rsif.royalsocietypublishing.org
ResearchCite this article: Pahlevan NM, Tavallali P,
Rinderknecht DG, Petrasek D, Matthews RV,
Hou TY, Gharib M. 2014 Intrinsic frequency for
a systems approach to haemodynamic
waveform analysis with clinical applications.
J. R. Soc. Interface 11: 20140617.
http://dx.doi.org/10.1098/rsif.2014.0617
Received: 9 June 2014
Accepted: 13 June 2014
Subject Areas:biomechanics, bioengineering, biomathematics
Keywords:pulse wave analysis, systems science,
cardiovascular disease, ventricular/arterial
coupling, instantaneous frequency,
arterial pressure
Author for correspondence:Morteza Gharib
e-mail: [email protected]
Electronic supplementary material is available
at http://dx.doi.org/10.1098/rsif.2014.0617 or
via http://rsif.royalsocietypublishing.org.
& 2014 The Author(s) Published by the Royal Society. All rights reserved.
Intrinsic frequency for a systems approachto haemodynamic waveform analysiswith clinical applications
Niema M. Pahlevan1, Peyman Tavallali2, Derek G. Rinderknecht3,Danny Petrasek4, Ray V. Matthews5, Thomas Y. Hou2 and Morteza Gharib3
1Medical Engineering, Division of Engineering and Applied Sciences, California Institute of Technology,1200 East California Boulevard, MC 301-46, Pasadena, CA 91125, USA2Applied and Computational Mathematics, Division of Engineering and Applied Sciences, California Institute ofTechnology, 1200 East California Boulevard, MC 9-94, Pasadena, CA 91125, USA3Graduate Aerospace Laboratories, Division of Engineering and Applied Sciences, California Institute ofTechnology, 1200 East California Boulevard, MC 205-45, Pasadena, CA 91125, USA4Medical Engineering, Division of Engineering and Applied Sciences, California Institute of Technology,1200 East California Boulevard, MC 217-50, Pasadena, CA 91125, USA5Keck School of Medicine, University of Southern California, 1510 San Pablo Street, Suite 322, Los Angeles,CA 90033, USA
The reductionist approach has dominated the fields of biology and medicine for
nearly a century. Here, we present a systems science approach to the analysis of
physiological waveforms in the context of a specific case, cardiovascular physi-
ology. Our goal in this study is to introduce a methodology that allows for novel
insight into cardiovascular physiology and to show proof of concept for a new
index for the evaluation of the cardiovascular system through pressure wave
analysis. This methodology uses a modified version of sparse time–frequency
representation (STFR) to extract two dominant frequencies we refer to as intrin-
sic frequencies (IFs; v1 and v2). The IFs are the dominant frequencies of the
instantaneous frequency of the coupled heart þ aorta system before the closure
of the aortic valve and the decoupled aorta after valve closure. In this study, we
extract the IFs from a series of aortic pressure waves obtained from both clinical
data and a computational model. Our results demonstrate that at the heart rate at
which the left ventricular pulsatile workload is minimized the two IFs are equal
(v1 ¼ v2). Extracted IFs from clinical data indicate that at young ages the total
frequency variation (Dv ¼ v1 2 v2) is close to zero and that Dv increases with
age or disease (e.g. heart failure and hypertension). While the focus of this
paper is the cardiovascular system, this approach can easily be extended to
other physiological systems or any biological signal.
1. IntroductionThe reductionist approach has dominated the science of biology and medicine.
While this approach has provided valuable information, it has not necessarily
translated into increased understanding. From a clinical standpoint, medicine
has been driven by the perpetuation of a normal physiological system,
yet symptoms are often managed through addressing singular parameters to
be corrected and not viewed as indicative of wider systemic problems. In this
context, we will present a systems science approach to cardiovascular physi-
ology by developing the concept of intrinsic frequency (IF) as a systems
parameter. Although generally speaking this methodology is applicable to
any physiological waveform, our focus in this paper will be its application to
arterial pressure waveforms. In doing so, we will show that using such a par-
ameter as a complement to well-established pathological markers may help
us create new pathways for discovery and increase our ability to predict
clinically relevant outcomes.
Analysing arterial waveforms is clinically important because it provides
information about states of health and disease [1]. Significant efforts have been
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
2
made in the past to elucidate the complex interaction betweenthe left ventricular (LV) and wave dynamics of the large cen-
tral arteries such as the aorta [1–5]. However, extracting
reliable information from haemodynamic waves about health
or disease conditions remains a significant challenge to
modern medicine. It is well accepted that the dynamics of
the left ventricle, arterial wave dynamics and the interaction
between the two determine the arterial pressure wave [1,2,6].
This means the pressure wave contains information about
these dynamic systems and their optimum coupling. This opti-
mum coupling can be impaired owing to increased arterial
stiffness, ageing, smoking or disease conditions, such as hyper-
tension and heart failure (HF) [1]. Therefore, it follows that
these waves also carry information about the diseases of the
heart, vascular disease (VD), as well as the coupling of the
heart and the arterial network [2,6–8].
There are several methods for analysing arterial pulse
waveforms [1,9]. Some of these methods are based on fre-
quency domain methodologies such as the impedance
method (Fourier method), whereas others are based on time
domain methodologies such as wave intensity analysis [10].
Both frequency and time domain methods give complemen-
tary results [9,11]. These methods however require both
pressure and flow waves to be measured simultaneously at
the same location, which is clinically difficult, if not impossible.
Our approach is based on a newly developed sparse
time–frequency representation (STFR) method [12] and an
analysis of the systemic coupling between the dynamics of
LV contraction and the dynamics of waves in the arterial net-
work. The STFR method is inspired by the empirical mode
decomposition (EMD) method [13] and provides a more
systematic way to define instantaneous frequency. Similar
to the EMD method, the STFR method is well suited to ana-
lyse nonlinear non-stationary data, is less sensitive to noise
perturbation and preserves some of the intrinsic physical
properties of the signal [12,14]. Although the application of
the EMD method to biological problems has been introduced
by Huang et al. [15,16], this paper will show the potential of
this concept to diagnose heart and VDs as well as its potential
to quantify the optimum coupling between the heart and
arterial system.
We begin with the premise that the left ventricle of the
heart and aorta construct a coupled dynamic system before
the closure of the aortic valve. The onset of aortic valve clo-
sure is marked by the dicrotic notch on the aortic input
pressure wave. This coupled dynamic system has a domi-
nant frequency that the instantaneous frequency oscillates
around (note that this frequency is not the resonant fre-
quency) which is not necessarily constant over the cardiac
cycle. This dominant frequency is influenced by the dynamics
of both the heart and the aorta. After valve closure, the heart
and the aorta are decoupled from each other. This means that
the dominant frequency is dictated only by the dynamics
of the aorta and its branches (arterial network). From this
point forward in this paper, we will refer to the heart and
arterial network as the heart þ aorta system.
By applying the STFR method, it is possible to compute the
instantaneous frequency of the coupled heart þ aorta system
and the decoupled aorta system from the aortic pressure
wave alone. The application of this technique to aortic pressure
waves led us to the observation that the instantaneous
frequency oscillates around different dominant frequencies
before and after the dicrotic notch, or closure of the aortic
valve. These instantaneous frequencies are not necessarily con-
stant in time, but do represent the dominant frequencies at any
instant time throughout the cardiac cycle beginning with the
coupled system of the heart and aorta prior to the dicrotic
notch and the decoupled system of the aorta itself afterwards.
We refer to these dominant frequencies as intrinsic frequencies
(IFs; v1 and v2). To extract the IF directly from the pressure
waveform, a modified version of STFR was developed using
a norm-2 (L2) minimization method and a brute-force algor-
ithm was applied to solve the problem. This algorithm
considers all possible values of frequencies to ensure that the
corresponding minimizer frequencies are in fact the unique
minimizers of the problem. In this regard, the piecewise con-
stant frequency before the dicrotic notch is the IF of the
heart–aorta system and the one after the dicrotic notch is the
IF of the aortic system. The main advantage of this method
in contrast to well-known and widely used impedance and
wave intensity methods is that only one arterial pressure wave-
form is required to perform the analysis [1,10].
Here, we show proof of concept of the IF as a new medical
index for the identification of the optimum left ventriculoar-
terial coupling and for diagnosis of cardiovascular disease
(CVD). The idea of the IF concept is based on our obser-
vations that when we apply the adaptive STFR method to
an aortic pressure wave to extract the instantaneous fre-
quency ( _u1(t) ¼ du1=dt) of the first intrinsic mode function
(IMF; where u1 is the phase angle of the first IMF) we see
that a dominant instantaneous frequency exists on either
side of the dicrotic notch. A computational fluid dynamics
(CFD) model was constructed to study the link between the
IFs (v1 and v2) and LV pulsatile power workload across a
range of heart rates (HRs) and aortic rigidities. Finally, our
analysis method will be applied to a small sample set of clini-
cal data from human subjects to establish proof of concept for
the IF method in the diagnosis of CVD.
2. Material and methods2.1. Adaptive method of sparse time – frequency
representationThe notion of the IMF was first introduced by Huang et al. [13].
A more mathematical definition of the IMF is given by Hou & Shi
[12], as follows.
A signal f (t) is called an IMF if there exists an envelope, a(t) .
0, and a phase function, u(t), satisfying three properties: (i) a(t)is smoother than cosu(t), (ii) u(t) is strictly increasing in
time, and (iii) the IMF has only one extremum between two
consecutive zeros,
f (t) ¼ a(t) cosu(t), t [ <: (2:1)
A real signal s(t) is called an intrinsic signal if it can be decom-
posed into a finite sum of IMFs
s(t) ¼XM
i¼1ai(t) cos ui(t): (2:2)
The essential idea behind the STFR is to find the sparsest rep-
resentation of multi-scale data within the largest possible
dictionary of IMFs. This huge dictionary consists of elements
(or bases) that are not defined a priori. The use of an infinite
dimensional highly redundant data-driven basis is what makes
the STFR truly adaptive. Based on an approximation, the STFR
method can be reduced to an L2 minimization problem [14] for
r
3
periodic signals. The description of the L2-STFR algorithm isprovided in appendix A.
sif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
2.2. Modified sparse time – frequency representation forheart – aorta system: intrinsic frequency algorithm
In our proposed method, we assume that the instantaneous fre-
quency of the coupled heart–aorta and decoupled aorta are
piecewise constant in time. This enables us to extract the IFs
directly from the arterial pressure wave. The IF is the frequency
that carries the maximum power in equation (2.2). To extract the
IF, we propose a simple but effective norm-2 (L2) minimization
method. The envelopes of the IMF are also assumed to be piece-
wise constant in time to distinguish between the two systems.
Hence, the L2 minimization problem, for the extraction of the
trend and frequency content of the input aortic pressure wave,
is proposed as follows:
min:
f (t)� x(0, T0)s1(t)� x(T0, T)s2(t)� ck k22, (2:3)
subject to:
a1 cos (v1T0)þ b1 sin (v1T0) ¼ a2 cos (v2T0)þ b2 sin (v2T0), (2:4)
a1 ¼ a2 cos (v2T)þ b2 sin (v2T), (2:5)
s1(t) ¼ a1 cos (v1t)þ b1 sin (v1t) (2:6)
and s2(t) ¼ a2 cos (v2t)þ b2 sin (v2t): (2:7)
Here, x(a, b) ¼ 1 a � t � b0 otherwise,
�(2:8)
and c is a constant.
This problem is now reduced to solving for a1, a2, c, b1, b2, v1
and v2. Equations (2.4) and (2.5) are linear constraints that ensure
the continuity of the trend at the time T0 (dicrotic notch) and the
periodicity of the trend, respectively. This minimization states
that the aortic input pressure wave can be approximated by
two incomplete sinusoids with different frequencies (v1 and
v2), which we refer to as IFs. Where v1 is the IF for the
heart þ aorta system (before aortic valve closure ¼ before dicro-
tic notch), and v2 is the IF for the decoupled aorta (after aortic
valve closure ¼ after dicrotic notch).
The original minimization problem is not convex. Thus, we
may have several local minima. To find the global minimum,
we use a brute-force algorithm over all possible values of
frequencies to ensure that the corresponding minimizer frequen-
cies (v1, v2)m are in fact the unique global minimizer frequencies
of the original minimization problem. The details of the
brute-force algorithm are provided in appendix B.
2.3. Computational aortaA physiologically relevant computational fluid dynamics (CFD-
FSI) model of the aorta with fluid–solid interaction (FSI) was
used. The methods as well as the physical parameters of the
model were the same as those described in Pahlevan & Gharib
[7], in which full details of the computational model were pro-
vided. Simulations were performed for different levels of aortic
rigidities (compliances) labelled E1 through E3, where E1 is the
aortic rigidity of a 30-year-old healthy individual [1]. All other
Ei are multiplicative factors of E1 defined as follows: E2 ¼
1.25E1 and E3 ¼ 3E1. At each Ei, simulations were completed
for eight HRs: 70.5, 75, 89.5, 100, 120, 136.4, 150 and 187.5
beats per minute (bpm). Information about the physical model,
mathematical model, inflow boundary condition and outflow
boundary condition as well as all other model parameters such
as cardiac output (CO), terminal resistance, terminal compliance
and the shape of the inflow wave are detailed in appendix C.
2.4. Clinical dataTo examine the potential clinical relevance of the IF method, data
were first gathered from published works [1]. In addition to the
publically available data, invasive blind clinical data were
obtained from patients having clinically indicated procedures in
the cardiac catheterization laboratory at Keck Medical Center, Uni-
versity of Southern California, USA (USC). Retrospective de-
identified data were analysed for 16 consecutive blinded patient
datasets. The data were collected as part of routine medical pro-
cedures using 6F fluid-filled catheters. All clinical data were
abstracted from the Keck Medical Center cardiac catheterization
laboratory research database and approved by the University of
Southern California Institutional Review Board.
3. Results3.1. The intrinsic frequency of aortic pressure wavesA series of aortic pressure waves were examined to observe
the behaviour of the adoptive STFR method. It was observed
that the instantaneous frequency oscillates around one domi-
nant frequency range at the beginning of the cardiac cycle
and then shifts and oscillates around a second range of domi-
nant frequencies (see appendix A). This implies that there is a
different dominant frequency within each band, the first
associated with the heart–aorta system and the second
with the arterial system alone. It must be mentioned that
the IFs are the dominant instantaneous frequencies and in
this regard are fundamentally different from resonant
frequencies. To seamlessly extract these dominant frequen-
cies, we created a modified version of the STFR for the
heart–aorta system called the IF algorithm. Figure 1 shows
the application of this algorithm to a number of exemplary
aortic pressure waveforms. Figure 1a shows a typical aortic
pressure waveform as well as the location of the dicrotic
notch. Figure 1b shows the same aortic pressure waveform
with the corresponding piecewise reconstruction using only
the two IFs (v1 and v2) of the first mode IMF overlaid on
top of the original pressure waveform. For clarity, the portions
of the reconstructed waveform that correspond to v1 and v2,
namely the systolic and diastolic phases, are shown in purple
and green, respectively. To further illustrate this behaviour, over-
lays of two other types of aortic pressure waveforms and their
reconstructions are provided in figure 1c,d. In all cases shown
in figure 1, we see good agreement between the shape of the sys-
tolic and diastolic portions produced by the IFs and the original
aortic pressure waveform.
3.2. Optimum heart rate prediction from the intrinsicfrequencies
A CFD model was constructed to examine the relevance of the
IFs to the pulsatile power workload on the left ventricle. Pulsa-
tile power Ppulse was calculated using the following equation:
Ppulse ¼1
T
ðT
0
p(t)q(t) dt� pmeanqmean, (3:1)
where p(t) is the pressure, q(t) is the flow, pmean is the mean
pressure, qmean is the mean flow and T is the period of the car-
diac cycle. The computational model used to generate the aortic
pressure waveforms is described in Pahlevan & Gharib [7]. The
results of this investigation are shown in figure 2 for three levels
of aortic rigidity: E1, E2¼ 1.5E1 and E3 ¼ 3E1. When the two IF
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.070
80
90
100
110
120
130
time (s)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.850
60
70
80
90
100
110
120
130
140
150
time (s)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.880
90
100
110
120
130
time (s)
pres
sure
(m
m H
g)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.850
60
70
80
90
100
110
120
130
140
150
time (s)
pres
sure
(m
m H
g)
(a) (b)
(c) (d )
Figure 1. The IF reconstruction of aortic pressure waveforms. (a) A typical aortic pressure waveform as well as the location of the dicrotic notch (marked by the redline). (b) The piecewise reconstruction of the pressure wave given in (a) overlaid on top of the original pressure waveform (blue). The portions of the reconstructedwaveform represented by v1 and v2 are shown in purple and green, respectively. (c) An overlay of a different type of aortic pressure waveform (blue) and itsreconstruction ( purple and green). (d ) An overlay of another type of aortic pressure waveform (different characteristics from (b) and (c)) and its reconstruction. Onlythe two IFs (v1 and v2) of the first mode IMF have been used in the reconstruction of systolic (v1) and diastolic (v2) portions. A good agreement between theshape of the systolic and diastolic portions produced by the IFs and the original aortic pressure waveform can be seen.
60 80 100 120 140 160 180 20040
60
80
100
120
heart rate (bpm)
puls
atile
pow
er (
mW
)
50
100
150
200
250
300
350(a)(i) (i) (i)
(ii) (ii) (ii)
(b) (c)
IF (
bpm
)
60 80 100 120 140 160 180 20040
60
80
100
120
heart rate (bpm)
50
100
150
200
250
300
350
60 80 100 120 140 160 180 20050
100
150
200
250
300
350
400
heart rate (bpm)
50
100
150
200
Figure 2. IFs (a(i), b(i) and c(i)) and pulsatile power (a(ii), b(ii) and c(ii)) versus HR. (a) The aortic rigidity of a healthy 30 year old, E1, yields two IF curves thatintersect at an optimum HR � 110 bpm at which the pulsatile power is minimized. (b) The effect of a 50% increase in aortic rigidity (E2 ¼ 1.5E1) yields two curveswhich cross each other at an optimum HR � 140 bpm. (c) An aortic rigidity of E3 ¼ 3E1 yields two IF curves that intersect an optimum HR � 190 bpm. Note: v1
denoted by the red curve is the IF for the coupled heart þ aorta, whereas v2 denoted by the blue curve is the IF for the decoupled aorta.
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
4
0 10 20 30 40 50 60 70 80 90−20
−10
0
10
20
30
40
50
60
70
80
90
100
110
120
age (years)
Dw (
bpm
)
+
+
+
+
+
+
+
++ +
+
++
++
healthy (published data)linear fit (healthy published data)VD (blind clinical data)HF (blind clinical data)HF (published data)
Figure 3. TFV (Dv ¼ v1 2 v2) versus age for healthy and CVD conditions. The IFs, v1 and v2, are close to one another at young ages therefore the differenceis very near zero. A linear fit shows that the difference, Dv, increases with age in healthy subjects. Furthermore, it can be observed that CVD increases Dv as thedisease shifts the ventricular – arterial system out of optimum coupling. In this study, all CVD patients displayed a Dv . 60 bpm. All published waveform dataused to produce this plot were taken from Nichols et al. [1]. Detailed explanations about the ages for the data marked with a plus sign (þ) are given in theelectronic supplementary material. HF refers specifically to HF with LV systolic dysfunction. (Online version in colour.)
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
5
curves, v1 and v2, are graphed as a function of HR, remarkably,
the two curves always intersect at the optimum computed HR
at which the LV pulsatile workload is minimized (figure 2). In
other words, the LV pulsatile workload reaches its minimum
when the two IFs become equal.
The plots of IF and pulsatile power versus HR in figure 2
also clearly show that at increased levels of aortic rigidity the
optimum HR shifts to the right. For example, changing aortic
stiffness threefold increases the optimum HR from 110 bpm
to approximately 185 bpm (E1 versus E3). Additionally,
from figure 2 it can be noted that high aortic rigidities have
a greater effect on pulsatile workload in the range of physio-
logical resting HRs. For example, given a resting HR of
80 bpm, an aortic stiffness of E1, E2 and E3 results in pulsatile
power workloads of 50, 95 and 325 mW, respectively.
3.3. Total frequency variation (Dv): an indexfor cardiovascular health and disease
When a similar analysis examining Dv ¼ v1 2 v2 is applied
to a survey of published clinical data taken from healthy sub-
jects of increasing age as shown in figure 3, we see a clear
physiological pattern. This suggests that Dv is near zero at
young ages when the heart–arterial system is operating
close to the optimum state and that Dv increases with age.
The survey was further extended to include published clinical
data from subjects with HF with LV systolic dysfunction in
addition to clinical data from subjects with vascular disease
(VD) and with HF with LV systolic dysfunction gathered
through collaboration with the catheterization laboratory at
USC. The analysis of these aortic pressure waveforms,
shown in figure 3, demonstrates that, in addition to ageing,
CVD also increases Dv owing to the ventricular–arterial
system shifting from its optimum coupling.
3.4. First intrinsic frequency (v1): a medical indexfor heart disease
After observing the behaviour of Dv in response to ageing
and CVD, we were motivated to investigate the physiologi-
cal information contained in the individual IFs. Since the
dynamics of the heart–arterial system are dominated by
the dynamics of the heart before aortic valve closure, we
anticipated that v1 would be affected by pathophysiological
conditions that impair the pumping dynamics of the heart
such as HF with LV systolic dysfunction. As shown in
figure 4, by examining the v1 for the subset of subjects includ-
ing the published healthy and HF data from figure 3, we
observe that v1 becomes elevated in HF with LV systolic dys-
function and otherwise remains relatively constant under
healthy conditions as age advances. For example, all subjects
with HF in our data population exhibited a v1 above
120 bpm. By contrast, normal healthy subjects displayed a
v1 below 112 bpm.
3.5. Second intrinsic frequency (v2): a medical indexfor vascular disease
The aorta and arterial networks dominate the dynamics of
the heart–arterial system after aortic valve closure. Hence,
v2 is likely to be affected by VDs such as arterial stiffening
and hypertension. As seen in figure 5, if we examine v2 for
the subset of subjects including the published healthy and
VD data displayed in figure 3, we observe that among
healthy individuals v2 decreases with age, which can be
indicative of increasing arterial rigidity [8]. Figure 5 also
shows that v2 drops significantly with certain VDs such as
hypertension and peripheral VDs, in most cases dropping
below 36 bpm (figure 5).
10 20 30 40 50 60 70 8080
90
100
110
120
130
140
150
160
age (years)
w1
(bpm
)
+
+
+
+
+
+
++
++
+
+
+
+
+
healthy (published data)HF (blind clinical data)HF (published data)
Figure 4. v1 at different ages and under healthy and HF conditions. All subjects with HF (LV systolic dysfunction) show v1 . 120 bpm (above the top dashedblack line). Normal subjects show v1 , 112 bpm (below the bottom dashed red line). Data points referred to as published were computed from waveforms takenfrom Nichols et al. [1]. Detailed explanations about the ages for the data marked with a plus sign (þ) are given in the electronic supplementary material. (Onlineversion in colour.)
20 30 40 50 60 70 80 900
20
40
60
80
100
120
age (years)
w2
(bpm
)
++
+ ++
+
+++
+
healthy (published data)
linear fit (healthy published data)
VD (blind clinical data)
Figure 5. v2 at different ages under healthy and VD conditions. All VD patients show v2 , 36 bpm (below the horizontal dashed red line). A linear fit shows thatv2 decreases with age by approximately 10 bpm per decade. Data points referred to as published were computed from waveforms taken from Nichols et al. [1].Detailed explanations about the ages for the data marked with a plus sign (þ) are given in the electronic supplementary material. (Online version in colour.)
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
6
4. DiscussionIn this study, we formulated a modified version of the STFR
method allowing the direct extraction of the dominant instan-
taneous frequency ( _u1(t) ¼ du1=dt) of the first IMF from an
aortic pressure wave. In addition to directly outputting the
two IFs (v1 and v2), the main advantage of this method is
that only one arterial waveform, namely the pressure wave,
is required to perform the analysis in contrast to well-
known and widely used impedance and wave intensity
methods where both pressure and flow waves are required
[1,10]. Additionally, as only the shape of the waveform is
required to calculate the IFs, a wide range of both invasive
and non-invasive arterial pressure waveform measurement
techniques can be used.
4.1. Total frequency variation and optimum heart rateTo examine the physiological significance of the IFs, a
computational model was constructed to explore the relation-
ship between the IFs and the pulsatile workload on the heart.
Additionally, to isolate the pulsatile power contributions of
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
7
the aorta, parameters related to the left ventricle were keptconstant (see Material and methods and appendix C) [7].
Although this model is not entirely physiological, since
both HR and stroke volume increase in response to an
increase in required CO, it provides a framework with
which to explore the IF method and illustrates the importance
of aortic wave dynamics on the workload of the cardiovascu-
lar system. As shown in figure 2, based on the conditions of
the aorta, changing aortic stiffness by a factor of 3 from that of
a healthy 30 year old can increase the pulsatile workload on
the heart from 50 to 325 mW at a resting HR of 80 bpm. The
potential significance of this additional load on the heart is
clear when one considers that this is nearly a sevenfold
increase in pulsatile power at an HR of 80 bpm and that the
average hydraulic power of the heart is only about approxi-
mately 1 W [7,17]. Figure 2 also shows that regardless of
aortic stiffness the two IF curves intersect at an HR at
which the pulsatile power on the heart is minimized. In
other words at this optimum HR, the IFs of the heart–aorta
system before and after decoupling are equal (Dv ¼ 0).
These results reiterate those of our previous work, which
suggested that there is an optimum HR at which LV pulsatile
power is minimized, and this optimum HR shifts to a higher
value as the aortic rigidity increases [7]. Additionally, these
findings are generally in agreement with those observed pre-
viously by researchers examining ventriculoarterial matching
and optimal power output by the left ventricle [18–20].
4.2. The intrinsic frequencies as indicesfor cardiovascular disease
In the light of the relationship between the IFs and minimal
pulsatile workload on the heart, it follows that the total fre-
quency variation (TFV) should be very close to zero at young
ages when there is an optimum balance between heart pump-
ing dynamics and the dynamics of the aorta and its branches.
From the clinical data presented in figure 3, we observed that
TFV (Dv ¼ v1 2 v2) increases naturally with age as optimum
coupling is disrupted and that in cases of HF or VDs we see
TFV more rapidly deviate from the ageing line. This means
that subjects with very different ages can have the same TFV
given the severity of their CVD. In this regard, the results of
the clinical data suggest that the TFV can be considered as a
possible marker of left ventricle–arterial coupling as well as
being strongly correlated to CVD.
Taken individually, the IFs also contain information
related to the respective systems which are engaged during
the cardiac cycle. Namely, v1 reflects the dynamics of the
heart and v2 the dynamics of the aorta and arterial network.
For example, in figure 4 it was shown that v1 increases above
120 bpm in patients with HF with LV systolic dysfunction. By
contrast, in healthy individuals, v1 remains below 112 bpm.
Future work will aim at confirming the above observations
using more diverse clinical datasets at various stages of HF.
Likewise, changes in the dynamics of the aorta and arterial
network due to ageing or VD will be reflected in the value of
v2. As seen in figure 5, v2 decreases linearly with age. This be-
haviour is similar to the observations of other researchers
monitoring arterial stiffness through techniques such as
pulse wave velocity [8]. On the contrary, however, it is crucial
to point out that v2 is indirectly proportional to arterial rigid-
ity or in other words v2 decreases with increasing arterial
stiffness. In this regard, it is important to note that these IFs
should not be confused with the resonance frequency from
classical dynamical systems (e.g. mass–spring system), which
increases with rigidity. Additionally, as illustrated in figure 5
under VD conditions v2 prematurely drops below 36 bpm
independent of age. Although a more rigorous population
study is needed, these results suggest that v2 has potential as
a marker of vascular ageing as well as for diagnosis of VD
and the quantification of their severity (e.g. hypertension).
Clinical studies commonly challenge medical science to
interpret confounding or paradoxical results. A relevant
example is the observation noted in the Framingham Heart
Study that the risk of sudden death was increased by three-
fold in treated hypertensive subjects compared with
untreated [21]. The subjects were treated with a thiazide
diuretic and the explanation was relegated to a potential
electrolyte imbalance. In the light of our new method of eval-
uating the dynamic vascular physiology, a new explanation
may be forthcoming. When the aetiology of the hypertension
is a result of the vasculature alone, the effect of diuretic
therapy (reducing preload) may affect the v1 in an unfavour-
able way and shift the Dv to a level that is unsustainable and
cause sudden death. Clearly, more data are needed to evalu-
ate these possibilities, but the methodology presented here
may prove to be a very useful clinical tool.
4.3. Critique of methodsIn this study, we have proposed a new method for analysing
cardiovascular physiology using aortic pressure waveforms.
The clinical study data used for the analysis were collected
from previously published work or from retrospective
blinded patient datasets (see Material and methods). The
number of aortic pressure waveforms we could attain were
limited and not from a designed study. In this regard, a
focused clinical study would be required before extracting
any true statistical correlations. Our analysis however
shows a general trend that fulfils our intention of demonstrat-
ing a proof of concept. With regards to the CFD model, the
following assumptions have been made: (1) the blood was
assumed to be an incompressible Newtonian fluid; (2) the
aortic wall was assumed to be elastic and isotropic; (3) the
aortic arch and bifurcations were excluded; (4) the truncated
vasculature was modelled with an extension tube boundary
model [22]; and (5) the left ventricle was assumed to be a
flow source [2,23,24]. The effect of these modelling assump-
tions has been thoroughly explained in Pahlevan & Gharib
[7]. Nevertheless, the results from both clinical data and
CFD data are complementary.
4.4. ConclusionWe have shown the proof of concept for a new medical index,
the IF, and introduced a quantitative method based on instan-
taneous frequency theory. Using only one pressure waveform,
the IF concept can be used to quantify the impaired balance
between the heart and aorta under various disease conditions.
One important advantage of this method is that only the
shape of the pressure waveform, not the magnitude, is required
to extract the IFs. In this study, the IFs of the cardiovascular
system were extracted from clinical data under resting con-
ditions. From these data, we observed that the two IFs, v1 and
v2, representing the coupled heart þ aorta system and
decoupled aortic system, respectively, are close at young ages
and gradually deviate through the progression of age or disease.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
20
40
60
80
100
120
140
time (s)
inst
anta
neou
s fr
eque
ncy
(bpm
)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.850
60
70
80
90
100
110
120
130
140
150(a) (b)
time (s)
pres
sure
(m
m H
g)
Figure 6. Instantaneous frequency of the pressure wave’s first IMF. The range of instantaneous frequency oscillation (marked by the grey band) changes after the dicroticnotch (marked by the vertical red line). (a) The aortic input pressure. (b) The instantaneous frequency of the aortic input pressure wave. (Online version in colour.)
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
8
Additionally, we established a link between the closeness of
these frequencies and minimal pulsatile workload on the left
ventricle. Examined individually, the IFs contain information
relating to LV systolic dysfunction and VD. Further investi-
gations are needed to analyse the IF indices under non-resting
conditions such as exercise. Future studies are planned to
verify the predictive value of this concept in the detection of
CVD states. While this paper was focused on the heart–arterial
system, these principles may be extended to the full vascular
system including venous return or generalized to other systems:
for example, the gastrointestinal system, where there are natural
rhythm and waves amenable to a similar analysis.
Funding statement. The authors (N.M.P., D.G.R. and M.G.) acknowledgethe support from Caltech innovative initiative grant (CII).
Appendix A. Sparse time – frequencyrepresentationA.1. Sparse time – frequency representation algorithmThe adaptive STFR method consists of two major steps. The
first step is to construct a highly redundant dictionary of all
IMFs, D. The second step is to find the sparsest decomposition
by solving a nonlinear optimization problem
min: Msubject to: s(t) ¼
PMi¼1 ai(t) cos ui(t), ai(t) cos ui(t) [ D,
(i ¼ 1, . . . , M):
9>=>;
(A 1)
This problem is an L0 minimization problem. Solving this
problem is extremely difficult. It is a nonlinear and non-
convex optimization problem [12,14]. To overcome this diffi-
culty, a nonlinear matching pursuit method is proposed to
approximate the original L0 minimization problem. Based
on an approximation, the STFR method can be reduced to
an L2 minimization problem [14]. A brief description of this
algorithm is as follows:
min: f(t)� a(t) cos u(t)k k22, (A 2)
subject to: a(t) cos u(t) [ D: (A 3)
In this formulation, the dictionary D is defined as
D ¼ a(t) cos u(t):du
dt¼ _u(t) � 0, a(t), _u(t) [ V(u)
� �,
(A 4)
where V(u) is a linear space consisting of functions smoother
than cos u(t):
V(u) ¼ span 1, cosku
2Lu
� �, sin
ku2Lu
� �, k ¼ 1, . . . , Lu
� �:
(A 5)
More detail about the dictionary, D, can be found in Hou &
Shi [14].
At each step of the algorithm, an IMF is extracted. The
residual is treated as a new signal and the L2 minimization
is again applied to the residual. By this nonlinear matching
pursuit method, one can extract the different scales of a
multi-scale, non-stationary and nonlinear signal [14].
A.2. Instantaneous frequency of aortic pressure wavesA demonstration of the STFR method applied to an exemp-
lary aortic pulse pressure waveform and the corresponding
instantaneous frequency curve of the first IMF are shown in
figure 6. In both plots, the location of the dicrotic notch is
denoted by a vertical line. As shown in figure 6, on either
side of the dicrotic notch there is a distinct band of
frequencies around which the instantaneous frequency
oscillates, marked by the grey band.
Appendix B. Brute-force algorithmIn order to solve the modified STFR problem, a brute-force
algorithm was used. First, the domain D was taken as
D ¼ {(v1, v2) such that 0 , v1, v2 � C}: (B 1)
In domain D, the frequencies v1 and v2 are bounded above
by some constant C. This is a valid assumption since the
aortic pressure wave signal has a certain level of smoothness,
and the signal is not rough; therefore, certain frequencies
rsif.royalsocietypublishing.orgJ.R.Soc.
9
cannot be accepted physically and mathematically as the sol-ution of the problem.
Next, we discretize D for pairs of (v1 and v2). For each
point (v1 and v2) in the discretized domain, the modified
STFR problem is solved and the solution is stored as P(v1,
v2). Note that the minimum of the modified STFR problem
for the whole domain D corresponds to the minimum of
P(v1, v2) over (v1, v2); this is a simple search problem. The
corresponding minimum frequencies are denoted as (v1,
v2)m. The original minimization problem is not convex.
Thus, we may have several local minima. However, the
brute-force algorithm looks over all possible values of fre-
quencies and ensures that the corresponding minimizer
frequencies (v1, v2)m are in fact the unique global minimizer.
Interface11:20140617Appendix C. Computational fluid dynamics/fluid – solid interaction model of aortaC.1. Physical modelThe methods, the physical parameters of the model as well as
the relevance and accuracy of the model assumptions were
described previously [7]. The geometrical data such as length,
diameter and wall thickness were all within the average physio-
logical range [25]. The change in rigidity along the wall of the
aorta and tapering of the aorta were considered in the model;
however, the aortic arch and bifurcations were excluded. The
blood was assumed to be an incompressible Newtonian fluid.
The aortic wall was assumed to be elastic and isotropic. The
material properties of the wall were taken from Nichols et al. [1].
C.2. Mathematical and computational modelAn arbitrary Lagrangian–Eulerian (ALE) formulation was
applied to solve the FSI problem. In an ALE formulation,
the Navier–Stokes equations (for an incompressible fluid)
take the following form [22,26]:
r � V ¼ 0
rf
@V
@tþ (V �W)rV
� �þrp ¼ mfr2V þ Fb,
8<: (C 1)
where W is the mesh velocity, V is the flow velocity, p is the
static pressure, mf is the dynamic viscosity of the fluid and Fb
is the body force.
A no-slip boundary condition was assumed at the wall.
The coupling equations, applied to the solid–fluid interface,
were displacement compatibility and traction equilibrium at
the wall.
Large deformation–small strain theory was considered
for the solid domain (wall of the aortic model). The solid
mechanics equations, constitutive relation (equation (C 2))
and balance of momentum (equation (C 3)), for a linear elastic
isotropic material in Lagrangian form, were used to calculate
the dynamic motion of the elastic wall [27],
sij ¼ l1kkdij þ 2ml1ij (C 2)
and
sij,j þ Fi ¼ rs€ui: (C 3)
In these equations, sij is the wall stress tensor, F is the external
force, u is the displacement vector, rs is the wall density and
l, ml are Lame constants.
The finite-element method with the direct two-way coup-
ling method of FSI was used. The time integration scheme
was the implicit Euler method. The commercial package
ADINA, v. 8.6 (ADINA R&D, Inc., MA, USA) was used to
run the simulations. Full details of the formulation of the
FSI model and numerical method can be found in our
recent publication [7].
C.3. Inflow and outflow boundary conditionA physiological flow wave with a flat velocity profile, the same
as Pahlevan & Gharib [7], was imposed at the inlet. It was
scaled to give a CO of 4.6 l min21 for any desired HR. The
choice of an outflow boundary condition is important since
aortic waves can be greatly affected by changes in the radial
arteries. These arteries can affect wave dynamics in the aorta
by altering the wave arrival time at the inlet of the aorta as
well as by changing the terminal volume compliance and
resistance. We used the extension tube boundary model for
the outflow boundary condition [22]. This outflow boundary
model involves extending the computational domain by an
elastic tube connected to a rigid contraction tube. This outflow
boundary model takes into account the effects of the truncated
vasculature (resistance, compliance and wave reflection). The
geometrical and material properties of the outflow boundary
model are the same as in Pahlevan & Gharib [7].
References
1. Nichols W, O’Rourke M, Vlachopoulos C. 2011McDonald’s blood flow in arteries: theoretical,experimental and clinical principles, 6th edn. BocaRaton, FL: CRC Press.
2. Denardo SJ, Nandyala R, Freeman GL, Pierce GL,Nichols WW. 2010 Pulse wave analysis of the aorticpressure waveform in severe left ventricular systolicdysfunction. Circ. Heart Fail. 3, 149 – 156. (doi:10.1161/circheartfailure.109.862383)
3. Garcia D, Barenbrug PJC, Pibarot P, Dekker ALAJ,van der Veen FH, Maessen JG, Dumesnil JG, DurandL-G. 2005 A ventricular-vascular coupling model inpresence of aortic stenosis. Am. J. Physiol. Heart Circ.
Physiol. 288, H1874 – H1884. (doi:10.1152/ajpheart.00754.2004)
4. Ooi H, Chung W, Biolo A. 2008 Arterial stiffness andvascular load in heart failure. Congest. Heart Fail. 14,31 – 36. (doi:10.1111/j.1751-7133.2008.07210.x)
5. Wang Y-YL, Sze W-K, Bau J-G, Wang S-H, Jan M-Y, HsuT-L, Wang W-K. 2008 The ventricular-arterial couplingsystem can be analyzed by the eigenwave modes of thewhole arterial system. Appl. Phys. Lett. 92, 153 901 –153 903. (doi:10.1063/1.2911746)
6. Manisty C, Mayet J, Tapp RJ, Parker KH, Sever P,Poulter NH, Thom SAM, Hughes AD. 2010 Wavereflection predicts cardiovascular events in
hypertensive individuals independent of bloodpressure and other cardiovascular risk factors: anASCOT (Anglo-Scandinavian Cardiac Outcome Trial)substudy. J. Am. Coll. Cardiol. 56, 24 – 30. (doi:10.1016/j.jacc.2010.03.030)
7. Pahlevan NM, Gharib M. 2011 Aortic wave dynamicsand its influence on left ventricular workload.PLoS ONE 6, e23106. (doi:10.1371/journal.pone.0023106)
8. Safar M, O’Rourke MF. 2006 Arterial stiffness inhypertension. Amsterdam, The Netherlands: Elsevier.
9. Avolio AP, Butlin M, Walsh A. 2010 Arterial bloodpressure measurement and pulse wave analysis—their
rsif.royalsocietypublishing.orgJ.R.Soc.Interface
11:20140617
10
role in enhancing cardiovascular assessment. Physiol.Meas. 31, R1. (doi:10.1088/0967-3334/31/1/R01)10. Parker KH, Jones CJH. 1990 Forward and backwardrunning waves in the arteries: analysis using themethod of characteristics. J. Biomech. Eng. 112,322 – 326. (doi:10.1115/1.2891191)
11. Hughes A, Parker K. 2009 Forward and backwardwaves in the arterial system: impedance or waveintensity analysis? Med. Biol. Eng. Comput. 47,207 – 210. (doi:10.1007/s11517-009-0444-1)
12. Hou TY, Shi Z. 2011 Adaptive data analysis viasparse time-frequency representation. Adv. Adapt.Data Anal. 3, 1 – 28. (doi:10.1142/S1793536911000647)
13. Huang NE, Shen Z, Long SR, Wu MC, Shih HH,Zheng Q, Yen N-C, Tung CC, Liu HH. 1998 Theempirical mode decomposition and the Hilbertspectrum for nonlinear and non-stationary timeseries analysis. Proc. R. Soc. Lond. A 454, 903 – 995.(doi:10.1098/rspa.1998.0193)
14. Hou TY, Shi Z. 2012 Data-driven time-frequencyanalysis. Appl. Comput. Harmon. Anal. 35,284 – 308. (doi:10.1016/j.acha.2012.10.001)
15. Huang W, Shen Z, Huang NE, Fung YC. 1998Engineering analysis of biological variables: anexample of blood pressure over 1 day. Proc. Natl
Acad. Sci. USA 95, 4816 – 4821. (doi:10.1073/pnas.95.9.4816)
16. Huang W, Shen Z, Huang NE, Fung YC. 1998 Use ofintrinsic modes in biology: examples of indicialresponse of pulmonary blood pressure to+ stephypoxia. Proc. Natl Acad. Sci. USA 95, 12 766 – 12771. (doi:10.1073/pnas.95.22.12766).
17. Kelly R, Fitchett D. 1992 Noninvasive determination ofaortic input impedance and external left ventricularpower output: a validation and repeatability study of anew technique. J. Am. Coll. Cardiol. 20, 952 – 963.(doi:10.1016/0735-1097(92)90198-V)
18. Sunagawa K, Maughan WL, Sagawa K. 1985Optimal arterial resistance for the maximal strokework studied in isolated canine left ventricle. Circ.Res. 56, 586 – 595. (doi:10.1161/01.res.56.4.586)
19. O’Rourke MF. 1967 Steady and pulsatile energy losses inthe systemic circulation under normal conditions and insimulated arterial disease. Cardiovasc. Res. 1, 313 – 326.(doi:10.1093/cvr/1.4.313)
20. Kass D. 2002 Age-related changes in venticular –arterial coupling: pathophysiologic implications.Heart Fail. Rev. 7, 51 – 62. (doi:10.1023/a:1013749806227)
21. Kannel WB, Cupples LA, D’Agostino RB, Stokes J. 1988Hypertension, antihypertensive treatment, and sudden
coronary death. The Framingham Study. Hypertension11, II45. (doi:10.1161/01.HYP.11.3_Pt_2.II45)
22. Pahlevan N, Amlani F, Hossein Gorji M, Hussain F,Gharib M. 2011 A physiologically relevant, simpleoutflow boundary model for truncated vasculature.Ann. Biomed. Eng. 39, 1470 – 1481. (doi:10.1007/s10439-011-0246-0)
23. Pahlevan NM, Gharib M. 2014 A bio-inspiredapproach for the reduction of left ventricularworkload. PLoS ONE 9, e87122. (doi:10.1371/journal.pone.0087122)
24. Westerhof N, O’Rourke MF. 1995 Haemodynamicbasis for the development of left ventricular failurein systolic hypertension and for its logical therapy.J. Hypertens. 13, 943 – 952. (doi:10.1097/00004872-199509000-00002)
25. Caro CG, Pedley JG, Schroter RC, Seed WA. 1978 Themechanics of the circulation. Oxford, UK: OxfordUniversity Press.
26. Bathe KJ, Zhang H, Wang MH. 1995 Finite elementanalysis of incompressible and compressible fluidflows with free surfaces and structural interactions.Comput. Struct. 56, 193 – 213. (doi:10.1016/0045-7949(95)00015-9)
27. Bathe KJ. 2006 Finite element procedures.Englewood Cliffs, NJ: Prentice-Hall.