NeuroImage 47 (2009) 1371–1380

Contents lists available at ScienceDirect

NeuroImage
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / y n i m g

A time-invariant visco-elastic windkessel model relating blood flow and

blood volume
Ying Zheng ⁎, John Mayhew
Centre for Signal Processing in Neuro-imaging and Systems Neuroscience, Department of Psychology, University of Sheffield, Northumberland Road, Sheffield S10 2TP, UK

a r t i c l e i n f o a b s t r a c t

Article history: The difference between the rate of change of cerebral blood volume (CBV) and cerebral blood flow (CBF)
Received 12 January 2009 following stimulation is thought to be due to circumferential stress relaxation in veins (Mandeville, J.B.,
Revised 2 March 2009 Marota, J.J.A., Ayata, C., Zaharchuk, G., Moskowitz, M.A., Rosen, B.R., Weisskoff, R.M., 1999. Evidence of a
Accepted 4 April 2009
cerebrovascular postarteriole windkessel with delayed compliance. J. Cereb. Blood Flow Metab. 19, 679–689).
Available online 14 April 2009
In this paper we explore the visco-elastic properties of blood vessels, and present a dynamic model relating
changes in CBF to changes in CBV. We refer to this model as the visco-elastic windkessel (VW) model. A novel
feature of this model is that the parameter characterising the pressure–volume relationship of blood vessels
is treated as a state variable dependent on the rate of change of CBV, producing hysteresis in the pressure–
volume space during vessel dilation and contraction. The VW model is nonlinear time-invariant, and is able
to predict the observed differences between the time series of CBV and that of CBF measurements following
changes in neural activity. Like the windkessel model derived by Mandeville, J.B., Marota, J.J.A., Ayata, C.,
Zaharchuk, G., Moskowitz, M.A., Rosen, B.R., Weisskoff, R.M., 1999. Evidence of a cerebrovascular postarteriole
windkessel with delayed compliance. J. Cereb. Blood Flow Metab. 19, 679–689, the VW model is primarily a
model of haemodynamic changes in the venous compartment. The VW model is demonstrated to have the
following characteristics typical of visco-elastic materials: (1) hysteresis, (2) creep, and (3) stress relaxation,
hence it provides a unified model of the visco-elastic properties of the vasculature. The model will not only
contribute to the interpretation of the Blood Oxygen Level Dependent (BOLD) signals from functional
Magnetic Resonance Imaging (fMRI) experiments, but also find applications in the study and modelling of
the brain vasculature and the haemodynamics of circulatory and cardiovascular systems.
© 2009 Elsevier Inc. All rights reserved.

Introduction (Fogliardi et al., 1996; Jones, 1969; Molino et al., 1998; Orosz et al.,
1999; Segers et al., 2008; Stergiopulos et al., 1999; Westerhof et al.,
The visco-elastic properties of blood vessels have been extensively 1971). The windkessel model was further extended into multi-
investigated. The pressure–volume (P–V) relationships of blood compartment models (Linehan et al., 1986; Ursino et al., 2000) for
vessels have been measured under static and dynamic conditions in circulation modelling.
vitro in many studies, e.g. Alexander et al. (1953), Bergel (1961), More recently the windkessel model has been incorporated in
Linehan et al. (1986), Porciuncula et al. (1964), and Remington and functional studies investigating the relationship between changes in
Alexander (1955). The dynamic P–V relationship was characterised by cerebral blood flow (CBF) and cerebral blood volume (CBV) due to
hysteresis during cycles of vessel dilation and contraction. evoked changes in neural activity (Buxton et al., 2004, 1998; Kong
Many models of blood vessels are based on the windkessel theory et al., 2004; Mandeville et al., 1999). Variants of such CBF–CBV models
(Frank, 1930). In the simplest form, the relationship between blood have been used by others to analyse Blood Oxygen Level Dependent
pressure and flow was modelled in electrical terms by a resistor and a (BOLD) functional Magnetic Resonance Imaging (fMRI) data and to
capacitor connected in parallel. This was known as the two-element infer neuronal activity (Deneux and Faugeras, 2006; Friston et al.,
windkessel model. The capacitor models the blood vessel compliance 2000; Vakorin et al., 2007; Vazquez et al., 2006; Zheng et al., 2005,
and the resistor models the vessel resistance encountered by blood 2002). It is often observed that after stimulus cessation, the time
flowing through the vascular system. Extensions of the simple model series of CBV returns to baseline more slowly than that of CBF
into multi-element windkessel models were developed to better (Herman et al., 2008; Jones et al., 2002; Kennerley et al., 2005; Kida
characterise systemic as well as pulmonary circulation systems et al., 2007; Mandeville et al., 1999). Some studies also showed that for
long stimulus duration, the CBV response continues to increase during
⁎ Corresponding author. Fax: +44 114 276 6515. the stimulus onset period while CBF response has plateaued (Herman
E-mail address: ying.zheng@shef.ac.uk (Y. Zheng). et al., 2008; Kida et al., 2007; Mandeville et al., 1999). These

1053-8119/$ – see front matter © 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.neuroimage.2009.04.022
1372 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380

differences in the time series of CBF and CBV are thought to be due to noise. The LDF time series were collected concurrently with spectro-
the visco-elastic properties of blood vessels, and the term ‘delayed graphic data for the following experimental paradigms.
compliance’ is often loosely used to refer to the mismatch between the Electrical stimulation of the whisker pad was delivered via
time series of CBF and CBV. tungsten electrodes inserted in an anterior direction each side of the
This paper presents a novel physiologically plausible model of the D1 whisker and ∼ 2 mm along the D barrel row. All stimuli were
CBF–CBV coupling which incorporates the visco-elastic properties of presented at 1.2 mA with a 0.3 ms individual pulse width. Stimuli of 1,
blood vessels. The model not only exhibits the visco-elastic characte- 2, 3, 4 and 5 Hz were randomly interleaved and presented for 2 s, with
ristics (i.e., creep loading, stress relaxation and hysteresis), but also stimulus onset at 8 s after the start of each trial, which lasted 23 s, and
can be used to predict time series of changes in physical quantities data were averaged over 30 trials. Experiments were conducted on six
such as the transmural pressure. Furthermore, explicit expressions for animals, one of which had missing data after 15 s of each trial and
static and dynamic compliances of blood vessels can be derived. A hence was excluded from the analysis. Also one animal had missing
further distinguishing feature is that a parameter characterising the data for the 4 Hz stimulus condition, hence was excluded from the
steady state P–V curve is treated as an additional state variable of the analysis under 4 Hz stimulation.
model. Based on the visco-elastic properties of blood vessels, this Fig. 1 shows the CBF and CBV time series of one animal
parameter has its static value on the steady state P–V curve at steady (representative) under the five stimulus conditions. To emphasise
states, but during transient states (i.e., following changes in CBF) it the difference between the CBF and CBV responses during the return-
takes on different values depending on the rate of change of CBV. The to-baseline phase, the time series are normalised between 0 and 1. The
assumptions made in deriving the model are more appropriate for the difference can be observed for all data with higher stimulus
venous compartment of the vasculature although the arteries and frequencies. However at lower stimulus frequencies (e.g., 1 Hz and
arterioles also have visco-elastic properties and the concepts of static 2 Hz), some data sets show no evidence of ‘delayed compliance’. This
and dynamic compliances apply equally to the arterial compartment. may partly be due to the relatively low signal-to-noise ratio of the LDF
We will demonstrate that the model can predict the slow return- measurement, resulting in elevated CBF at low frequencies during
to-baseline characteristics of the CBV time series satisfactorily. The stimulus cessation in some animals.
model will be referred to as the visco-elastic windkessel (VW) model
throughout this paper. The VW model

Materials and methods Mandeville's windkessel model

The VW model is based primarily on the windkessel model
Animal preparation proposed by Mandeville et al. (1999). In their paper, Mandeville et al.
modelled the venous compartment using a two-element windkessel
The experimental data used in this paper are from Martindale et al. model so that the venous blood volume and flow are related by
(2003). The CBF and CBV data were obtained concurrently via laser
α + β
Doppler flowmetry and optical imaging spectroscopy respectively. The dV P V
=F− w =F− ; βN1 ð1Þ
experimental procedures for concurrent measurement of CBF and CBV dt Rw KAβ
were detailed in Jones et al. (2001). They are briefly reviewed here.
The animals used were female Hooded Lister rats weighing where V is the venous blood volume, F is the inflow to the venous
between 250 and 400 g, anaesthetised with urethane (1.25 g/kg, compartment, Fw is the pressure drop across the windkessel. The
intraperitoneal injection), and atropine (0.4 mL/kg, subcutaneous transmural pressure of the venous compartment was assumed to be
injection). The whisker barrel was first located using single wave- the same as the pressure drop Pw because intracranial pressure and
length (∼ 590 nm) illumination. Then the slit spectrograph mounted central venous pressure in the large veins are approximately the same.
on the camera was sited over the centre of the barrel region. This was Rw is the windkessel resistance. α and β are coefficients, with α = 2
followed by the placement of an LDF probe (Perimed, Stockholm, for lamina flow, and the constraint on β N 1 to ensure diminished
Sweden; fibre separation 0.25 mm) over the barrel region (b1 mm volume reserve at high transmural pressure. K is constant dependent
from the skull surface) to measure CBF changes. The LDF spectrometer on the baseline values of V and Rw.
included a low-pass filter with a 0.2 second time constant and 12 kHz In deriving the above model, Mandeville et al. utilised two other
bandwidth (Nilsson, 1984) to reduce errors caused by measurement models. One relates the windkessel resistance to the inverse power of

Fig. 1. The time series of the CBF (grey trace) and CBV (black trace) changes normalised between 0 and 1. The stimulation duration is indicated by the black bar on the time axis. The
five data sets are from a representative subject (subject 3) under conditions of different stimulation.
 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380 1373

the blood volume as Rw ∝ (1 / V)α, the other models the pressure– changes in transmural pressure, hysteresis in the P–V plane almost
volume (P–V) relationship of the blood vessel as V = AP1w/ β where A is disappears (Porciuncula et al., 1964) and hence the trajectory of the
a constant. After normalising each variable with respect to their dynamic P–V curve converges to the steady state P–V curve. Thus the
baseline values, Mandeville et al. showed that the delayed characte- state variable A is dependent on the rate of change of the physical
ristics of the blood volume could be predicted if different venous quantities in the system. We selected the following simple first order
transient times were used for specific time intervals of stimulus onset system to model the state variable A:
and cessation. To explore this further, Mandeville et al. modified the

P–V relationship so that the constant A was modelled as an increasing dA dV
τA + A = ASS exp −b0 ; b0 N 0 ð3Þ
exponential function of time, lasting for a short time interval dt dt
beginning several seconds after stimulus onset and then again after
stimulus cessation. A single time constant was found to be sufficient to where ASS is the steady state value for A, τA is the time constant of the
account for the delayed characteristics observed in CBV (see the model and b0 reflects the amount of influence the rate of change in
Appendix of Mandeville et al., 1999). blood volume has on the parameter A.
If the normalised quantities are used: f = F / F0, v = V / V0 ,
Visco-elastic windkessel (VW) model w = A / ASS where the subscript 0 denotes baseline values and the
The windkessel model proposed by Mandeville et al. is effectively a subscript SS denotes steady state values, then Eqs. (2) and (3) result in
‘time varying’ model because a parameter of the model (e.g. the time the following VW model:
constant or the parameter A) varies with time. Furthermore the /
dv v
variation is not only discontinuous in time, but the issue of the τv =f −
dt w ð4Þ
switching of the relevant parameter values was not addressed. This

paper proposes a time-invariant dynamic model relating time series of dw dv
τw + w = exp −b
CBF to CBV. We re-examine the P–V relationship of visco-elastic dt dt
material using a version of the above P–V model slightly modified for where τv is the venous transit time (V0 / F0), the exponent ϕ = α + β is
notational convenience. The P–V relationship is modelled in this paper a constant, b = b0V0 is the gain parameter modelling the degree of
as V = (APw)1 / β, hence Eq. (1) above can be re-written as influence exerted by dv on w, and τw = τA is the time constant of w.
dt
The state variable w reflects the normalised changes in the muscle
dV P Vα + β
=F− w =F− ; βN1 ð2Þ tone of blood vessels. The VW model captures several properties of
dt Rw KA
visco-elastic material:
The variable A may be thought of as representing the vascular tone (1) At any steady state, dv
dt
= dw = 0, hence w = 1. This preserves the
dt
of a particular blood vessel type (Klabunde, 2005). By varying the steady state P–V relationship of the blood
 vessel.

parameter A, a family of P–V curved can be obtained, as depicted in d
(2) During vessel dilation, dt N 0, hence exp −b ddvt b1 (for b N 0). This
v

Fig. 2. It is important to recognise that these curves model the steady implies that w will decrease from its baseline value of unity,
state P–V relationships of blood vessels. During vessel dilation and resulting in a P–V trajectory ‘below’ the steady state P–V curve
contraction, however, the P–V curve forms a hysteresis loop around during dilation. By the same reasoning, the P–V trajectory during
the steady state curve due to the visco-elastic properties of a blood vessel contraction is ‘above’ the steady state curve (Fig. 2), thus
vessel (Porciuncula et al., 1964). The directions of the hysteresis are generating a hysteresis loop in the P–V plane as that observed by
shown in Fig. 2. This suggests that the hysteresis loop generated by Porciuncula et al. (1964).
vessel dilation and contraction may be modelled by modelling the (3) The use of the exponential function provides desirable con-
parameter A so that it decreases from its steady state value during straints on the range of values of w. As dv ranges from − ∞
dt
vessel dilation, but increases during vessel contraction. Effectively the to + ∞, the exponential function varies from 0 to +∞. Since w is
parameter A is treated as a new state variable of the model. modelled as a first order dynamic system driven by a non-
Another important characteristic of visco-elastic material is that negative exponential function, this ensures that w will never
the shape of the hysteresis in the P–V plane is dependent on the rate at become negative. Consequently the parameter A will be
which a vessel dilates or contracts. For instance, for slow sinusoidal guaranteed to lie within the range (0, + ∞), and the P–V
trajectories predicted by the model will always lie within the
first quadrant of the P–V plane.
Furthermore by clamping w = 1 for all t, the VW model reduces to

dv /
τv =f −v : ð5Þ
dt

This model is often referred to as the simple ‘balloon’ model

(Friston et al., 2000; Kong et al., 2004; Zheng et al., 2002), with the
parameter ϕ equal to the inverse Grubb exponent (Grubb et al., 1974).
By clamping w = 1, the model effectively assumes that there is no
hysteresis in the P–V plane and that the blood vessel has only elastic
properties. Hence in this paper we will refer this model as the elastic
windkessel (EW) model. The EW model is a special case of the VW
model.

Static and dynamic compliances

The compliance of a blood vessel is determined by the amount of
Fig. 2. An illustration of the hysteresis loop on a family of steady state pressure–volume
volume change per unit change in transmural pressure, and hence is
(P–V) curves. The hysteresis is the P–V trajectory formed during vessel dilation and the gradient of the P–V curve. If the curve is nonlinear, the vessel
contraction. compliance varies at every steady state. During the transient period,
1374 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380

Fig. 3. Steady state P–V curves. (a) Static compliance defined as the gradient on the steady state P–V curve. (b) Dynamic compliance defined as the gradient on the hysteresis loop of
the P–V trajectory formed during dynamic phases of dilation and contraction.

the P–V relationship of a blood vessel forms a hysteresis loop in the P–V each trial, yielding normalised changes in CBV and CBF as v = V / V0
plane. Based on these observations, we classify the compliance of and f = F / F0 respectively.
a blood vessel into two categories: one is static compliance given by The parameters of the VW model were estimated for each of the
the gradient on the steady state P–V curve (Fig. 3a), the other dy- five animals and under each of the five stimulation frequencies (1, 2, 3,
namic compliance given by the gradient along the hysteresis loop in 4 and 5 Hz), with the exception of animal 4 which has only data for
the P–V plane (Fig. 3b). Whereas the static compliance maybe used to four stimulus frequencies (1, 2, 3 and 5 Hz). Hence there are 24 data
model blood vessels with mainly elastic characteristics such that sets each with 173 data points. Parameter optimisation was carried out
the hysteresis effect is minimum, the dynamic compliance reflects using a nonlinear least squares algorithm (Levenberg–Marquardt
the visco-elastic properties of a blood vessel during dilation and algorithm, Matlab™ function “lsqnonlin”). The performance of the
contraction and is dependent on the rate at which it expands or VW model was compared with that of the EW model using the
contracts. following three methods.
From the steady state P–V model V = (APw)1 / β, it can be shown
(a) For each model, we computed the sum of squares of errors (SSE)
that the static compliance is given by
for each data set:
dV V
Cs = j = : ð6Þ X
N X
N
dP SS βP SSE =
2
ei = ðv̂ − vÞi
2
ð10Þ
i=1 i=1
Furthermore using Eq. (3) it can be shown that the dynamic
compliance is given by:
where v and v̂ are the measured CBV time series and the model
: predicted CBV time series respectively, and N is the number of
V V̇A2
CD = : =  : : ð7Þ data points used. Then the difference scores for the SSE
P V β − 1
βV̇A − VA
ΔSSE = SSEðVWÞ − SSEðEWÞ
Their normalised forms are
were computed.
CS v 1 1 (b) The corrected Akaike's information Criterion (AICc) was calculated
cs = = = 1−1=β = β−1 ð8Þ
CS0 p p v for each model under each condition. This criterion provides a
measure of goodness-of-fit of a model to empirical data, and offers
and a trade-off between model prediction and model complexity. The
2 equation used for the calculation is:
CD βw v̇
cd = = β−1 ð9Þ
CD0 v ðβwv̇ − vẇ Þ

SSE 2K ðK + 1Þ
AICc = N log + 2K + ð11Þ
respectively. Note that if a blood vessel behaves like (nonlinear) N N−K−1
elastic material, there will be little hysteresis in the P–V plane during
vessel dilation and contraction, hence the dynamic compliance as where K is the number of parameters in the model, including the
expressed above will converge to the static compliance. estimation of the SSE (Burnham and Anderson, 2002). Model
selection was determined by the difference scores in AICc
Data analysis between models. Hence using the EW model as the reference
model, we computed the difference scores
Spectroscopic data were captured at 7.5 Hz and analysed using a
path length scaling algorithm, as described in Mayhew et al. (1999). ΔAICC = AICC ðVWÞ − AICC ðEWÞ:
This algorithm used a modified version of the Beer–Lambert Law
which included terms of differential path lengths over the range of (c) In order to compare the predictive properties of the EW and the
wavelengths used. The LDF data were captured at 30 Hz and sub- VW models, we re-optimised the 4 Hz data with the two
sampled to 7.5 Hz. Baseline values were calculated from the first 8 s of compliance model parameters (b and τw) clamped at values
 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380 1375

Fig. 4. Comparison of the EW and VW model predictions of the CBV time series (normalised w.r.t. baseline values). The normalised CBV data (grey trace) are superimposed with the
VW model predicted CBV (black broken trace) and the EW model predicted CBV (black trace). The stimulation duration is indicated by the black bar on the time axis. The five data
sets are from a representative subject (subject 3) under conditions of different stimulation.

optimised for the 5 Hz data. In other words we optimised the same The difference scores for SSE and AICc for each animal under each
model parameters ϕ and τv for both models. This was done for stimulation frequency were calculated. All ΔSSE are negative, except
each animal, and the model predicted CBV time series were one (animal 5, 1 Hz) which has a difference score of zero. A t-test on
compared. the SSE scores showed that the SSE for the VW model are significantly
smaller than those of the EW model (p b 0.01), hence the VW model is
Results better at predicting the CBV time series.
The ΔAICc scores are shown in Table 1. As a rule of thumb, a model
The VW model shows much improved performance compared to is significantly better than other models if the corrected AIC score is at
the EW model. Fig. 4 shows the normalised CBV and the VW model least 10 units lower than those of the other models (Burnham and
predicted CBV, superimposed with the EW model predicted CBV for a Anderson, 2002). In 19 out of 24 data sets, ΔAICC are negative with a
representative subject, which clearly illustrates the improvement of magnitude much bigger than 10. In these cases the VW model is
model predictions provided by the VW model at higher stimulus superior than the EW model. For the remaining five data sets, the
frequencies. At lower stimulation frequencies, in this case at 1 Hz, the ΔAICC scores are within ±10, hence the performances of the two
predictions from both models are almost identical because there is models are not significantly different. Note that these five data sets are
little ‘delayed compliance’ in the data. all from conditions with lower stimulus frequencies: three of
them from 1 Hz, and one of each from the 2 and 3 Hz conditions.
The two EW model parameters (ϕ,τv) and the four VW model
parameters (ϕ,τv ,b,τw) were optimised for all 24 data sets and their
Table 1
The difference scores of the corrected Akaike's information criterion for five subjects values are displayed as histograms shown in Fig. 5. Also shown in
under five stimulation frequencies 1, 2, 3, 4 and 5 Hz. Fig. 5 are the means and standard deviations calculated for each
histogram. The statistics for panel (f) of Fig. 5 were calculated by
1 Hz 2 Hz 3 Hz 4 Hz 5 Hz
excluding the outlier. It can be seen that the standard deviations of the
S1 − 7.4⁎ − 2.6⁎ − 123.6 − 12.0 − 328.0
S2 − 15.5 − 31.3 4.2⁎ − 372.8 − 304.9 two parameters common to both models (ϕ,τv), are smaller for the
S3 1.9⁎ − 161.6 − 232.2 − 269.8 − 404.8 VW model than for the EW model. However the differences were not
S4 − 89.4 − 347.7 − 99.4 N/A − 354.9 significant at the p b 0.05 level. The issue of the outliers in the
S5 4.2⁎ − 60.2 − 411.1 − 277.6 − 280.9 distributions for the two compliance model parameters (b,τw) will be
The insignificant scores (with magnitude ≤ 10) are marked with ⁎. addressed in the Discussion section.
1376 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380

Fig. 5. Histograms of model parameters. The mean and the standard deviation of each histogram are displayed inside each subplot. (a) and (b) are histograms of the two parameters
associated with the EW model. (c–f) are histograms of the four parameters associated with the VW model respectively. In panel (f), the symbol ⁎ is used to highlight the fact that
these statistics were calculated excluding one outlier.

Fig. 6 shows the normalised CBV data under the 4 Hz stimulation In summary an important advantage of the VW model over the
frequency, the EW model predicted CBV and that using the VW model EW model is that the values of the two compliance model parameters
but with the compliance parameters (b,τw) clamped at their values (b,τw) capture the variation of the difference between the CBF and the
optimised for the 5 Hz stimulation frequency. It can be seen that apart CBV time series as they return to baseline.
from subject 1, the VW model predictions for the other three subjects Another important feature of the VW model is that it converges to
are significantly better than the predictions of the EW model. The data the Grubb's relationship CBV ∝ CBFΦ (Grubb et al., 1974) at steady
for subject 1 showed a large mismatch between the CBF and CBV time state. The parameter ϕ in the VW model is the inverse of the Grubb's
series at 5 Hz, but only slight difference at 4 Hz. Thus clamping the VW exponent Φ. As ϕ is the summation of the two parameters: α given by
model compliance parameters optimised under the 5 Hz condition did the blood vessel resistance–volume relationship, and β given by the
not produce good predictions for the 4 Hz data for this subject. P–V curve, and both parameters have different values for different
blood vessel types, it means that the Grubb's exponent is unlikely to
Discussion be the same across different vascular compartments. Furthermore,
during transient phases of vessel dilation and contraction, the
Using the Akaike's Information Criterion, we have shown that the relationship between CBF and CBV is a dynamic one which is not
VW model is superior to the EW model, and the two additional model simply governed by the power law described by Grubb et al. (1974).
parameters (b,τw) improved model predictions during the return-
to-baseline phase of the CBV time series. The physiological basis of the CBF–CBV mismatch
There was one outlier in the distribution of the parameter τw, with
a value of τw = 33. For this particular data set (2 Hz, animal 2), there In deriving the VW model, we have attributed the observed
was little mismatch between the two time series of CBF and CBV, i.e., difference between the time series of CBF and CBV to the visco-elastic
they returned to baseline at roughly the same rate. A large τw means properties of the blood vessels. This is based on the following
that the state variable w has very slow dynamics, or hardly varies reasoning.
during the time course of the stimulation. During neural activation, dilation in the arterial compartment
The histogram of the parameter b also highlights some cases in reduces the vascular resistance and hence blood flow increases. In the
which much larger values of b were necessary in order to capture the venous compartment, the increased flow from upstream increases the
data. Closer examination of the associated data revealed that these pressure inside the blood vessels. Assuming intracranial pressure
cases correspond to the time series of CBV remaining elevated remains constant during local hyperaemia (by replacement of the
throughout the post-stimulation period that the data was collected, CSF), the transmural pressure in the venous compartment will be
while CBF did return to its baseline. The role of the parameter b is to increased. Because the venous compartment is passive and the venous
amplify the influence of dv / dt on the state variable w. In these cases, blood vessels are visco-elastic, the volume of the vessels increases
b has to be sufficiently larger to prevent the state variable w from quickly initially due to the elastic component of the vessel walls, but
returning to the steady state value of unity. this expansion slows down considerably as the transmural pressure
 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380 1377

Fig. 6. Comparison of the EW and VW model predictions of the CBV time series using clamped parameters. The normalised CBV data (grey trace) are superimposed with the VW
model predicted CBV (black broken trace) and the EW model predicted CBV (black trace). The four data sets are from four individual subjects (1, 2, 3 and 5) with 4 Hz stimulation
frequency. The VW model predictions were obtained by clamping the two compliance model parameters (b, τw) at values optimised for the 5 Hz condition.

(and hence CBF) reaches a new steady state (Porciuncula et al., 1964). track the sign of the changes of the CBV time series. A further
This is the ‘creep loading’ characteristic of a visco-elastic material. This weakness of the ‘Buxton’ model is that it effectively assumes that
process also occurs when there is a decrease in transmural pressure, during vessel dilation, only the elastic response is present (hence
hence there is a mismatch between the time series of CBF and CBV the shorter time constant τMTT), and during vessel contraction, only
during both the inflation and deflation phases. A recent study by Royl the visco-elastic response is present (hence the longer time constant
et al. (2008) on the effect of elevated intracranial pressure on τMTT + τvisco). Data obtained from long stimulation studies (Herman
neurovascular coupling demonstrated that by increasing the intra- et al., 2008; Kennerley et al., 2005; Kida et al., 2007; Mandeville et al.,
cranial pressure (hence decreasing the transmural pressure) within 1999) show that during stimulation, there are marked differences
an activated region of the rat cortex the temporal mismatch between between the CBF and CBV time series, i.e., typically the CBF reaches a
CBF and CBV was reduced. This suggests that transmural pressure plateau while CBV is still increasing. However for short stimulation
plays an important part in the differences between the time courses of studies, this difference in the time series of CBF and CBV may not be
CBF and CBV. observed because there is insufficient time during the onset phase
Although many optical imaging and MRI studies demonstrated the before the return-to-baseline phase begins. During the return-to-
mismatch between the time series of CBF and CBV, this phenomenon baseline phase of the haemodynamic responses, the initial rates of
is not always observed (Donahue et al., 2008; Jin et al., 2006). Some decrease in CBF and CBV are in fact similar and the dominant observed
studies even showed no measurable changes of CBV in the venous mismatch between the CBF and CBV time series only occurs later.
compartment with respect to evoked changes in neural activity (e.g. To overcome the problem of the hidden switching function (in
Hillman et al., 2007; Kim et al., 2007). This inconsistency across both the windkessel model and the Buxton's model), Kong et al.
laboratories and across studies could be due to different modalities (2004) extended Mandeville's windkessel model by introducing a
used for measuring the haemodynamic responses and different new state variable, which was a function of CBV (but not the rate of
stimulus types. But this alone may not explain the considerable change of CBV), and captured some of the desired time varying
variability in the published physiological data. It is possible that other characteristics of the dynamic compliance. This model was nonlinear
physiological mechanisms, such as changes in haematocrit, also affect time-invariant with four model parameters, much the same as the VW
the observed mismatch. This is still an area of active research. model in terms of model complexity. It was also able to predict
well the slow return-to-baseline characteristics of the CBV time series.
Comparison with other existing CBF–CBV models The model was used (Zheng et al., 2005) to develop a multi-
compartmental model linking stimulus to the BOLD signal. However,
Two other models relating CBF to CBV are often used as part of an because the model by Kong et al. (2004) was a formal model and not
effort to establish a model linking changes in neural activity to the based on the mechanical properties of visco-elastic materials, it was
BOLD signal measured in fMRI experiments. One is similar to the EW difficult to relate the additional state and the parameters of the model
model, except that the time constant τv took on different values to the known physiology.
during blood vessel dilation and constriction (Buxton et al., 2004). In contrast, the VW model presented here is capable of producing
This model has been used in modelling and analysis of the BOLD fMRI both elastic and visco-elastic responses during stimulus onset and
data (Deneux and Faugeras, 2006; Vakorin et al., 2007; Vazquez et al., cessation because of the introduction of the state variable w which
2006). Like the original windkessel model by Mandeville et al. (1999), varies in time. The VW model is in the elastic phase when w is close to
this model is time varying. A hidden switching function is required to unity and transforms itself into the visco-elastic phase as w moves
1378 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380

Fig. 7. The visco-elastic properties of the VW model using a simulated step change in CBF (dotted trace). The computed CBV response (solid trace) shows creep loading characteristics,
and the computed transmural pressure (broken trace) show characteristics of stress relaxation.

 
away from unity. Hence, there is no need to decide when to ‘manually’ w with the term exp −b dv to ensure that within the theoretical
dt
switch to a different time constant. The compliance parameters (b range of dv að−∞; + ∞Þ, the state variable w is always positive.
dt
and τw) determine the speed of this transformation and by doing so On the other hand, many published data showed that the steady
control the degree of the mismatch between the two time series CBF state relationship between the amplitude of changes in CBF and CBV is
and CBV. Furthermore, the model is capable of predicting physiolo- primarily linear (Ito et al., 2003, 2001; Lee et al., 2001; Risberg et al.,
gically meaningful variables. For instance, from the relationship V = 1969; Smith et al., 1971). The dynamic relationship between changes
(APw)1 / β, we can write the normalised expression v = (wpw)1 / β, thus in CBF and CBV was explored (Wang et al., 2007) by linearising the
predicting the normalised transmural pressure Pw across the vessel modified windkessel model (Kong et al., 2004). They showed that
wall (assuming, say lamina flow with α = 2). Fig. 7 shows the results within the range of changes tested (up to 25% change in CBF and 6%
of a simulation of the VW model whose parameters were selected as change in CBV), the linearised model was adequate at predicting the
the mean values of their distributions given in Fig. 5. The normalised time series of changes in CBV from the CBF changes. More recently a
CBF time series was the driving input to the model, and the model linear system identification technique was used (Wei et al., in press)
predicted time series of the normalised CBV and transmural pressure which further demonstrated that changes in CBF and CBV could be
were computed. A step increase in CBF was chosen to demonstrate the modelled by a linear dynamic system. The drawback of such
visco-elastic properties of the VW model: the CBV response showed techniques, however, is that the relationships between the model
the creep loading characteristics, and the transmural pressure showed parameters and any physiological and biophysical constructs are
the stress relaxation characteristics of visco-elastic vessels. unclear.
In addition, the VW model has the unique feature that both the
static and dynamic compliances are emergent properties of the model. Limitations of the VW model
Although these compliances are not explicitly used in the VW model,
they are important concepts and directly relevant in modelling the The VW model is guided by the mechanical properties of visco-
brain vasculature and the blood circulation network (Boas et al., 2008; elastic material, independent of the underlying physiological mecha-
Huppert et al., 2007; Linehan et al., 1986; Ursino et al., 2000). The use nisms coupling haemodynamic changes to evoked changes in neural
of static instead of dynamic compliance implies the assumption that activity. Its function is to account for the mismatch between the time
blood vessels are elastic. This assumption may need to be examined series of CBF and CBV, not to model the temporal shape of the CBF
when multi-compartment models are used, as the assumption may itself due to evoked neural activity. The model is most suited to relate
stand for the arterial compartment, but is less appropriate for the the time series of changes in CBF and CBV in the venous compartment
venous compartment. due either to evoked neural responses or as part of the circulation
vascular network. Unlike the venous compartment which can be
Nonlinear vs. linear models modelled as a passive leather bag (windkessel), the cerebral arterial
network is actively involved in its role as a cerebral auto-regulation
The VW model relates the normalised change in CBF to that of CBV. mechanism (Kontos, 1981; Paulson et al., 1990). It has been shown
It is constrained by two nonlinear relationships: (1) the steady state that cerebral arteries of certain diameter range will dilate with a drop
relationship between the normalised CBF and CBV which is widely in perfusion pressure but constrict with an increase in perfusion
accepted as a power law relationship (Grubb et al., 1974) rather than a pressure in order to maintain cerebral blood flow (Kontos et al., 1978).
linear one, and (2) the steady state pressure–volume curve of blood This is in direct contrast to the P–V relationship of blood vessels
vessels which is also well known to be nonlinear; with changes in measured in vitro. Thus the development of arterial models will be
volume diminishing at high transmural pressure. The VW model needed for detailed multi-compartment modelling of the relationship
introduced a further nonlinearity by modelling the state variable between changes in blood flow and volume in the cerebral system.
 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380 1379

It is also worthwhile noting that the steady state P–V relationship Hillman, E.M.C., Devor, A., Bouchard, M.B., Dunn, A.K., Krauss, G.W., Skoch, J., Bacskai, B.J.,
Dale, A.M., Boas, D.A., 2007. Depth-resolved optical imaging and microscopy of vascular
V = (AP)1 / β is a qualitative description of the P–V curve of blood compartment dynamics during somatosensory stimulation. NeuroImage 35, 89–104.
vessels rather than an optimal fit to experimental data. It seems Huppert, T.J., Allen, M.S., Benav, H., Jones, P.B., Boas, D.A., 2007. A multicompartment
appropriate for describing the P–V relationship of a vein (Porciuncula vascular model for inferring baseline and functional changes in cerebral oxygen
metabolism and arterial dilation. J. Cereb. Blood Flow Metab. 27, 1262–1279.
et al., 1964), but other studies showed that the P–V relationship of an Ito, H., Takahashi, K., Hatazawa, J., Kim, S.-G., Kanno, I., 2001. Changes in human regional
arterial blood vessel maybe better described by a sigmoid function cerebral blood flow and cerebral blood volume during visual stimulation measured
(Bergel, 1961; Roy, 1881). Notwithstanding, the strategy of modelling by positron emission tomography. J. Cereb. Blood Flow Metab. 21, 608–612.
Ito, H., Kanno, I., Ibaraki, M., Hatazawa, J., Miura, S., 2003. Changes in human cerebral blood
the parameter of the steady state P–V curve can be extended to other flow and cerebral blood volume during hypercapnia and hypocapnia measured by
forms of P–V relationships. positron emission tomography. J. Cereb. Blood Flow Metab. 23, 665–670.
Jin, T., Wang, T., Zhao, F., Wang, P., Tasker, M., Kim, S.-G., 2006. Spatiotemporal
characteristics of BOLD, CBV and CBF responses in the cat visual cortex. Proc. Intl.
Conclusion
Soc. Mag. Reson. Med. 14, 2762.
Jones, R.T., 1969. Blood flow. Annu. Rev. Fluid Mech. 1, 223–244.
This paper proposed a novel strategy for modelling the relation- Jones, M., Berwick, J., Johnston, D., Mayhew, J., 2001. Concurrent optical imaging
ship between the CBF and the CBV time series. The strategy is based on spectroscopy and laser-Doppler flowmetry: the relationship between blood flow,
oxygenation, and volume in rodent barrel cortex. NeuroImage 13, 1002–1015.
the observation that the pressure–volume relationship of a visco- Jones, M., Berwick, J., Mayhew, J., 2002. Changes in blood flow, oxygenation, and
elastic tube at steady state differs from the relationship during volume following extended stimulation of rodent barrel cortex. NeuroImage 15,
transient states of dilation and contraction. The key point of the VW 474–487.
Kennerley, A.J., Berwick, J., Martindale, J., Johnston, D., Papadakis, N., Mayhew, J., 2005.
model is that a parameter characterising the steady state P–V curve Concurrent fMRI and optical measures for the investigation of the hemodynamic
becomes time varying during transient states, and the model response function. MRM 54, 354–365.
predicted P–V trajectory form a hysteresis loop. This parameter is Kida, I., Rothman, D.L., Hyder, F., 2007. Dynamics of changes in blood flow, volume, and
oxygenation: implications for dynamic functional magnetic resonance imaging
treated as a state variable and modelled by a simple first order calibration. J. Cereb. Blood Flow Metab. 27, 690–696.
nonlinear dynamic system driven by the rate of change of volume, and Kim, T., Hendrich, K.S., Masamoto, K., Kim, S.G., 2007. Arterial versus total blood volume
the overall model is nonlinear time-invariant. changes during neural activity-induced cerebral blood flow change: implication for
BOLD fMRI. J. Cereb. Blood Flow Metab. 27, 1235–1247.
Although the VW model has two more parameters than the Klabunde, R.E., 2005. Cardiovascular Physiology Concepts. Lippincott Williams &
simpler elastic balloon model, based on Akaike's information criterion, Wilkins, US.
it is superior. The values of these parameters account for the difference Kong, Y.Z., Zheng, Y., Johnston, D., Martindale, J., Jones, T., Billings, S., Mayhew, T., 2004. A
model of the dynamic relationship between blood flow and volume changes during
between the time series of CBF and CBV, thus capturing the ‘delayed
brain activation. J. Cereb. Blood Flow Metab. 24, 1382–1392.
compliance’ phenomenon. Kontos, H.A., 1981. Regulation of the cerebral-circulation. Annu. Rev. Physiol. 43,
397–407.
Acknowledgments Kontos, H.A., Wei, E.P., Navari, R.M., Levasseur, J.E., Rosenblum, W.I., Patterson, J.L., 1978.
Responses of cerebral-arteries and arterioles to acute hypotension and hypertension.
Am. J. Physiol. 234, H371–H383.
The authors are grateful to Drs. J Berwick, M. Jones and J. Martindale Lee, S.P., Duong, T.Q., Yang, G., Iadecola, C., Kim, S.G., 2001. Relative changes of cerebral
for providing the physiological data, and to Dr David Johnston for his arterial and venous blood volumes during increased cerebral blood flow:
implications for BOLD fMRI. Magn. Reson. Med. 45, 791–800.
advice over the years on mathematical rigour. The work was supported Linehan, J.H., Dawson, C.A., Rickaby, D.A., Bronikowski, T.A., 1986. Pulmonary vascular
by the Engineering and Physical Science Research Council grant: EP/ compliance and viscoelasticity. J. Appl. Physiol. 61, 1802–1814.
D001218, the National Institute of Health grant: ROI-NS44567-01 and Mandeville, J.B., Marota, J.J.A., Ayata, C., Zaharchuk, G., Moskowitz, M.A., Rosen, B.R.,
Weisskoff, R.M., 1999. Evidence of a cerebrovascular postarteriole windkessel with
the Medical Research Council Coop group grant: 9825307. delayed compliance. J. Cereb. Blood Flow Metab. 19, 679–689.
Martindale, J., Mayhew, J., Berwick, J., Jones, M., Martin, C., Johnston, D., Redgrave, P.,
References Zheng, Y., 2003. The hemodynamic impulse response to a single neural event.
J. Cereb. Blood Flow Metab. 23, 546–555.
Alexander, R.S., Edwards, W.S., Ankeney, J.L., 1953. The distensibility characteristics of Mayhew, J., Zheng, Y., Hou, Y.Q., Vuksanovic, B., Berwick, J., Askew, S., Coffey, P., 1999.
the portal vascular bed. Circ. Res. 1, 271–277. Spectroscopic analysis of changes in remitted illumination: the response to
Bergel, D.H., 1961. The static elastic properties of arterial wall. J. Physiol. London 156, increased neural activity in brain. NeuroImage 10, 304–326.
445–457. Molino, P., Cerutti, C., Julien, C., Cuisinaud, G., Gustin, M.P., Paultre, C., 1998. Beat-to-beat
Boas, D.A., Jones, S.R., Devor, A., Huppert, T.J., Dale, A.M., 2008. A vascular anatomical estimation of windkessel model parameters in conscious rats. Am. J. Physiol. Heart
network model of the spatio-temporal response to brain activation. NeuroImage 40, Circ. Physiol. 43, H171–H177.
1116–1129. Nilsson, G.E., 1984. Signal processor for laser Doppler tissue flowmeters. Med. Biol. Eng.
Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference, Second Comput. 22, 343–348.
ed. Springer. Orosz, M., Molnarka, G., Toth, M., Nadasy, G.L., Monos, E., 1999. Viscoelastic behavior of
Buxton, R.B., Wong, E.C., Frank, L.R., 1998. Dynamics of blood flow and oxygenation vascular wall simulated by generalized Maxwell models — a comparative study.
changes during brain activation: the balloon model. Magn. Reson. Med. 39, Med. Sci. Monit. 5, 549–555.
855–864. Paulson, O.B., Strandgaard, S., Edvinsson, L., 1990. Cerebral autoregulation. Cerebrovasc.
Buxton, R.B., Uludag, K., Dubowitz, D.J., Liu, T.T., 2004. Modeling the hemodynamic Brain Metab. Rev. 2, 161–192.
response to brain activation. NeuroImage 23, S220–S233. Porciuncula, C.I., Armstrong, G.G., Stone, H.L., Guyton, A.C., 1964. Delayed compliance in
Deneux, T., Faugeras, O., 2006. Using nonlinear models in fMRI data analysis: model external jugular vein ofdog. Am. J. Physiol. 207, 728–732.
selection and activation detection. NeuroImage 32, 1669–1689. Remington, J.W., Alexander, R.S., 1955. Stretch behavior of the bladder as an approach to
Donahue, M.J., Stevens, R.D., de Boorder, M., Pekar, J.J., Hendrikse, J., van Zijl, P.C.M., vascular distensibility. Am. J. Physiol. 181, 240–248.
2008. Hemodynamic changes after visual stimulation and breath holding provide Risberg, J., Ancri, D., Ingvar, D.H., 1969. Correlation between cerebral blood volume and
evidence for an uncoupling of cerebral blood flow and volume from oxygen cerebral blood flow in the cat. Exp. Brain Res. 8, 321–326.
metabolism. J. Cereb. Blood Flow Metab. 29, 176–185. Roy, C.S., 1881. The elastic properties of the arterial wall. J. Physiol. 3, 125–159.
Fogliardi, R., Di Donfrancesco, M., Burattini, R., 1996. Comparison of linear and nonlinear Royl, G., Fuchtemeier, M., Leithner, C., Kohl-Bareis, M., Dirnagl, U., Lindauer, U., 2008.
formulations of the three-element windkessel model. Am. J. Physiol. Heart Circ. Influence of intracranial pressure on neurovascular coupling: clarifying mecha-
Physiol. 40, H2661–H2668. nisms of the BOLD post-stimulus undershoot. Soc. Neurosci. 2008 (Abstract 482.6).
Frank, O., 1930. Evaluation of the shock volume of the human heat on the basis of waves Segers, P., Rietzschel, E.R., De Buyzere, M.L., Stergiopulos, N., Westerhof, N., Van Bortel,
and Windkessel theory. Z. Biol. 90, 405–409. L.M., Gillebert, T., Verdonck, P.R., 2008. Three- and four-element windkessel
Friston, K.J., Mechelli, A., Turner, R., Price, C.J., 2000. Nonlinear responses in fMRI: the models: assessment of their fitting performance in a large cohort of healthy middle-
balloon model, volterra kernels, and other hemodynamics. NeuroImage 12, aged individuals. Proc. Inst. Mech. Eng., H J. Eng. Med. 222, 417–428.
466–477. Smith, A.L., Neufeld, G.R., Ominsky, A.J., Wollman, H., 1971. Effect of arterial CO2 tension on
Grubb, R.L., Raichle, M.E., Eichling, J.O., Terpogos, M.M., 1974. Effects of changes in Paco2 cerebral blood flow, mean transit time, and vascular volume. J. Appl. Physiol. 31, 701–707.
on cerebral blood volume, blood flow, and vascular mean transit time. Stroke 5, Stergiopulos, N., Westerhof, B.E., Westerhof, N., 1999. Total arterial inertance as the fourth
630–639. element of the windkessel model. Am. J. Physiol. Heart Circ. Physiol. 276, H81–H88.
Herman, P., Sanganahalli, B.G., Hyder, F., 2008. Multimodal measurements of blood Ursino, M., Ter Minassian, A., Lodi, C.A., Beydon, L., 2000. Cerebral hemodynamics
plasma and red blood cell volumes during functional brain activation. J. Cereb. during arterial and CO2 pressure changes: in vivo prediction by a mathematical
Blood Flow Metab. 29, 19–24. model. Am. J. Physiol. Heart Circ. Physiol. 279, H2439–H2455.
1380 Y. Zheng, J. Mayhew / NeuroImage 47 (2009) 1371–1380

Vakorin, V.A., Krakovska, O.O., Borowsky, R., Sarty, G.E., 2007. Inferring neural activity data-based modelling approach. IEEE Trans Biomedical Engineering. doi 10.1109/
from BOLD signals through nonlinear optimization. NeuroImage 38, 248–260. TBME.2009.2012722. http://www.ncbi.nlm.nih.gov/pubmed/19174333.
Vazquez, A.L., Cohen, E.R., Gulani, V., Hernandez-Garcia, L., Zheng, Y., Lee, G.R., Kim, S.G., Westerhof, N., Elzinga, G., Sipkema, P., 1971. An Artificial arterial system for pumping
Grotberg, J.B., Noll, D.C., 2006. Vascular dynamics and BOLD fMRI: CBF level effects hearts. J. Appl. Physiol. 31, 776–781.
and analysis considerations. NeuroImage 32, 1642–1655. Zheng, Y., Martindale, J., Johnston, D., Jones, M., Berwick, J., Mayhew, J., 2002. A model
Wang, K.Y., Zheng, Y., Billings, S., Coca, D., Johnston, D., Mayhew, J., 2007. Estimating CBF of the hemodynamic response and oxygen delivery to brain. NeuroImage 16,
changes from measurements of CBV changes: a linear dynamic modelling 617–637.
approach. Soc. Neurosci. 2007 (Abstract 753.17). Zheng, Y., Johnston, D., Berwick, J., Chen, D.M., Billings, S., Mayhew, J., 2005. A three-
Wei, H.L., Zheng, Y., Pan, Y., Coca, D., Li, L.M., Mayhew, J., Billings, S., in press. Model compartment model of the hemodynamic response and oxygen delivery to brain.
estimation of cerebral hemodynamics between blood flow and volume changes: a NeuroImage 28, 925–939.