Sei sulla pagina 1di 10

Biomedical Signal Processing and Control 6 (2011) 251–260

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control


journal homepage: www.elsevier.com/locate/bspc

Mathematical model-based markers of autonomic nervous activity during the


Valsalva Maneuver and comparison to heart rate variability
Michel Kana a,∗ , Jiri Holcik b,c
a
Department of Biomedical Informatics, Czech Technical University in Prague, Nam. Sitna 3105, 27201 Kladno, Czech Republic
b
Institute of Biostatistics and Analyses, Masaryk University Brno, Kamenice 3, 62500 Brno, Czech Republic
c
Institute of Measurement Science, Slovak Academy of Sciences, Dubravska cesta 9, 841 04 Bratislava, Slovakia

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

Article history: This work exhibits limitations of some time–frequency markers of heart rate variability (SDNN: standard
Received 29 March 2010 deviation of normal-to-normal intervals, RMSSD: root mean square successive difference of normal-to-
Received in revised form 22 March 2011 normal intervals, LF: power of low frequency components 0.015–0.15 Hz, HF: power of high frequency
Accepted 5 May 2011
components 0.15–0.4 Hz) as tool for assessing autonomic nervous activity on the cardiovascular system
Available online 2 June 2011
during the Valsalva Maneuver (VM) and proposes a new approach based on a physiologically based
mathematical model.
Keywords:
Our mathematical model fits measured heart rate and galvanic skin response during intra-thoracic
Markers of autonomic nervous activity
Parasympathetic tone
and intra-abdominal pressure elevation, simulates the baroreflex and estimates parameters that should
Sympathetic tone quantify mean autonomic tone of parasympathetic activity (MATpar ), maximal autonomic tone of
Valsalva Maneuver parasympathetic activity (MATpar , total autonomic tone of sympathetic activity (TATsym ) and autonomic
Heart rate variability tone balance (ATB) on 5 healthy subjects.
Mathematical modeling An extended statistical analysis including multivariable linear regression analysis and correlation anal-
ysis was performed in order to study patterns of variation of our model-based estimates and possible
relationships with time–frequency markers of heart rate variability during the tachycardia phase 2 and
bradycardia phase 4 of the VM. It was found that, MATpar significantly decreased by 32 ± 8% (p < 0.001)
from 0.87 ± 0.02 n.u. (normalized unit) in phase 1 to 0.58 ± 0.05 n.u. in phase 2 and was found to be signif-
icantly correlated to HF (R = 0.87, p < 0.05) that decreased from 0.5 ± 0.21 n.u. to 0.1 ± 0.01 n.u. Meanwhile
TATsym increased by 2135 ± 711% (p < 0.001) from 0.14 ± 0.05 n.u. to 2.89 ± 0.7 n.u. and was significantly
correlated to LF (R = 0.95, p < 0.05) that increased from 0.49 ± 0.21 n.u. to 0.89 ± 0.01. MATpar significantly
increased by 87 ± 63% (p < 0.001) from 0.53 ± 0.18 n.u. in phase 3 to 0.89 ± 0.02 n.u. in phase 4 and was
found to be significantly correlated to RMSSD (R = 0.91, p < 0.05) that increased from 18.61 ± 5.23 ms
to 36.31 ± 7.02 ms. Meanwhile TATsym significantly decreased by 82 ± 13% (p < 0.01) from 1.46 ± 1.4 n.u.
to 0.14 ± 0.05 n.u. and was significantly correlated to SDNN (R = 0.91, p < 0.05) that decreased from
28.79 ± 8.17 ms to 22.62 ± 9.63 ms.
An additional Wilcoxon sign test suggested the following markers that might provide the most signif-
icant assessment of autonomic nervous activity in a given phase of VM: SDNN in phases 0 & 5, TATsym in
phases 1 & 2 and ATB in phases 3 & 4 for sympathetic activity assessment; RMSSD in phases 0 & 5, MATpar
in phases 1 & 2, MATpar in phases 3 & 4 for parasympathetic activity assessment.
Results suggest that statistical quantities derived from mathematical models of cardiovascular control
provide useful information for assessing the distinct contribution of sympathetic and parasympathetic
branches on heart rate variability during the Valsalva Maneuver.
© 2011 Elsevier Ltd. All rights reserved.

1. Introduction (1677–1723), which consists of an abrupt transient increase of


intra-thoracic and intra-abdominal pressure provoked by blowing
The Valsalva Maneuver (VM) is a cardiac autonomic test, into a resistance of approx. 40 mmHg for about 15 s. The VM is suit-
named after the physician and anatomist Antonio Maria Valsalva able for assessing the integrity of autonomic nervous control on the
cardiovascular system [1]. The VM consists of several phases which
are graphically depicted in Fig. 1. During phase 1 a brief increase
in arterial blood pressure and decrease in heart rate is observed. In
∗ Corresponding author. Tel.: +420 312 608 207; fax: +420 312 608 204.
phase 2 sustained high intra-thoracic and intra-abdominal pressure
E-mail address: kana@fbmi.cvut.cz (M. Kana).

1746-8094/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.bspc.2011.05.001
252 M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260

2. Materials

Three healthy women (age 15, 21 and 46 years) and 2 men


(both age 23 years) underwent the Valsalva Maneuver. Each sub-
ject performed the maneuver several times until a typical heart rate
response was visible, which characterizes the VM. We selected 10
datasets, corresponding to 2 measurements per subject.
A Biopac MP35 [11] system was used for data acquisition. Three
leads electrocardiogram (ECG) was recorded using SS2L electrodes
placed on the left and right wrists and on the left part of the
Fig. 1. Mean arterial pressure (plain line) and heart rate (dashed line) changes dur-
ing the VM. The maneuver covers the timeline from the 15th to the 30th second. abdomen just above the belt. Two electrodes SS57L were fixed
Phase 1 goes from 15th to 16th second, phase 2 ends at 30th second, followed by around the index (ground) and middle finger (+) for galvanic skin
phase 3. The last phase 4 runs from 31st to 38th second. response (GSR) measurement. Baseline measurements were per-
formed during 15 s prior to the maneuver. After taking a breath the
subject expired against a resistance during 15 s while signals were
hinders venous return, leading to a rapid decrease in arterial blood recorded. After the forced expiration was released, measurements
pressure and increased heart rate. During phase 3, the forced expi- were resumed for another 30 s. Thus we obtained 60 s ECG and
ration is released resulting in a sudden decrease in arterial blood GSR records for each subject. The signals were sampled at 500 Hz
pressure. During phase 4, the accumulated venous blood returns using the Biopac Student Lab PRO acquisition software. They were
to heart and is pumped into the constricted arteries causing an exported in time series ASCII files for further processing in Matlab.
“overshoot” of arterial pressure. R-peaks were detected from the ECG signal using an implemen-
There are signal analysis methods to evaluate cardiovascular tation of the non-syntactic QRS detection algorithm [12]. For this
response during the Valsalva Maneuver including T-wave width purpose the ECG signal was sampled down to 400 Hz and filtered
alteration [2], QT-RR regression lines [3] and baroreflex open- to get rid of shifting baseline and 60 Hz noise using a Butterworth
loop gain [4]. Although such methods provide insight information, band pass filter. RR intervals durations were calculated as dif-
the contribution of neural pathways involved is not covered. An ferences of timestamps of consecutive R-peaks. The obtained RR
attempt using frequency domain analysis of heart rate variability time series was de-trended, sampled down to 10 Hz and stored
(HRV) with Fast Fourier Transform (FFT) was made in [5,6]. FFT in a vector RRmeas for further analysis. The activity of the sweat
analysis of autonomic balance during the VM can held erroneous glands in response to sympathetic nervous stimulation results in
results because of the very short duration of the test (15–30 s). an increase in the level of skin conductance. The measured gal-
Recording of approximately 1 min is needed to access the high vanic skin response signal was sampled down to 10 Hz and stored
frequency (HF) components of HRV; while 2 min are needed to in a vector GSRmeas for further analysis.
evaluate the low frequency (LF) components. Very low frequency
(VLF) assessment from short-term recordings (less than 5 min) is a
dubious measurement and should be avoided [7]. Discrete Wavelet 3. Limitations of heart rate variability
Transform (DWT) has been suggested to provide a more accurate
frequency analysis of very short signals. DWT has been applied in Heart rate variability presents one of the most promising mark-
[8] to decompose the non stationary heart rate signals obtained ers of quantitative assessment of autonomic activity. Variables can
during VM over time, yielding both frequency and time resolu- be measured in frequency domain of NN intervals (normal-to-
tion. Disadvantages of DWT method include its limitation to clearly normal interval between adjacent QRS complexes in the ECG). The
demonstrate the distinct contribution of LF (0.067–0.125 Hz) and spectral values for heart rate variability can be calculated for low
HF (0.125–0.25 Hz) components between phase 2 and phase 3 of (LF: 0.04–0.15 Hz) and high (HF: 0.15–0.5 Hz) frequency bands. Low
VM. Furthermore the reported increase to 30% in LF band dur- frequency heart rate oscillations are considered to be mediated by
ing phase 4, is contradictory with the observed bradycardia in combined sympathetic and parasympathetic activity while there
that phase. It appears that both FFT and DWT methods are not is a predominance of sympathetic tone during stressful conditions.
able to delineate the contributions of parasympathetic and sympa- High frequency heart rate oscillations are associated with respi-
thetic activities in the LF band. Physiologically based mathematical ratory sinus arrhythmia and reflect parasympathetic activity. The
models look as a promising tool for distinctly evaluating sympa- physiological background of very low frequency (VLF) components
thetic and parasympathetic activity during each single phase of is still under discussion. The spectrum of the HRV signal is usu-
the VM. Authors of [9,10] proposed such integrated cardiovascu- ally calculated using non-parametric methods based on Fast Fourier
lar and respiratory models aiming to quantify the response to VM. Transforms or other methods such as Lomb periodogram, which is
Both models include multiplicity of mechanisms for simulating a method of spectral analysis for unevenly sampled series [13]. A
hemodynamic response; atrial, ventricular and lung mechanics. rigorous evaluation requires at least 60 s ECG. This basic condition
Because of the lack of direct experimental evidence, the authors is not met by the Valsalva Maneuver, which is usually performed
acknowledge that their equations for autonomic nervous con- for less than 30 s. Therefore time-domain analysis is more suitable
trol are largely empirical in nature and not readily applicable to in the evaluation of HRV during VM, for example the standard devi-
other studies without modification. It is not clear how to interpret ation of NN intervals (SDNN) which reflects the overall variability,
the simulated firing rates of efferent neurons, since a quantita- but strongly depends on the duration of the signal. SDNN increases
tive comparison with common methods such as HRV was not with the length of recording. The square root of the mean square
included. successive differences of NN intervals (RMSSD) estimates high fre-
Our objective is to develop a simpler mathematical model of quency variations of heart rate and is proposed as a measure of
cardiovascular control during VM and to derive model-based esti- parasympathetic activity.
mates that aim to quantify sympathetic and parasympathetic tone Time-domain and frequency-domain estimates of HRV were
on the cardiovascular system. A comparison with HRV should estimated for the 10 data sets and values are presented in Fig. 2 (top
bring evidences for validity and eventually superiority of this panels). The SDNN and RMSSD for each single phase were calculated
method. on the portion of NN intervals of that phase. Changes in SDNN are
M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260 253

Fig. 2. Time-domain parameters (SDNN – top left, RMSSD – top right) and normalized frequency-domain parameters (LF – bottom left, HF – bottom right) of HRV during
each phase of the VM for 10 data sets and their mean values (dashed line).

visible from one phase to the next phase of the maneuver. Greatest 4. Model development
heart rate variability occurs in the phase 2, where average SDNN
equals 38.66 ± 3.67 ms compare to 21.27 ± 5.78 ms before the 4.1. Circulatory component
maneuver. Phase 2 is also characterized by a decreased RMSSD from
30.14 ± 8.11 ms before the maneuver to 23.81 ± 6.66 ms, which A block diagram of our model is depicted in Fig. 3. The circulatory
suggests a withdrawal of vagal activity and increased sympathetic component includes four compartments following the approach
activity as response to the transient decrease in blood pressure. suggested in [15]. Large arteries (e.g. aorta) are lumped together
Phase 4 bradycardia is characterized by an increase of RMSSD from in an arterial systemic compartment, characterized by the pres-
18.61 ± 5.22 ms in phase 3 to 36.31 ± 7.02 ms in phase 4. sure Pas in mmHg. The aorta branches into smaller arteries and a
Frequency-domain estimates of HRV were calculated for each network of capillaries which form the peripheral systemic com-
phase of the VM using the Lomb periodogram. The power of low partment, characterized by the resistance PR in mmHg min/l. After
frequency components (LF band 0.015–0.15 Hz) and high frequency leaving the capillaries blood flows into veins, then into both vena
components (HF band 0.15–0.4 Hz) are depicted in normalized cavae, which join together to form the venous systemic compart-
units in Fig. 2 (bottom panels). They represent the relative value ment, characterized by the pressure Pvs in mmHg. From there blood
of each power component centered at the frequency of interest enters the atria, lungs and ventricles, which we lump together into
(LF or HF) and divided by total power. At baseline, on average a heart compartment, characterized by the stroke volume Vstr in
22.54 ± 11.02% of the energy corresponds to LF components. Their liter and the heart rate HR in beats/min.
activity increases by 49.46 ± 21.74% and 89.03 ± 1.2% of the total The volume of blood ejected by the heart per beat (stroke vol-
energy of phases 1 and 2, respectively, suggesting an increase ume in liter) depends on the pressure in both venous and arterial
in sympathetic activity. A similar increase was obtained in [8] systemic compartments. We capture the effect of contractility and
using wavelets transform. Phase 4 is characterized by predominant compliance of the ventricles using a weighting factor Kstr in liter:
energy in HF components, 61.52 ± 6.4% what supports the observed pvs
bradycardia. Vstr = kstr · (1)
pas
Although LF, HF, RMSSD and SDNN provide a measure of the
degree of autonomic modulation, they do not indicate the level of The blood flow generated by the heart, the cardiac output in
autonomic tone regulating heart rate for each single phase of the l/min, is given by:
VM [14]. They can also provide contradictory results, for exam-
Qin = HR · Vstr (2)
ple in Fig. 2, LF increases in phase 1 while a decreased SDNN is
observed. RMSSD surprisingly increases from 17.24 ± 10.18 ms in The blood flow in l/min through the peripheral systemic com-
phase 1 to 23.81 ± 6.66 ms in phase 2 although phase 2 should be partment can be modeled as:
characterized by a decrease of parasympathetic activity. Phase 3 is
characterized by an increase of HF by 80% while RMSSD is reporting 1
Qout = · (Pas − Pvs ) (3)
a decrease by 20%. Such inconsistency motivated us to investigate PR
the usage of estimates of autonomic tone based on mathematical If we assume that the arterial systemic compartment is charac-
models. terized by a compliance Cas in l/mmHg, then we can partly use the
254 M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260

Fig. 3. Block diagram describing the model of heart rate, blood pressure and peripheral vascular resistance control during the Valsalva Maneuver. Intra-thoracic pressure
change is communicated to venous systemic and peripheral systemic compartments via pressure variables Ps , P1 and P3 . Change in venous pressure Pvs affects stroke volume
Vstr , which has a direct influence on arterial pressure Pas . Baroreceptors firing rate N signalizes pressure fluctuations to autonomic nervous centers for cardiovascular control.
Response is communicated via efferent sympathetic and parasympathetic outflows Tsym and Tpar ; this has the effect of modulating heart rate HR, cardiac output Qin , blood
flow through the peripheral systemic compartment Qout and peripheral vascular resistance PR.

Grodins model of the mechanical part of the cardiovascular system 4.3. Intra-thoracic pressure component
[16] and write the differential equation:
During the phase 1, the gradual increase in intra-thoracic and
dPas 1 intra-abdominal pressure will cause aortic and peripheral vessels
= · (Qin − Qout ) (4)
dt Cas compression. This results in slight increased peripheral resistance
that is modeled using a Gaussian function P1 :
4.2. Autonomic nervous control component (t−dPtime −(P1dur /8))2
2(P1dur /8)2
P1 = P1amp · e (10)
The change in arterial blood pressure is sensed by the carotid and
aortal baroreceptors, which adjust their firing rate accordingly. We where P1amp is the amplitude of pressure increase in mmHg,
adapt the non-linear mathematical model for baroreflex regulation dPtime is the timestamp in seconds when the VM starts with con-
proposed in [17] as follows, stant value 15 s, P1dur is the duration of the phase 1.
Therefore Eq. (9) can be rewritten as:
dPas M−N
= (Pas − Pasmean ) · N · − N0 (5)
dt (M/2)2 PR = PR0 · (1 + Kpr · Tsym ) + P1 (11)

where N is the firing rate of the baroreceptors in Hz corresponding The sustained high intra-thoracic and intra-abdominal pressure
to the deviation of the arterial pressure from its mean value Pasmean ; is modeled using a logistic function Ps :
M is the maximum firing rate with constant value 120 Hz. N0 is the

⎨ Psmax · 10−1 · edPrate ·t
baseline firing rate with constant value 100 Hz. if 0 ≤ t ≤ dPtime + dPdur
Ps = −1
· (edPrate ·t + 1) (12)
The baroreceptors firing rate information is integrated in ⎩ 0Psmax + 10 if t > dPtime + dPdur
the cardiovascular autonomic nervous centers, which generate a
parasympathetic and sympathetic response. The level of parasym-
where Psmax is the maximum intra-thoracic pressure with
pathetic tone is given by Tpar . It is assumed to follow the direct
constant value 40 mmHg, dPrate is the growing rate of pressure
law, thus it is linearly dependent on the baroreceptors firing rate
development at the beginning of the expiration and dPdur is the
N. The sympathetic tone is modeled by Tsym . The inhibitory effect
duration of the forced expiration with constant value 15 s.
of parasympathetic tone on the sympathetic response is taken into
We assume that this intra-thoracic pressure has a linear effect
account using a dampening factor ˇ. The following equations were
on the pressure Pvs in the venous systemic compartment, thus we
proposed by Ottesen et al. [17]:
can write the following:
N + N0 Pvs = Pvs0 + kvs · Ps (13)
Tpar = (6)
M
where Pvs0 is the mean venous pressure in mmHg prior to VM
1 − ((N + N0 )/M) and kvs is a weighting factor.
Tsym = (7)
1 + ˇ · Tpar During the phase 2, the sustained intra-thoracic and intra-
abdominal pressure hinders venous return. We can rewrite (1) and
The heart rate HR deviation from its mean value HR0 depends
include the resulting decrease of transmural pressure in the venous
on the sympathetic and parasympathetic tone on sinoatrial node.
systemic compartment, as follow:
We model this response using scaling factors Ms and Mp :
Pvs0 − kvs · Ps
HR = HR0 · (1 + Ms · Tsym − Mp · Tpar ) (8) Vstr = kstr · (14)
Pas

The sympathetic nervous stimulation additionally results in During the phase 3, the forced expiration is released, resulting
increased peripheral vascular resistance PR. We assume local con- in a decrease in venous pressure. We model this marked decrease
trol and hormone control of arteriolar resistance to be negligible using a negative Gaussian function:
and model the modulation of peripheral resistance by sympathetic (t−dPtime −dPdur −(P3dur /4))2
outflow Tsym using the scaling factor Kpr with PR0 being the baseline z·(P3dur /8)2
P3 = −P3amp · e (15)
peripheral resistance:
where P3amp is the amplitude of pressure drop in mmHg, P3dur
PR = PR0 · (1 + Kpr · Tsym ) (9) is the duration of the phase 3 in seconds.
M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260 255

Therefore, Eq. (14) can be rewritten as sympathetic tone over total parasympathetic tone for a given phase
using the following formulas.
Pvs0 − kvs · Ps + P3
Vstr = kstr · (16)
Pas TATsym (i)
ATB(i) = (27)
During phase 4, the accumulated venous blood returns to heart TATpar (i)
and is pumped into the constricted arteries causing an overshoot
of stroke volume above the normal level. We can rewrite (16) and 4.5. Sensitivity analysis
include the overshoot effect, as follows:
⎧ We study how sensitive the model outputs are to small per-
⎪ Pvs0 − kvs · Ps + P3

⎪ Vstr = kstr · turbation of parameters values. The sensitivity of model output

⎪ Pas


HR is obtained by partial differentiation of Eq. (8) with respect to
⎪ if 0 ≤ t < dPtime + dPdur
⎪ model parameters. For example this will held following sensitivity
⎨ Pvs0 − kvs · Ps + P3 + P4amp
Vstr = kstr · functions for parameters HR0 , Ms and Mp :
P (17)


as
if dPtime + dPdur ≤ t < dPtime + dPdur + P3dur + P4dur

⎪ ∂

⎪ Pvs0 − kvs · Ps + P3 + (P4amp /5) S1 = HR = 1 + Ms · Tsym − Mp · Tpar (28)

⎪ Vstr = kstr · ∂HR0

⎩ Pas
if t ≥ dPtime + dPdur + P3dur + P4dur ∂
S2 = HR = HR0 · Tsym (29)
∂Ms
where P4amp is a weighting factor to quantify the overshoot effect
on the stroke volume, P4amp is the duration of phase 4. ∂
S3 = HR = −HR0 · Tpar (30)
∂Mp
4.4. Model-based estimates of autonomic tone
The sensitivity of model output PR is obtained by partial differ-
entiation of Eq. (11) with respect to model parameters. For example
As an alternative to heart rate variability we extend the mathe-
this will held following sensitivity functions for parameters PR0 and
matical model previously developed with markers of autonomic
Kpr :
activity during the Valsalva Maneuver. For this purpose we use
model equations (6) and (7) to calculate the trajectory of parasym- ∂
pathetic tone Tpar and sympathetic tone Tsym on a normalized scale S4 = PR = 1 + Kpr · Tsym (31)
∂PR0
of 0–1.
The first class of model-based estimate is the so-called mean ∂
S5 = PR = PR0 · Tsym (32)
autonomic tone (MAT) which is calculated as the average ∂Kpr
parasympathetic or sympathetic tone for a given phase of the Val-
Sensitivity Eqs. (28)–(32) were resolved together with models
salva Maneuver using the formulas below. The index i represents
equations in Matlab. A sensitivity ranking of the parameters using
the phase of the maneuver which can be baseline (i.e. before the
the maximum absolute value of the sensitivity function as criterion
maneuver), phase 1, phase 2, phase 3, phase 4 and post VM (i.e. after
shows that sensitive parameters can be ranked from the most to the
the maneuver). Tpar {i} is the portion of the parasympathetic signal
less sensitive as follows: Mp , Ms , HR0 , Kpr , PR0 . The remaining 13
corresponding to the phase i. Tsym is the portion of the sympathetic
model parameters are more or less sensitive and have been ranked
signal corresponding to the phase i. MATpar and MATsym are the
closed to the middle of the ranking scale. This suggests performing
average parasympathetic tone and average sympathetic tone in the
mean arterial blood pressure measurement in order to be able to
phase i, respectively.
identify those parameters more efficiently.
MATpar (i) = mean(Tpar {i}) (18)
5. Results
MATsym (i) = mean(Tsym {i}) (19)

A second class of model-based estimate is the so-called total 5.1. Parameters estimation
autonomic tone (TAT) which is calculated as the sum of all indi-
vidual tones within the portion of parasympathetic or sympathetic Measured heart rate was derived in beats/min from measured
signal for a given phase using the following formulas. RR time series (RRmeas ) using the following formula:

TATpar (i) = sum(Tpar {i}) (20) 60 000


HRmeas = (33)
RRmeas
TATsym (i) = sum(Tsym {i}) (21)
Measured galvanic skin response signal (GSRmeas ) was used to
TAT (i) = TATpar (i) + TATsym (i) (22) approximate measured peripheral vascular resistance as follows:
The third and fourth class of model-based estimates are the so- 1
called minimal autonomic tone (MIAT) and maximal autonomic PRmeas = (34)
GSRmeas
tone (MAAT) which are calculated respectively as the minimum
and maximum tone within the portion of parasympathetic or sym- The model equations (2)–(8), (10)–(15) and (17) were imple-
pathetic signal for a given phase using the following formulas. mented using the Systems Biology Toolbox 2 for Matlab [18]. The
toolbox calls Matlab ode23s solver for resolving the ordinary differ-
MIATpar (i) = min(Tpar {i}) (23) ential Eqs. (4) and (5) with initial conditions Pas (0) = 80 mmHg and
MIATsym (i) = min(Tsym {i}) (24) N(0) = 0 Hz. ode23s is a one-step differential equation solver based
on a modified Rosenbrock formula of order 2 [19].
MAATpar (i) = max(Tpar {i}) (25) In order to estimate the parameter values, a weighted least-
squares formulation minimizing the errors between measured data
MAATsym (i) = max(Tsym {i}) (26)
(HRmeas , PRmeas ) and model outputs (HR, PR) was applied. The corre-
The last class of model-based estimate is the so-called auto- sponding cost function is given in Eq. (35) below. n is the number of
nomic tone balance (ATB) which is calculated as the ratio of total data points in the time series (value is 60 s times 10 Hz, i.e. 600 data
256
Table 1
Estimated values of model parameters.

Mean ± STD Min Max Data set 1 Data set 2 Data set 3 Data set 4 Data set 5 Data set 6 Data set 7 Data set 8 Data set 9 Data set 10

Pasmean 85.37 ± 11.05 mmHg 70 120 102.50 70.50 82.59 76.62 105.42 77.76 88.70 80.17 85.17 84.32
ˇ 1.5 ± 0.42 1 2 1.66 1.98 1.95 1.05 1.31 1.13 1.04 1.85 1.95 1.03
HR0 82.49 ± 7.61 beats/min 60 100 86.29 86.75 85.78 84.35 61.20 82.96 82.64 83.79 85.79 85.33
Ms 1.74 ± 1.15 0.1 5 1.04 1.06 1.30 1.16 4.86 1.66 1.94 1.80 1.70 0.93
Mp 0.005 ± 0.002 0.001 5 0.006 0.001 0.007 0.004 0.004 0.003 0.006 0.004 0.004 0.006
Kpr 1.5 ± 0.59 1 5 1.03 1.01 1.19 1.04 2.72 1.46 2.26 1.69 1.59 1.01
PR0 12.97 ± 1.12 ␮Mho-1 5 20 13.63 13.79 13.62 13.46 10.03 13.29 12.83 13.34 12.34 13.35
Cas 0.16 ± 0.03 0.1 1 0.19 0.16 0.17 0.18 0.10 0.22 0.11 0.18 0.14 0.17
Kvs 0.28 ± 0.07 0.1 1 0.36 0.26 0.33 0.29 0.12 0.30 0.23 0.26 0.36 0.32
Pvs0 11.1 ± 2.14 mmHg 5 20 15.21 8.61 13.04 11.21 9.57 12.17 9.43 9.58 9.38 12.76

M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260


Kstr 0.56 ± 0.12 0.1 1 0.58 0.62 0.46 0.47 0.77 0.49 0.73 0.47 0.57 0.44
P1dur 2.37 ± 0.29 s 2 5 2.30 2.91 2.55 2.22 2.82 2.24 2.21 2.15 2.25 2.09
P1amp 1.14 ± 0.19 ␮Mho-1 1 5 1.05 1.06 1.61 1.36 1.02 1.15 1.08 1.02 1.04 1.03
P3dur 4.05 ± 1.04 s 2 5 2.40 4.52 4.99 4.43 2.92 4.79 2.41 4.90 4.50 4.63
P3amp 2.26 ± 0.3 ␮Mho-1 2 5 2.05 2.01 2.03 2.47 3.00 2.26 2.21 2.06 2.16 2.32
P4dur 3.78 ± 0.92 s 2 5 3.08 4.59 4.73 2.52 2.65 4.67 3.21 4.76 4.36 3.22
P4amp 3.04 ± 0.06 ␮Mho-1 3 5 3.03 3.00 3.02 3.00 3.01 3.05 3.05 3.02 3.22 3.05
dPrate 2.22 ± 0.23 0.5 5 2.29 2.30 2.31 2.25 1.57 2.31 2.35 2.31 2.30 2.24

Table 2
Wilcoxon sign/range test for each time–frequency variables of HRV and model-based estimates. p > 0.05 is considered to be non-significant (NS). p < 0.05 is considered to be significant. Most significant (p < 0.001) differences
between consecutive phases of VM are indicated in bold. Percentage of increase ↑ or decrease ↓ is shown. Values are average over all data sets.

Phase LF HF SDNN RMSSD MATpar MATsym TATpar TATsym TAT ATB MIATpar MIATsym MAATpar MAATsym

0→1 ↑130 ± 95% ↓36 ± 25% ↓36 ± 25% ↓42 ± 34% NS NS ↓83 ± 2% ↓85 ± 2% ↓84 ± 2% NS ↑6 ± 2% NS NS ↓29 ± 10%
p < 0.01 p < 0.01 p < 0.01 p < 0.01 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.01
1→2 ↑116 ± 97% ↓69 ± 28% ↑237 ± 136% NS ↓32 ± 8% ↑415 ± 280% ↑253 ± 49% ↑2135 ± 711% ↑361 ± 76% ↑536 ± 193% ↓50 ± 14% NS NS ↑599 ± 259%
p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001
2→3 ↓43 ± 15% ↑352 ± 130% ↓25 ± 23% NS ↓22 ± 12% NS ↓83 ± 10% ↓61 ± 17% ↓77 ± 10% ↑268 ± 336% NS ↑464 ± 309% ↓38 ± 23% NS
p < 0.001 p < 0.001 p < 0.01 p < 0.01 p < 0.001 p < 0.001 p < 0.001 p < 0.05 p < 0.001 p < 0.001
3→4 ↓15 ± 40% ↑33 ± 35% NS ↑111 ± 77% ↑83 ± 46% ↓16 ± 102% ↑233 ± 302% ↓63 ± 14% NS ↓82 ± 13% NS ↓78 ± 22% ↑87 ± 63% ↓27 ± 26%
p < 0.05 p < 0.05 p < 0.001 p < 0.001 p < 0.01 p < 0.01 p < 0.001 p < 0.01 p < 0.001 p < 0.001 p < 0.05
4→5 NS NS ↓4 ± 2% ↓13 ± 19% NS ↓35 ± 24% ↑705 ± 201% ↑322 ± 219% ↑647 ± 176% ↓45 ± 29% ↑77 ± 55% ↑80 ± 87% ↓5 ± 6% ↓69 ± 23%
p < 0.05 p < 0.05 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.05 p < 0.05 p < 0.001
M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260 257

Multivariable linear regression analysis (*) considering each model-based estimate (row headers) as dependent variable and time–frequency estimates of HRV (column headers) as independent variables. Coefficient of regression points). HRmeas and PRmeas are mean of values in vectors contain-
ing individual heart rate and peripheral vascular resistance values
MAATsym

1 → 2**
r = 0.67
measured during the experiments on a beat-to-beat basis. HR and
PR are their corresponding simulated values.

NS

NS

NS
1 
n

J = |HRmeas (i) − HR(i)|2


n · HRmeas
i=1

b = −0.12
MAATpar

R = 0.91,


p < 0.05
n

3 → 4*
1
+ |PRmeas (i) − PR(i)|2 (35)
NS

NS

NS
(R), beta coefficients (b) and p-value (p) are reported only if p < 0.05, otherwise a correlation analysis (**) is performed and correlation coefficient (r) is reported only if r > 0.5. NS: non-significant.

n · PRmeas
i=1

In order to find the parameters values that minimize the


cost function (35) the Nelder–Mead algorithm [20] was used for
b = −2.24
R = 0.81,
MIATsym

bound-constrained optimization. It is based on cost functional eval-


p < 0.05

4 → 5**
r = 0.52
3 → 4*

uations of sequences of downhill simplexes. The model includes 18


NS

NS
unknown parameters which were estimated for each data set after
an average of 2983 iterations and 38 468 functions evaluations with
a final optimal cost function value in order of 0.000228002. Mean
parameters values are presented in Table 2 for 12 data sets obtained
from 5 subjects. The table includes parameters constraints in form
R = 0.91,

p < 0.05
b = 0.47
MIATpar

1 → 2*

of a finite range (a minimum and a maximum value). Optimized


NS

NS

NS

starting condition for model state Pas was found to be on aver-


age Pas (0) = 82 mmHg obtained with bounds constraints from 80
to 120 mmHg.
b = −7.14

b = −1.09

5.2. Model validation


R = 0.91,

R = 0.91,

R = 0.91,
R = 0.90,
p < 0.05

p < 0.05

p < 0.05

p < 0.05
b = 1.33

b = 0.31
2 → 3**
r = 0.60
3 → 4*

1 → 2*

3 → 4*

3 → 4*
ATB

NS

Figs. 4 and 5 show a good fit of model outputs to data obtained


from individual measurements on one representative 23 years old
healthy man. 97.16% and 97.44% of residuals for heart rate and
peripheral vascular resistance respectively are within the con-
fidence interval with 95% significance level. The corresponding
0 → 1**
r = 0.60

parameter estimates are included in Table 1 under data set 5. The


TAT

NS

NS

NS

average fit over all 10 datasets was found to be 94.23 ± 4.2%.


During phase 1 of the maneuver, the gradual increase in intra-
thoracic and intra-abdominal pressure causes aortic and peripheral
vessels compression. It follows an increase in peripheral resistance
b = −26.1
b = −5.09

resulting in a mild increase in blood pressure from 15th to 16th


R = 0.95,

R = 0.95,
p < 0.05

p < 0.01

1 → 2**

3 → 4**
r = 0.62

r = 0.69
1 → 2*

1 → 2*
TATsym

second (see Fig. 1).


NS

During phase 2, the sustained high intra-thoracic and intra-


abdominal pressure hinders venous return, leading to a decrease
in cardiac output and consequently a marked decrease in arterial
systolic pressure from the 16th to the 20th second (see Fig. 1).
This fall in arterial pressure is detected by baroreceptors in the
0 → 1**
r = 0.61

carotid artery sinus and aorta. The autonomic response includes


TATpar

NS

NS

an increase of sympathetic tone (see Fig. 6b) to the heart and ves-
sels as well as a decrease of parasympathetic tone (see Fig. 6a) in
the heart. This causes a reflex tachycardia and peripheral vasocon-
striction (see increased peripheral vascular resistance in Fig. 5). The
sympathetically mediated response results in some rise of arterial
1 → 2**
MATsym

r = 0.68

blood pressure from 20th to 30th second (see Fig. 1), which how-
ever does not overcome the elevated intra-thoracic pressure Since
NS

NS

NS

the blood pressure is still below the normal level, the baroreflex
remains active in the whole phase 2, resulting in a continuous rise
of heart rate [21].
During phase 3, the forced expiration is released, resulting in a
R = 0.87,

p < 0.05
b = 0.02

2 → 3**
r = 0.53
MATpar

decrease in intra-thoracic pressure, thus the pressure around the


1 → 2*

aorta decreases and consequently arterial blood pressure slightly


NS

NS

decreases from 30th to 31st second (see Fig. 1). Meanwhile heart
rate is still elevated.
During phase 4, the accumulated venous blood returns to the
heart and is pumped into the constricted arteries causing an “over-
RMSSD

shoot” of arterial pressure above the normal level from 31st to 38th
SDNN
Table 3

second (Fig. 1). This is detected by the baroreceptors and results in


HF
LF

a reflex bradycardia. Sympathetic discharge to the heart decreases


258 M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260

Fig. 4. Heart rate change during VM: (a) predicted value (dashed line) and experimental data set 5 (plain line) shows a 97.16% good fit (b).

Fig. 5. Changes in vascular resistance during VM: (a) predicted value (dashed line) and experimental data set 5 (plain line) shows a 97.44% good fit (b).

I II III IV I II III IV
0.18
0.85 (a) Tpar (b) Tsym

0.16
paraympathetic tone

sympathetic tone

0.8
0.14

0.12
0.75

0.1

0.7
0.08

0.06
0 10 20 30 40 50 60 0 10 20 30 40 50 60
seconds seconds

Fig. 6. Predicted parasympathetic (a) and sympathetic (b) activity during the Valsalva Maneuver for a 23 years old healthy man (data set 5). The maneuver covers the
timeline from the 15th to the 30th second. Phase 1 is characterized by slight increase of parasympathetic activity from 15th to 16th second; there is baroreflex mediated
parasympathetic withdrawal and increased sympathetic tone in phase 2 until 30th second. Bradycardia characterizes phase 4 which runs from 31st to 38th second. Sympathetic
activity is reduced in phase 4 in order to compensate arterial pressure overshot.

(see Fig. 6b) whereas parasympathetic discharge increases (see and suggest which variables are the most suitable for assessing
Fig. 6a), causing a decrease of heart rate. As result arterial blood dynamics of autonomic nervous activity for a given transition.
pressure returns to normal level. For example, the gradual increase in intra-thoracic pressure dur-
ing phase 1 significantly decreases the total autonomic tone TAT
5.3. Model-based estimates of autonomic tone and statistical by 84 ± 2% (p < 0.001) from its baseline value 13.88 ± 0.22. HF,
analysis SDNN, RMSSD and MAATsym decrease as well, but to a less signifi-
cant extend (p < 0.01). In phase 2, baroreflex activation following
Eqs. (18)–(27) were resolved in Matlab using built-in statistical sustained decrease in arterial blood pressure causes a signifi-
functions and following model-based estimates of autonomic ner- cant (p < 0.001) increase of LF (116 ± 97%), SDNN (237 ± 136%),
vous activity were calculated for each data set and each phase of MATsym (415 ± 280%), TAT (361 ± 76%), ATB (536 ± 193%), MAATsym
the Valsalva Maneuver. We therefore obtained 10 values per phase (599 ± 259%) from their average values in phase 1 (0.49 ± 0.21 n.u.,
for each model-based estimate and for each subject. 13.26 ± 5 ms, 0.057 ± 0.018 n.u., 2.28 ± 0.23 n.u., 0.066 ± 0.024 n.u.,
In the statistical analysis, variables taken into consideration 0.058 ± 0.018 n.u., respectively). Meanwhile HF, MATpar , MIATpar
were the percentage of increase/decrease of the normalized significantly decrease by 69 ± 28% (p < 0.001), 32 ± 8% (p < 0.001)
power in the frequency bands LF and HF; the percentage of and 50 ± 14% (p < 0.001), respectively. Prevalent parasympathetic
increase/decrease of time-domain indexes of heart rate variability activity in phase 4 is significantly reflected in increasing HF
SDNN and RMSSD; the percentage of increase/decrease of model- (33 ± 35%: p < 0.05), RMSSD (111 ± 77%: p < 0.001), MATpar (83 ± 46:
based estimates of autonomic tones when moving from one phase p < 0.001), TATpar (233 ± 302%: p < 0.01) and MAATpar (87 ± 63%:
of the maneuver to the consecutive one. p < 0.001) meanwhile LF, MATpar , TATsym , ATB and MAATsym signifi-
In order to determine whether there was some pattern of change cantly decrease (p < 0.05).
among the ranges of the statistical variables by comparing two In order to establish the degree of linear dependency of the
consecutive phases, a Wilcoxon range test was performed. Table 2 model-based estimates with respect to time–frequency estimates
shows which variables change significantly (p < 0.05) when tran- of HRV, different models of multi-linear regression analysis were
siting from one phase of the VM to the next consecutive one. used. For each phase, we took the percentage of change of each
The most significant (p < 0.001) variability are highlighted in bold model-based estimate as dependent variable and the percentage of
M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260 259

Fig. 7. Fit between HRV estimates HF & LF and model-based estimates MATpar & TATsym .

Fig. 8. Fit between HRV estimates RMSSD & SDNN and model-based estimates MAATpar & ATB.

Fig. 9. Mean values of significant model-based and HRV estimates of autonomic nervous activity in each phase transition of VM.

change of all time–frequency estimates as independent variables. between LF and TATsym (R = 0.95, p < 0.05) as illustrated by the line
Coefficient of regression (R), beta coefficients (b) and p-values are fit plot between their percentage of increase in Fig. 7 (right panel).
presented in Table 3 for significant regression (p < 0.05). In the cases The transition 2 → 3 was characterized by a significant correla-
where no significant regression (p > 0.05) was found, an additional tion between MATpar and RMSSD (r = 0.53); and a more significant
correlation analysis was performed. The correlation coefficient (r) correlation between ATB and SDNN (r = 0.60). The transition 3 → 4
is included in Table 3 for significant correlations (r > 0.50). NS is highlighted a significant regression between MAATpar and RMSSD
reported for non-significant correlations (r < 0.50). (R = 0.91, p < 0.05) on one hand and between ATB and SDNN on the
In the transition 0 → 1, from baseline to phase 1, TATpar and other hand (R = 0.91, p < 0.05). This means that, knowing the per-
TAT were found to be correlated to RMSSD (r = 0.6). The transition centage of increase in RMSSD and the percentage of decrease in
1 → 2 was characterized by a significant regression between MATpar SDNN when transiting from phase 3 to phase 4, we were able to
and HF (R = 0.87, p < 0.05) as illustrated in the line fit plot in Fig. 7 predict the percentage of decrease in MAATpar and the percentage
(left panel). Knowing the percentage of decrease in the power of of increase in ATB, respectively, with a 91% fit, as depicted in Fig. 8.
high frequency components of heart rate variability HF between Finally a less significant correlation (r = 0.52) was found between
phase 1 and phase 2, we were able to predict the percentage of MIATsym and SDNN in the transition 4 → 5, i.e. from phase 4 to post
decrease of the model-based estimate of mean parasympathetic VM phase.
tone MATpar with a 87% fit. A significant regression was also found
260 M. Kana, J. Holcik / Biomedical Signal Processing and Control 6 (2011) 251–260

Based on results of the Wilcoxon sign test, linear regression and could be extended in a future effort to other mathematical models
correlation analysis above, we were able to select the most signif- of cardiovascular control and to other experimental setups such as
icant estimates of autonomic nervous activity in each single phase active change of posture. Extending the model to study the effect
of the VM. Results are shown in Fig. 9. For each transition, the plots of gender, age and health status on the value range of model-based
show the mean value and confidence interval of the selected signif- estimates is also part of a future research.
icant estimates of sympathetic and parasympathetic activity in the
previous and current phase; as well as the percentage of increase Acknowledgments
or decrease of those estimates. It was found that SDNN and RMSSD
respectively provide the most significant estimates of sympathetic The research was partially granted by the research program
and parasympathetic activity in phase 1 compared to phase 0 (base- No. MSM6840770012 “Transdisciplinary Research in the Field of
line) and in phase 5 (post VM) compared to phase 4. Similarly LF Biomedical Engineering II.” and by the project VEGA No.2/0210/10
and HF are the estimates of choice when it respectively comes to “Methods and systems for multichannel measurement and evalu-
assess sympathetic and parasympathetic activity in phase 3 with ation of bioelectric signals of heart and brain”.
respect to phase 2. However, our model-based estimates TATsym
and ATB were found to provide the most significant estimates of References
sympathetic activity in phase 2 and 4, with respect to phases 1
and 3, respectively. MATpar was found to provide the most signifi- [1] S.H. Hohnloser, T. Klingenheben, Clinical guide to cardiac autonomic tests, in:
Clinical Guide to Cardiac Autonomic Tests, Kluwer Academic Publisher, 1998,
cant estimate of parasympathetic activity in phase 2 with respect p. 56.
to phase 1, similarly MAATpar was found to assess parasympathetic [2] A. Minchole, J.P. Martinez, P. Arini, M. Risk, P. Laguna, T wave width alter-
activity in phase 4 with respect to phase 3 in a significant way. ations during Valsalva Maneuver in diabetic patients, Computers in Cardiology
(September 17–20) (2006) 709–712.
[3] S. Kujanik, M. Rakarova, J. Petrovicova, QT-RR regression lines during Val-
6. Conclusion salva Maneuvre in young healthy women are different from resting conditions,
Scripta Medica 77 (2004) 135–144.
[4] R. Burattini, P. Borgdorff, N. Westerhof, Baroreflex open-loop gains estimated
In this work we proposed markers of autonomic nervous activ-
in humans by graded Valsalva maneuver, Computers in Cardiology (September
ity using a mathematical model of regulation processes performed 23–26) (1991) 689–691.
by the autonomic nervous system on the cardiovascular system [5] M. Mendez, M.J. Gaitan, S. Carrasco, Time–frequency assessment of heart rate
during the Valsalva Maneuver. The model could fit measured beat- variability under Valsalva and Mueller maneuvers, in: 25th Annual Intem-
ational Conference of the IEEE EMBS (CDROM), Cancun, Mexico, 2003.
to-beat heart rate and galvanic skin response signals with low [6] G. Ramos, S. Carrasco, V. Medina, Time–frequency analysis of the heart rate
residuals error (average fit 94.23 ± 4.2%). We defined model-based variability during the Valsalva manoeuvre, Journal of Medical Engineering &
estimates for quantitative approximation of mean, total, minimal Technology 24 (2000) 73–82.
[7] M. Malik, Heart rate variability. Standards of measurement, physiological inter-
and maximal autonomic activity, as well as autonomic balance for pretation and clinical use, Circulation 93 (March (5)) (1996) 1043–1065.
each phase of VM. We performed an extended statistical analysis [8] M. Risk, M. Berghoff, R. Freeman, Characterization of the valsalva maneuver
including Wilcoxon sign test, multivariable linear regression analy- using wavelet transform, Computers in Cardiology (2000) 411–414.
[9] K. Lu, J.W. Clark, F.H. Ghorbel, D.L. Ware, A. Bidani, A human cardiopulmonary
sis and correlation analysis in order to study patterns of variation of system model applied to the analysis of the Valsalva maneuver, Journal of
our model-based estimates and possible relationships with current Physiology: Heart and Circulatory Physiology 281 (December (6)) (2001).
time–frequency markers of heart rate variability. Results show that [10] F. Liang, H. Liu, Simulation of hemodynamic responses to the Valsalva Maneu-
ver: an integrative computational model of the cardiovascular system and the
when it comes to assessing autonomic nervous activity as absolute autonomic nervous system, The Journal of Physiological Science 56 (1) (2006).
quantity within a single phase with disregard to transition to the [11] I. Biopac Systems, BSL Hardware Guide, 2010.
next phase, we can select SDNN (standard deviation of NN inter- [12] W.J. Pan, J. Tompkins, A real-time QRS detection algorithm, IEEE Transactions
on Biomedical Engineering 32 (1985) 230–236.
vals) in phases 0 & 5, TATsym (total autonomic tone of sympathetic
[13] N.R. Lomb, Least-squares frequency analysis of unequally spaced data, Astro-
activity) in phases 1 & 2 and ATB (autonomic tone balance) in phases physics and Space Science 39 (1976) 447–462.
3 & 4 for sympathetic activity; RMSSD (mean square successive dif- [14] J.J. Goldberger, M.W. Ahmed, M.A. Parker, A.H. Kadish, Dissociation of heart rate
ferences of NN intervals) in phases 0 & 5, MATpar (mean autonomic variability from parasympathetic tone, American Journal of Physiology: Heart
and Circulatory Physiology 266 (1994) H2152–H2157.
tone of parasympathetic activity) in phases 1 & 2, MAATpar (max- [15] J. Batzel, F. Kappel, D. Schneditz, H.T. Tran, Cardiovascular and Respiratory
imal autonomic tone of parasympathetic activity) in phases 3 & 4 Systems – Modeling, Analysis and Control, Society for Industrial and Applied
for parasympathetic activity. Mathematics, Philadelphia, 2007.
[16] F.S. Grodins, Control Theory and Biological Systems, Columbia University Press,
The absolute values of model-based estimates presented in this New York, 1963.
work do not correspond to the real firing rates. They rather give [17] M.S. Olufsen, H.T. Tran, J.T. Ottesen, Modeling baroreflex regulation of heart
means of comparison of the change in efferent neural activity. rate during orthostatic stress, Journal of Physiology: Regulatory, Integrative
and Comparative Physiology 29 (1) (2006) 1355–1368.
Future research includes validating the predicted markers of auto- [18] H. Schmidt, M. Jirstrand, Systems Biology Toolbox for MATLAB: a computational
nomic tones against data from experiments with sympathetic and platform for research in systems biology, Bioinformatics 22 (4) (2006) 514–
parasympathetic blockades. We expect e.g. the model to return 515.
[19] L.F. Shampine, M.W. Reichelt, The MATLAB ODE suite, SIAM Journal on Scientific
MATpar closed to zero in the presence of atropine. The proposed Computing 18 (1997) 1–22.
model-based estimates are closely related to the design of the [20] J.A. Nelder, R. Mead, A simplex method for function minimization, The Com-
underlying mathematical model and the experimental setup (Val- puter Journal 7 (1965) 308–313.
[21] D. Taylor, The Valsalva Manoeuvre: a critical review, South Pacific Underwater
salva Maneuver in our case). The basic principles of our method
Medicine Society Journal 26 (1) (1996).

Potrebbero piacerti anche