Estimation of respiratory parameters

via fuzzy clustering

R. Babuska
a,*
, L. Alic
a
, M.S. Lourens
b
,
A.F.M. Verbraak
b
, J. Bogaard
b
a
Department of Information Technology and Systems, Control Engineering Laboratory,
Delft University of Technology, P.O. Box 5031, 2600 GA Delft, The Netherlands
b
Department of Pulmonary and Intensive Care Medicine, Erasmus Medical Centre, Rotterdam,
3015 GD Rotterdam, The Netherlands
Received 24 February 2000; received in revised form 12 July 2000; accepted 1 August 2000
Abstract
The results of monitoring respiratory parameters estimated from owpressurevolume
measurements can be used to assess patients' pulmonary condition, to detect poor patient
ventilator interaction and consequently to optimize the ventilator settings. A new method is
proposed to obtain detailed information about respiratory parameters without interfering with the
expiration. By means of fuzzy clustering, the available data set is partitioned into fuzzy subsets that
can be well approximated by linear regression models locally. Parameters of these models are then
estimated by least-squares techniques. By analyzing the dependence of these local parameters on
the location of the model in the owvolumepressure space, information on patients' pulmonary
condition can be gained. The effectiveness of the proposed approaches is demonstrated by
analyzing the dependence of the expiratory time constant on the volume in patients with chronic
obstructive pulmonary disease (COPD) and patients without COPD. #2001 Elsevier Science B.V.
All rights reserved.
Keywords: Respiratory mechanics; Mechanical ventilation; Parameter estimation; Respiratory resistance and
compliance; Expiratory time constant; Fuzzy clustering; Least-squares estimation
1. Introduction
The importance of monitoring respiratory mechanics in patients on ventilatory support is
generally accepted. Respiratory parameters can be used to assess patients' pulmonary
condition, to detect poor patientventilator interaction and consequently to optimize
Artificial Intelligence in Medicine 21 (2001) 91105
*
Corresponding author. Tel.: 31-15-2785117; fax: 31-15-2786679.
E-mail addresses: r.babuska@its.tudelft.nl (R. Babuska), verbraak@lonf.azr.nl (A.F.M. Verbraak).
0933-3657/01/$ see front matter # 2001 Elsevier Science B.V. All rights reserved.
PII: S0 9 3 3 - 3 6 5 7 ( 0 0 ) 0 0 0 7 5 - 0
ventilator settings. Monitoring respiratory mechanics is of paramount importance in
mechanically ventilated patients with chronic obstructive pulmonary disease (COPD).
In these patients, due to the disturbance in respiratory mechanics, it is difcult to set the
ventilator and weaning from the ventilator is cumbersome. By identifying these patients,
assessing their pulmonary condition and patientventilator interaction, ventilator settings
and medical treatment can be adjusted resulting in a more favorable outcome. The
expiratory time constant provides information on the mechanical properties of the
respiratory system, it is a measure of lung emptying and can be used to predict the
minimal time needed for complete expiration. Several methods have been proposed to
determine the expiratory time constant [13]. However, these methods either interfere with
the expiration [2,3], or assume a linear relationship between owand volume [1]. The latter
has to be questioned in patients with COPD, in view of the presence of ventilatory
inhomogeneity and expiratory ow limitation. In this article, a new method is proposed to
estimate respiratory parameters without interfering with the expiration. Moreover, the
method is very general in the sense that it does not assume any particular functional
relationship between the ow and volume.
The monitoring of lungs can be accomplished by analyzing the pressure, ow and
volume measurements. Typically, a low-order physical model of the respiratory dynamics
is assumed and parameters in this model are estimated from the data [1,46]. Two main
approaches can be distinguished in the literature: time-domain estimation and frequency-
domain estimation. In this article, the time-domain approach is followed, as it enables us to
gain insight into nonlinear phenomena and exploit for diagnosis the information obtained.
Frequency domain techniques, however, are limited to linear models and are less suitable
for this purpose.
The method presented in this paper uses a straightforward extension of the classical
linear single-compartment model [4,5]. This model describes the dynamic relation
between the pressure P (hPa), the air ow-rate

V (l/min) and the volume V (l) of the lungs
P = EV R

V P
0
X (1)
The respiratory elastance E (hPa/l), the resistance R (hPa s/l) and the elastic recoil pressure
P
0
(pressure when the volume V equals zero) are free parameters to be estimated fromdata.
It is well known that this linear model may yield too coarse an approximation of the
given data, especially for patients with pulmonary disorders (such as the COPD). There-
fore, modications of (1) have been proposed [4]. One can, for instance, use different
parameters for inspiration and expiration or nonlinearities can be introduced by consider-
ing E and R to be functions of the volume or the ow [5].
The approach proposed in this paper is based on automatic detection (localization) of
multiple local linear models. Hence, one does not need to make any assumptions about the
mathematical form or parameterization of the nonlinearity. By observing the dependence
of the local respiratory parameters on the location of the model in the owvolume
pressure space, information about the ventilated subject can be obtained.
The two main techniques used to obtain the parameters of the multiple models are fuzzy
clustering and linear least-squares estimation. By means of fuzzy clustering, the available
data set is partitioned into fuzzy subsets that can be well approximated by linear regression
models locally. Parameters of these models are then estimated by least-squares techniques.
92 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
The use of clustering techniques is motivated by the fact that they can reveal structures in
data without relying on assumptions common to conventional statistical methods, such as
the underlying statistical distribution. Clustering has been successfully used in a variety of
elds, including classication, image processing, pattern recognition, modeling and
identication. An increasing number of applications can be found in the medical eld.
Typical application domains include image processing for computer-aided diagnosis [79],
signal processing in evoked potentials estimation [10], analysis of time series for imaging
[11], and classication [12].
The current paper is organized as follows. The patients and the respiratory measure-
ments are described in Section 2. The applied methods and algorithms can be found in
Section 3, while Section 4 presents the results obtained for the estimation of the expiratory
time constant. Section 5 gives a discussion of the results, conclusions and suggestions for
future research.
2. Problem description
In this section, the conducted study is described in terms of the patients, the
mechanical ventilation conditions, and the collection and preprocessing of the respiratory
data.
2.1. Patients
Twenty patients admitted to the medical intensive care unit of the Erasmus Medical
Centre in Rotterdam were studied. Patients were included if they fullled the following
criteria: mechanical ventilation via an endotracheal or tracheostomy tube and the absence
of air leaks.
Ten of these patients had a history of severe chronic obstructive pulmonary
disease (COPD) according to the European Respiratory Society consensus; a clinical
diagnosis of COPD and previous lung function data showing a forced expiratory
volume in 1 s (FEV1) less than 50% of predicted normal value (mean 29% of predicted,
range 2137%) [13]. These 10 patients were ventilated because of respiratory failure due to
an exacerbation of their COPD. In the other 10 patients, underlying diseases included a
variety of medical conditions all complicated by respiratory failure and ventilator
dependency.
All patients were mechanically ventilated with the Siemens Servo 300 ventilator
(Siemens-Elema, Solna, Sweden). Ventilator settings were set by the primary physician
and remained unchanged during the study, with the exception that if positive end-
expiratory pressure (PEEP) was present, it was removed. All patients were ventilated
in the volume-controlled mode with an average minute volume of 8.5 l/min (range
6.515.0 l/min). The average respiratory rate was 12 breaths per minute (range 820).
The ratio between inspiratory and expiratory time was 35:65. During the study, the patients
were sedated with midazolam (Roche Nederland B.V., Mijndrecht, The Netherlands).
Informed consent was obtained fromthe patients or their next of kin. The measurements for
this study were approved of by the local ethics committee.
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 93
2.2. Respiratory measurements
The ow

V was measured by a heated pneumotachometer (Lilly, Jaeger, Wurzburg,
Germany) connected to the endotracheal tube. The airway opening pressure Pwas measured
proximal to the pneumotachometer by a pressure transducer (Validyne, Validyne Co.,
Northridge, USA). A 12-bit AD converter was used to convert the measured analog signals
to digital data at a sampling frequency of 100 Hz. Data were stored and analyzed using a
personal computer. On average, 3000 samples were used for the analysis of each patient.
The rst two panels of Fig. 1 show typical pressure and ow-rate signals for a
representative patient without COPD. The bottom panel gives the volume V obtained
by numerical integration of the sampled ow signal. Note that a drift in the volume signal
occurs, which is mainly caused by leakage and by the difference in temperature and relative
humidity between inspiration and expiration. This drift can be partially removed by adding
an offset signal obtained by tting a line or a low-order polynomial trough the volume
minima in the individual cycles.
3. Parameter estimation through fuzzy clustering
This section presents the methods and algorithms applied to the analysis of the data and
parameter estimation.
Fig. 1. Sample data set of a representative patient without COPD.
94 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
3.1. The data set to be clustered
In our application, the data are the sampled measurements of the pressure, ow and
volume. Generally, there can be any observations, each consisting of n measured variables,
grouped into an n-dimensional column vector z
k
= [z
1k
Y F F F Y z
nk
[
T
, z
k
R
n
. A set of N
observations is denoted by Z = z
k
[k = 1Y 2Y F F F Y N, and is represented as an n N
matrix:
Z =
z
11
z
12
z
1N
z
21
z
22
z
2N
F
F
F
F
F
F
F
F
F
F
F
F
z
n1
z
n2
z
nN
P
T
T
T
R
Q
U
U
U
S
X (2)
In the analysis of respiratory measurement, each column of Z is generally the ow
volumepressure triplet: z
k
= [V(k)Y

V(k)Y P(k)[
T
, although only some of these signals can
be used when clustering is applied locally to the inspiration or expiration phase. Details can
be found in Section 4.
3.2. Fuzzy partition
The objective of clustering is to partition the data set Z into c clusters. Fuzzy partition of
Z is a family of fuzzy subsets A
i
[1 _ i _ c. The subsets are dened by their membership
(characteristic) functions, represented in the partition matrix U = [m
ik
[
cN
. The ith row of
this matrix contains values of the membership function of the ith fuzzy subset A
i
of Z. The
partition matrix satises the following conditions:
m
ik
[0Y 1[Y 1 _ i _ cY 1 _ k _ NY (3a)

c
i=1
m
ik
= 1Y 1 _ k _ NY (3b)
0 `

N
k=1
m
ik
` NY 1 _ i _ cX (3c)
Eq. (3a) states the well-known fact that the membership degrees are real numbers from the
interval [0Y 1[. Condition (3b) constrains the sum of each column to 1, and thus the total
membership of each z
k
in all the clusters equals one. Eq. (3c) means that none of the fuzzy
subsets is empty nor it contains all the data.
Let us illustrate the concept of fuzzy partition by a simple example. Consider a data set
Z = [z
1
Y z
2
Y F F F Y z
9
[, shown in Fig. 2.
By using multiple-model regression, this data set can be approximated by two lines (data
points z
1
to z
4
and z
6
to z
9
, respectively). Note that point z
5
does not fully belong to either of
the models. The corresponding fuzzy partition of Z is:
U =
1X0 1X0 1X0 1X0 0X5 0X0 0X0 0X0 0X0
0X0 0X0 0X0 0X0 0X5 1X0 1X0 1X0 1X0
!
X
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 95
The rst row of U denes the membership function for the rst subset A
1
, the second
row denes the membership function of the second subset A
2
. Membership degrees of
z
5
reect the fuzziness of the partition. Note that the sum of the rows is one for all data
points.
3.3. Clustering algorithm
In this application, the fuzzy partition matrix is obtained by applying the Gustafson
Kessel (GK) algorithm [14], based on the minimization of the well-known fuzzy c-means
functional:
J(ZY UY VY A
i
) =

c
i=1

N
k=1
(m
ik
)
m
D
2
ikA
i
Y (4)
where U = [m
ik
[ is the fuzzy partition matrix of Z, V = [v
1
Y v
2
Y F F F Y v
c
[, v
i
R
n
is a vector
of cluster prototypes (centers), which have to be determined, and
D
2
ikA
i
= (z
k
v
i
)
T
A
i
(z
k
v
i
) (5)
is the squared inner-product distance norm. Matrices A
i
are computed in the optimization
algorithm using the local covariance of the data around each cluster center. This allows
each cluster to adapt the distance norm to the local distribution of the data. If the data
samples are distributed along a nonlinear hypersurface, the GK algorithm will nd clusters
that are local linear approximations of this hypersurface. The overlap of the clusters is
controlled by the user-dened parameter m [1Y ).
The minimization of (4) represents a nonlinear optimization problem that is usually
solved by using the Picard iteration through the rst-order conditions for stationary points
of (4). This algorithm is stated (without derivation) in Algorithm 1.
Algorithm 1. GustafsonKessel (GK) algorithm
Given the data set Z, choose the number of clusters 1 ` c ` N, the weighting exponent
m b 1 and the termination tolerance E b 0. Initialize the partition matrix randomly, such
that it satises conditions (3a)(3c).
Fig. 2. A sample data set in R
2
.
96 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
Repeat for l = 1Y 2Y F F F
Step 1: Compute cluster prototypes (means):
v
(l)
i
=

N
k=1
(m
(l1)
ik
)
m
z
k

N
k=1
(m
(l1)
ik
)
m
Y 1 _ i _ cX
Step 2: Compute the cluster covariance matrices:
F
i
=

N
k=1
(m
(l1)
ik
)
m
(z
k
v
(l)
i
)(z
k
v
(l)
i
)
T

N
k=1
(m
(l1)
ik
)
m
Y 1 _ i _ cX
Step 3: Compute the distances:
D
2
ikA
i
= (z
k
v
(l)
i
)
T
[det(F
i
)
1an
F
1
i
[(z
k
v
(l)
i
)Y 1 _ i _ cY 1 _ k _ NX
Step 4: Update the partition matrix:
for 1 _ k _ N
if D
ikA
i
b 0 for 1 _ i _ c,
m
(l)
ik
=
1

c
j=1
(D
ikA
i
aD
jkA
j
)
2a(m1)
Y
otherwise
m
(l)
ik
= 0 if D
ikA
i
b 0Y and m
(l)
ik
[0Y 1[ with

c
i=1
m
(l)
ik
= 1Y
until | U
(l)
U
(l1)
|` E.
Appendix A contains a MATLAB implementation of the algorithm.
The following parameters must be specied before clustering: the number of clusters, c,
the `fuzziness' exponent, m, the termination tolerance, E. The choices for these parameters
are now discussed.
3.3.1. Number of clusters
The number of clusters can either be selected a priori or it can be automatically
determined by using cluster validity measures [1517] or by iterative merging or insertion
of clusters [18,19]. In this paper, c is a user-dened parameter.
3.3.2. Fuzziness parameter
The weighting exponent m inuences the fuzziness of the resulting partition. As m
approaches one fromabove, the partition becomes hard (non-fuzzy) (m
ik
0Y 1) and as m
goes to , the partition becomes completely fuzzy (m
ik
= 1ac). The standard value m = 2
is used in this paper.
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 97
3.3.3. Termination criterion
The GK algorithm stops iterating when the norm of the difference between U in two
successive iterations is smaller than the termination parameter E. For the maximum norm
max
ik
([m
(l)
ik
m
(l1)
ik
[), the usual choice is E = 0X01.
3.4. Least-square estimation of the respiratory parameters
The GK algorithm yields a partition matrix which determines to what degree the
individual data samples belong to the individual linear submodels. In order to estimate the
parameters of these submodels, the fuzzy partition is rst converted into a crisp one by
applying an a-cut. This means that only the data samples that belong to the given cluster to
a degree greater than a specied threshold a [0Y 1[ are used to obtain the estimate:
Z
i
= z
k
[m
ik
b aY k = 1Y 2Y F F F Y NY i = 1Y 2Y F F F Y cX (6)
Recall that our Z contains the owvolumepressure triplets, z
k
= [V(k)Y

V(k)Y P(k)[
T
.
Hence, assuming a local linear model (1), any least-squares estimation method can be
applied to estimate the local respiratory elastance E, resistance R and the elastic recoil
pressure P
0
.
Note, that while all three parameters can easily be estimated in the global model (1), only
some of these parameters can be estimated locally. For instance, in the expiration phase,
there is no ow imposed and the expiration is a free-run process described by a rst-order
differential equation
V t

V = 0Y (7)
where t is a time constant to be estimated. The following section gives more details on the
analysis of expiration data.
4. Application of fuzzy clustering to respiratory data
Fuzzy clustering can be applied either to the entire respiratory cycle (or several cycles)
or to the inspiration and expiration separately. The former approach is useful for the
detection of those parts of the respiratory cycle that are well described by local linear
models and thus yield more reliable and accurate parameter estimates. This method is
briey outlined in Section 4.1. By clustering the inspiration and expiration data separately,
insight can be gained in the dependence of respiratory parameters on variables like the
volume, for instance. This method has been investigated in more detail and its application
to the analysis of the expiratory time constant is described in Section 4.2.
4.1. Clustering of the entire respiratory cycle
In this section, the application of fuzzy clustering to the entire respiratory cycle is
demonstrated. It is well known, that the parameters of (1) may vary during the respiratory
cycle [4]. This variability is caused by the fact that the underlying phenomena are actually
nonlinear. Rather than assuming a particular form of the nonlinearity, we propose to use
98 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
fuzzy clustering to automatically nd segments of the cycle that are well explained by local
linear models and thus yield reliable parameter estimates. Fig. 3 shows an example of a
partitioning obtained by clustering the variables [VY

VY P[ into four clusters (this number
was chosen quite arbitrarily).
Note that the obtained segmentation of the cycle is different from a manual splitting into
inspiration and expiration. For instance, the initial samples of inspiration and expiration are
not included. The reason for this is that these data samples are not well explained by the
local linear model estimated for the rest of the phase. This is typically due to nonlinear
phenomena taking place at the transition between the different phases. By adjusting the a-
level, the number of samples included in the given local linear model can be controlled.
Large a values will result in more separation between the segments (more data samples will
be discarded). The partitioning can, of course, be obtained from a data sequence consisting
of several cycles (see Fig. 1). This improves the quality of the partitioning, as well as the
subsequent estimation of the respiratory parameters.
4.2. Clustering of expiration data
In this section, only the clustering of expiration data is considered. The goal is to reveal
the dependency of the expiratory time constant on the volume. The expiration data are
obtained by extracting the parts of the cycles between the end of the inspiratory pause and
the beginning of the next cycle. This is done by a simple detection algorithm, clustering is
not used at this stage. The reason for not using fuzzy clustering is that here we are interested
in the nonlinearity. Hence, the complete expiration phase is extracted.
As the expiration is a free-run process described by a rst-order differential equation (7),
only the time constant t can be estimated and analyzed. The columns of Z are thus the
volumeow pairs: z
k
= [V(k)Y

V(k)[
T
. In each subset obtained by clustering, the time
constant t is estimated. The trend of t among the local models found yields useful
information about the subject's condition.
Fig. 3. Partitioning of the entire respiratory cycle. The phases are obtained by applying the a-cut (6) with
a = 0X8.
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 99
Fig. 4 shows a typical result obtained by clustering the expiratory data for a patient
without COPD. The expiratory owvolume curve is characterized by a smooth transition
between the initial and later part of the curve. Note the decreasing trend of the time constant
with decreasing volume (bottom plot).
A typical result for a patient with COPD is given in Fig. 5. An abrupt transition is
observed from a small to a large time constant between the initial and the later part of
expiration. The time constant increases with decreasing volume.
Fig. 6 clearly shows the distinction between the patients with and without COPD. In this
gure, the trend of the expiratory time constant t at the beginning of expiration
Dt =
t
2
t
1
v
2
v
1
(8)
is plotted versus the average expiration time constant in clusters 2 and 3, (t
2
t
3
)a2. Here,
v
1
and v
2
denote the volume coordinate of the rst and second cluster center, respectively.
The large difference between the time constant in cluster 1 and the consecutive clusters
in the patients with COPDis also apparent in Table 2. Table 1 shows the pattern for the non-
COPD group.
Fig. 4. Patient without COPD. Top: volumeow-rate data partitioned into four clusters. Middle: the
membership degrees of the four clusters plotted against the volume data. The a = 0X8 threshold is shown as well.
Bottom: the time constants of the four local models plotted against the volume coordinate v
i
of cluster centers
(linearly interpolated).
100 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
Fig. 5. Patient with COPD. Top: volumeow-rate data partitioned into four clusters. Middle: the membership
degrees of the four cluster plotted against the volume data. The a = 0X8 threshold is shown as well. Bottom: the
time constants of the four local model plotted against the volume coordinate of cluster centers (linearly
interpolated).
Table 1
Time constant for patients without COPD
Data set Cluster 1 Cluster 2 Cluster 3 Cluster 4
cont1 0X60 0X03 0X51 0X02 0X42 0X01 1X60 0X00
cont2 0X83 0X01 0X73 0X01 0X59 0X01 0X51 0X04
cont3 1X16 0X01 0X97 0X00 0X76 0X01 1X12 0X02
cont4 1X11 0X01 0X81 0X01 0X62 0X02 3X22 0X01
cont5 1X09 0X01 0X88 0X05 1X03 0X02 1X13 0X01
cont6 0X44 0X03 1X29 0X02 1X50 0X02 2X08 0X02
cont7 0X56 0X01 0X88 0X01 1X47 0X01 2X49 0X01
cont8 0X74 0X01 1X22 0X00 1X18 0X00 1X43 0X01
cont9 0X27 0X08 0X79 0X03 1X25 0X02 1X55 0X02
cont10 0X63 0X02 0X96 0X01 1X55 0X01 2X23 0X01
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 101
Fig. 6. The trend Dt at the beginning of expiration, dened by Eq. (8), plotted versus the average expiration
time constant in clusters 2 and 3 ((t
2
t
3
)a2). The two groups of patients are clearly visible (`o' with COPD,
`' without COPD).
Table 2
Time constant for patients with COPD
Data set Cluster 1 Cluster 2 Cluster 3 Cluster 4
copd1 0X22 0X19 3X59 0X02 6X75 0X01 8X04 0X01
copd2 0X32 0X06 1X39 0X01 3X03 0X01 6X29 0X01
copd3 0X32 0X05 1X40 0X01 2X05 0X01 2X47 0X01
copd4 0X14 0X15 1X55 0X01 2X21 0X01 3X01 0X01
copd5 0X15 0X10 0X84 0X04 3X36 0X01 12X34 0X00
copd6 0X51 0X03 2X42 0X01 3X77 0X01 4X93 0X01
copd7 0X18 0X06 1X53 0X01 2X49 0X01 3X15 0X01
copd8 0X14 0X08 2X11 0X01 2X95 0X01 5X15 0X01
copd9 0X51 0X03 1X69 0X01 2X86 0X01 4X15 0X01
copd10 0X77 0X05 3X49 0X01 4X98 0X02 5X17 0X02
102 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
5. Discussion and conclusions
This study shows that in mechanically ventilated patients with and without COPD, fuzzy
clustering can be applied to assess respiratory parameters. In previous studies we have
shown that the plot of the expiratory ow versus expired volume (expiratory owvolume
curve) provides information about mechanical properties of the respiratory system [20,21].
The inverse of the slope of these curves can be interpreted as a time constant, describing
lung emptying. Mostly, a specic part of the curve (the latter part) is used to calculate a
single time constant. In this study we applied a method based on automatic detection of
multiple local linear models which enables the description of the time constant behavior of
the whole curve.
There is a clear difference in the shape of the expiratory owvolume curve between
patients with and without COPD [20,2224]. In patients without COPD, the expiratory
owvolume curve is mostly characterized by a smooth transition between the initial and
later part of the curve. In patients with COPD, however, an abrupt transition is observed
from a small to a large time constant between the initial and the later part of expiration.
This is caused by airway compression, related to both a decreased lung elastic recoil and an
increased airway resistance in those patients. Fig. 6 clearly shows the distinction between
the groups with and without COPD. The large difference between the time constant in
cluster 1 and the consecutive clusters in the patients with COPD is also apparent in Table 2.
Table 1 shows the pattern for the non-COPD group. Patients 1 to 5 have normal or even stiff
(brotic) lungs. These patients were ventilated because of muscular weakness or pul-
monary brosis due to radiotherapy. Patients 6 to 10 were ventilated for various medical
conditions, but had either mild COPD (FEV1 between 50 and 70% of predicted) or
increased airway resistance. Besides the clear distinction between respiratory mechanics of
patients with and without COPD, fuzzy clustering may discriminate for other pulmonary
conditions, facilitating clinical diagnoses of mechanically ventilated patients in the future.
Besides these applications, the methodology may be useful in other nonlinear estimation
problems in the medicine.
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 103
Appendix A. MATLAB implementation of the GK algorithm
104 R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105
References
[1] Brunner JX, Laubscher TP, Banner MJ, Iotti G, Braschi A. Simple method to measure total expiratory time
constant based on the passive expiratory owvolume curve. Crit Care Med 1995;23:111722.
[2] Gottfried SB, Rossi A, Higgs BD, et al. Noninvasive determination of respiratory system mechanics during
mechanical ventilation for acute respiratory failure. Am Rev Respir Dis 1985;131:41420.
[3] Gottfried SB, Higgs BD, Rossi A, et al. Interrupter technique for measurement of respiratory mechanics in
anesthetized humans. J Appl Physiol 1985;59:64752.
[4] Peslin R, da Silva F, Duvivier C, Chabot F. Respiratory mechanics studied by multiple linear regression in
unsedated ventilated patients. Eur Respir J 1992;5:8718.
[5] Lauzon AM, Bates JHT. Estimation of time-varying respiratory mechanical parameters by recursive least
squares. J Appl Physiol 1991;71(3):115965.
[6] Avanzolini G, Barbini P. A comparative evaluation of three on-line identication methods for a respiratory
mechanical model. IEEE Trans Biomed Eng 1985;32(11):95763.
[7] Kanazawa K, Kawata Y, Niki N, Satoh H, Ohmatsu H, Kakinuma R, Kaneko M, Moriyama N, Eguchi K.
Computer-aided diagnosis for pulmonary nodules based on helical CT images. Comput Med Imaging
Graphics 1998;22(2):15767.
[8] Tolias YA, Panas SM. Fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering. IEEE
Trans Med Imaging 1998;17(2):26373.
[9] Lin JS, Cheng KS, Mao CW. Multispectral magnetic resonance images segmentation using fuzzy Hopeld
neural network. Int J Biomed Comput 1996;42(3):20514.
[10] Zouridakis G, Jansen BH, Boutros NN. Fuzzy clustering approach to EP estimation. IEEE Trans Biomed
Eng 1997;44(8):67380.
[11] Manseld JR, Sowa MG, Scarth GB, Somorjai RL, Mantsch HH. Fuzzy c-means clustering and principal
component analysis of time series from near-infrared imaging of forearm ischemia. Comput Med Imaging
Graphics 1997;21(5):299308.
[12] Lorenz A, Bluem M, Ermert H, Senge Th. Comparison of different neuro-fuzzy classication systems for
the detection of prostate cancer in ultrasonic images. In: Proceedings of the 1997 IEEE Ultrasonics
Symposium, vol. 2, Toronto, Canada, 1997, p. 12014.
[13] Siafakas NM, Vermeire P, Pride NB, et al. Optimal assessment and management of chronic obstructive
pulmonary disease (COPD). The European respiratory task force. Eur Respir J 1995;8:13981420.
[14] Gustafson DE, Kessel WC. Fuzzy clustering with a fuzzy covariance matrix. In: Proceedings of the IEEE
CDC, San Diego, CA, USA, 1979, p. 7616.
[15] Bezdek JC. Pattern recognition with fuzzy objective function. New York: Plenum Press, 1981.
[16] Gath I, Geva AB. Unsupervised optimal fuzzy clustering. IEEE Trans Pattern Anal Machine Intell
1989;7:77381.
[17] Pal NR, Bezdek JC. On cluster validity for the fuzzy c-means model. IEEE Trans Fuzzy Syst
1995;3(3):3709.
[18] Krishnapuram R, Freg C-P. Fitting an unknown number of lines and planes to image data through
compatible cluster merging. Pattern Recognition 1992;25(4):385400.
[19] Kaymak U, Babuska R. Compatible cluster merging for fuzzy modeling. In: Proceedings FUZZ-IEEE/
IFES'95, Yokohama, Japan, March, 1995, p. 897904.
[20] Lourens MS, van den Berg B, Hoogsteden HC, Bogaard JM. Flowvolume curves as measurement of
respiratory mechanics during ventilatory support: the effect of the exhalation valve. Intensive Care Med
1999;25:799804.
[21] Aerts JG, van den Berg B, Lourens MS, Bogaard JM. Expiratory owvolume curves in mechanically
ventilated patients with chronic obstructive pulmonary disease. Acta Anaesthesiologica Scandinavica
1999;43:3227.
[22] Morris MJ, Lane DJ. Tidal expiratory ow patterns in airow obstruction. Thorax 1981;36:13542.
[23] Aerts JG, van den Berg B, Bogaard JM. Ventilator-CPAP with the Siemens Servo 900C compared with
continuous ow-CPAP in incubated patients: effect on work of breathing. Anaesth Intensive Care
1997;25:48792.
[24] Morris MJ, Madgwick RG, Collyer I, Denby F, Lane DJ. Analysis of expiratory tidal ow patterns as a
diagnostic tool in airow obstruction. Eur Respir J 1998;12:11137.
R. Babuska et al. / Artificial Intelligence in Medicine 21 (2001) 91105 105