N, Marder E (1994) Mechanisms for oscilla-
A network model of respiratory rhythmogenesis. Am. J. Physiol. tion and frequency control in reciprocally inhibitory model neural
263:R962–R975. networks. J. Comput. Neurosci. 1:69–88.
Ostlund S, Rand D, Sethna D, Siggia E (1983) Universal properties of Smith JC (1996) Realistic computational models of respiratory neu-
the transition from quasi-periodicity to chaos in dissipative sytems. rons and networks. In: M Khoo, ed. Bioengineering Approaches
Physica 8D:303–342. to Pulmonary Physiology and Medicine. Plenum Press, New York.
Parker TS, Chua LO (1989) Practical Numerical Algorithms for pp. 77–92.
Chaotic Systems. Springer-Verlag, New York. p. 80. Smith JC, Ellenberger K, Ballanyi D, Richter DW, Feldman JL (1991)
Petrillo GA, Glass A (1984) A theory for phase locking respiration Pre-Bötzinger complex: a brain stem region that may generate res-
in cats to a mechanical ventilator. Am. J. Physiol. 246 (Regulatory piratory rhythm in mammals. Science 254:726–729.
Integrative Comp. Physiol. 15):R311–R320. Takahashi E, Tejima K, Tateishi I (1992) Synchronization of res-
Phillipson EA (1986) Control of breathing during sleep. In: NS piratory rhythm with mechanical ventilation by respiratory O2
Cherniack, JG Widdicombe, eds. Handbook of Physiology, Sec- oscillations in vagotomized dogs. In: T Honda, Y Miyamoto, K
tion 3: The Respiratory System, Control of Breathing, Part 2, Konno, JG Widdicomb, eds. Control of Breathing and Its Model-
American Physiological Society, Bethesda, MD., vol. II, pp. 649– ing Perspectives. Plenum Press, New York. pp. 147–150.
690. Tomita K (1986) Periodically forced nonlinear oscillators. In: AV
Phillipson EA, Duffin J, Cooper JD (1981) Critical dependence of Holden, ed. Chaos. Princeton University Press, Princeton, NJ. pp.
respiratory rhythmicity on metabolic CO2 load. J. Appl. Physiol. 257–269.
50:45–54. Vardhan A, Kachroo A, Sapru HN (1993) Excitatory amino acid re-
Pinsky PF, Rinzel J (1994) Intrinsic and network rhythmogenesis in ceptors in the nucleus tractus solitarius mediate the responses to
a reduced Traub model for CA3 neurons. J. Comput. Neurosci. the stimulation of cardio-pulmonary vagal afferent C fiber endings.
1:39–60. Brain Research 618:23–31.
Poon C-S (1995) Respiratory models and control. In: HJ Chizeck, ed. Widdicombe JG (1961) Respiratory reflexes in man and other mam-
Handbook of Biomedical Engineering, Section 15, Physiological malian species. Clin. Sci. 21:163–170.