Proceedings of the 29th Annual International
Conference of the IEEE EMBS
Cité Internationale, Lyon, France
August 23-26, 2007.
A Neuro-Fuzzy Controller for the estimation of Tidal Volume and
Respiration Frequency ventilator settings for COPD patients
ventilated in control mode.
A.Tzavaras, P.R.Weller, B. Spyropoulos
Abstract—Patients with chronic obstructive pulmonary
disease (COPD) are characterized by increased work of
breathing (WOB) and ventilatory muscle dysfunction.
Mechanical ventilation is applied to unload the WOB; rest
respiratory muscles decrease arterial partial pressure of
carbon dioxide (PaCO2) and treat hypoxemia. Since patients’
needs are not static, ventilator settings have to be adjusted
regularly. The aim of the present study was the development
and evaluation of a neuro-fuzzy controller, that utilizes non-
invasively acquired parameters for the determination of the
appropriate tidal volume (VT) and respiration frequency (RR)
ventilator settings for COPD patients. Forty three (43) hours of
non-invasively monitored physiology parameters and ventilator
settings, from four (4) different COPD patients ventilated in
control mode, were collected in two (2) General Hospitals in
Greece. Recorded data were randomly allocated into two sets,
namely training set (60%) and evaluation set (40%). A neuro-
fuzzy controller was developed and trained, by employing the
training set. The controller utilizes non-invasively measured
parameters, namely oxygen saturation (SpO2), lung compliance
(C) and resistance (R), Peak Inspiratory pressure (PIP) and
Plateau pressure (Pplateau), for predicting appropriate VT and
RR settings. The developed neuro-fuzzy controller was tested
against evaluation set. The Mean Square Error of the tidal
volume and the respiration rate was 0.222 ml/Kgr and 1.21
breaths per minute (bpm) respectively.
I. INTRODUCTION
M ECHANICAL ventilation support is provided to
critically ill ICU patients who are unable to maintain
gas exchange. ICU clinicians monitor and evaluate cardio-
respiratory related physiology parameters, in order to
evaluate adequacy of mechanical ventilation. Since a
patient’s needs are continuously changing, clinicians have to
adapt the ventilation strategy and drug administration on a
regular basis. This ongoing process is described as
ventilation-respiration management. Patients with chronic
obstructive pulmonary disease (COPD) are characterized by
increased work of breathing (WOB) and ventilatory muscle
dysfunction due to chronic airflow limitation from
Manuscript received April 2, 2007.
A.Tzavaras is with the Medical Instrumentation Technology Department,
Technological Education Institute of Athens, GR 12210 Athens, Greece
(phone: 30-2105385335; fax: 30-2105385302; e-mail: tzavaris@otenet.gr)
and a PhD student in the Centre for Health Informatics, City University.
P.R.Weller is with Centre for Health Informatics, City University
London EC1V OHB, (e-mail:P.R.Weller@soi.city.ac.uk).
B. Spyropoulos is with the Medical Instrumentation Technology
Department, Technological Education Institute of Athens, GR 12210
Athens, Greece (e-mail: basile@teiath.gr).
1-4244-0788-5/07/$20.00 ©2007 IEEE
FrP2D3.7
inflammation, airway hyperactivity, secretions and loss of
the structural integrity of the lung parenchyma [1].
A. Fuzzy Rule Based Systems and Neural Networks
Fuzzy Rule Based Systems (FRBSs), constitute an
extension to classical rule based systems. Fuzzy Sets (FSs),
were introduced by Lofti A. Zadeh in 1965 [2]. Knowledge
representation is performed with the use of linguistic
variables. Fuzzification is the process of interpreting a crisp
value as membership degree to the fuzzy sets which
compose the fuzzy space. The total of linguistic terms and
membership functions (µ) of a FRBS, forms the Data Base
(DB) of the system. The logic of how the system responds to
inputs, is formed as a collection of linguistic rules joined by
the also operator. The total of rules forms the Rule Base
(RB) of the system. In FRBSs such as Mandani [3], where a
fuzzy output is produced, it is common to translate the fuzzy
output to crisp values. This process is called defuzzification.
Fuzzy Logic (FL) demonstrates several advantages over
other methodologies. It can easily model complex systems,
by introducing a development methodology similar to
human communication; experts’ knowledge is encoded
directly in a form very similar to their decision making
process; the RB of a FRBS is evaluated in parallel, thus all
decision determinants are considered in the solution of a
problem; FL models uncertainty and imprecision in complex
models where understanding is limited and/or judgmental.
Neural networks consist of interconnected information
processing units called artificial neurons (Fig 1). The
structure of a neuron consist of external inputs (X1,..Xn),
synapses, dendrites (d), a soma and an axon, which transmits
output to other neurons [4]. Inputs are modified by weights
(Wnj), representing the synaptic junctions. Each synaptic
output is an input to the soma called dendritic input. Each
dendritic input is a transformed version of the external input.
The neuron produces an output when the aggregated
activity of all dendritic inputs exceeds a threshold value. The
neural output is performed with the activation or transfer
function. Activation function varies with the type of neuron
we choose to use. During neural network training an
external input set and corresponding output set is utilized for
adjusting weights and threshold values. The use of training
input – output sets is named supervised training and is the
prevailing methodology.
3765

Fig. 1. Schematic representation of a neuron (Where j is the neuron number
and n is the input number).
B. Synergism of FRBSs and NN.
The development of FRBSs, employed for Supervised
Machine Learning purposes, requires feedback from an
experimental data set and/or from the experts’ advice. The
employment of NN and of Genetic Algorithms (GA)
technologies, in the automation of the development process
of FRBS, based on available input and output training sets,
constitute, both, well established practices.
As we have reported [5], automatic generation of FRBSs
with the use of the GA approach is inferior to the NN in
modeling complex non linear mathematical functions.
Artificial Neural Networks have been extensively used in
fuzzy systems [6]. Neural Networks can be used in
determining membership functions. During the first stage a
NN is used for classification or clustering of domain data,
and during the second stage, fuzzification is used for
assigning fuzzy membership values to clusters.
Tagaki and Hayashi [7], suggested a NN Driven Fuzzy
reasoning methodology. According to this strategy a NN is
trained from a set of input – output data, not based on their
crisp values but rather on the degree of membership to
predefined input – output fuzzy sets.
Fig. 2. Architecture of the neural-fuzzy network proposed by XZ Wang et
al.
Jang & Gulley developed a toolbox for Matlab®, named
Adaptive network based fuzzy inference system (ANFIS),
[8], which is appropriate for learning fuzzy systems. ANFIS
is based on gradient descent optimization with feed forward
NN, for learning single output Tagaki-Sugeno (TSK)
systems. ANFIS constructs a fuzzy inference system (FIS)
whose membership function parameters are tuned (adjusted)
using either a back propagation algorithm alone, or in
combination with a least squares type of method.
3766
X.Z. Wang et al proposed a neural fuzzy network for RB
generation [9]. The Input and Output variable domains are
assigned to Linguistic Variables. Then, the Input–Output
data sets are translated into membership degrees (µn), and
they are being concurrently processed by the NN (Fig. 2).
The NN outputs are membership degrees for the output
variables.
Employment of NN-FRBS in ventilation management has
several examples in bibliography.
Kwok et al in 2003 [10] proposed the use of the Adaptive
Fuzzy Inference System (ANFIS), for the control of the
inspired FiO2. Their model utilized FiO2, PEEP and PaO2 as
inputs to their system.
The ANFIS methodology utilized training data from
different clinical scenarios. The scenario values at each
sampling point were used for training data for the ANFIS
TSK type fuzzy inference system. The training cost function
was the mean-square error between scenario value and
systems output.
Wang et al, in 2006, presented a new version of
Simulation of Patients under Artificial Ventilation (SOPA
Vent) model [11]. An ANFIS algorithm was used on data
from control ventilated ICU patients. In order to identify
model’s inputs, physiology and ventilation parameters were
correlated to dead space; those that scored higher were
incorporated as inputs to the model.
PaCO2, RR, tidal volume, Pinsp and PEEP were identified
as inputs and were utilized for tuning the rule base of the
fuzzy inference engine. The results have shown a good
estimation of dead space.
The same methodology was used for identifying input
parameters for the estimation of tidal volume. Patient’s
weight, PEEP, PIP, and RR were the input parameters for
developing and training with an ANFIS algorithm the
inference engine.
To validate the final model the authors have fed the
ventilator settings, ventilator measurement, demographic
data and blood gas measurements to the model and they
have evaluated the predictions of PaCO2 and PaO2 against
real measurements. Results show a good prediction of
arterial oxygen, but not so realistic estimation of arterial
CO2.
Liu et al, in 2006 [12], proposed a similar approach to
ours for the modeling the setting of the Fraction of Inspired
Oxygen (FiO2) by a neuro-fuzzy hybrid system. They
proposed an iterative rule reduction and tuning process to
produce FRs and maintain models’ error to minimum. The
training and evaluation set they have used, presented a
single patient under BIPAP ventilation mode. The authors
concluded that the model resulted in small errors between
suggested settings of FiO2 and recorded data, and the
extracted rules were interpretable to an ICU clinician.
In relation to the other approaches in ventilation
management, the proposed system was designed for specific
lung pathology, namely COPD, it was trained with real
patient data rather than scenarios/simulation data, and it
utilizes as inputs non-invasively acquired parameters, which
simplifies its future application in a clinical setting.
 The proposed architecture utilizes NN to implement the
rule base of a fuzzy system. Although both fuzzy and neural
approaches possess remarkable properties when employed
individually, their synergism is closer to human intelligence
[6]. The employment of NN in fuzzy systems provides them
the ability to learn, while maintaining all the advantages of
fuzzy systems.
II. METHODOLOGY
Forty three (43) hours of non-invasively monitored
physiology parameters and ventilator settings, from four (4)
different COPD ventilated in control mode, were collected
in two (2) general hospitals, namely Veteran’s General
Hospital of Athens (NIMTS), and University Hospital of
Heraclion, Crete (PEPAGNI).
The ICUs were equipped with Siemens-Draeger
ventilators, which are medibus serial interface enabled.
Furthermore the ICUs are equipped with a central
monitoring station, able to record patients’ parameters trends
for the duration of patient’s stay.
The ethics committees of the PEPAGNI and NIMITS
hospitals have approved the collection of patients;
physiology data. Patients’ data were collected directly from
digital outputs of medical equipment by using certified
medical software [13, 14].
We have designed and developed a MatLab
(Mathworks®) toolbox for automatically generating
Mandani FRBS [3] from available input(s) – output(s) data
in excel format.
The toolbox utilizes the NN toolbox of MatLab. The
toolbox was named FUzzy Neural (FUN) and allows for the
user to define fuzzy systems’ parameters as well as well as
NN settings (Fig. 3). The structure of the toolbox is based on
the architecture proposed by Wang et al [9].
Fig 3 A typical screen of the developed FUN Matlab (Mathworks ®)
toolbox for automatically generating FRBS from available input(s)-
output(s) recorded data.
The training process of the FUN toolbox is as follows:
User selects the columns of the excel data that will be used
as input(s)-output(s) variables to the system, and the
corresponding fuzzy sets domain space. This feature allows
the user to develop FRBSs with a subset of the recorded
3767
data. Membership Functions (MFs) are automatically
generated. The type of MFs is user defined, however the size
and position of the MFs is automatically adjusted to the
number of MFs and the fuzzy domain of each input-output
variable. Excel input(s) and output(s) numeric values are
translated into membership degrees; input-output
membership degrees form the training set for the NN. The
NN presents results as membership degrees for the output
fuzzy domain(s). Finally NN output(s) membership degrees
are compared to Microsoft Excel ® output(s) membership
degrees, for the NN performance evaluation.
FUN toolbox was developed for automatically generating
FRBSs with the synergism of NN, for a broad range of
applications, rather than being specific to the purpose of our
research.
Fig. 4. Input-Output membership functions.
Recorded data were randomly divided into training and
evaluating sets. Sixty percent (60%) of the available data
were used for the training of the NN-FRBS system. The
controller utilizes non-invasively measured parameters,
namely oxygen saturation (SpO2), lung compliance (C) and
resistance (R), Peak Inspiratory pressure (PIP) and Plateau
pressure (Pplateau) as inputs for the system. The outputs are
suggested values for VT and RR ventilator settings. VT was
expressed as ml/Kgr, in order to allow comparison between
different patients.
Each input-output variable domain was assigned with five
Triangular- Trapezoid shape membership functions (Fig.
4).The defuzzification method chosen was the Center of
Area (Centroid).
The NN Architecture is graphically described in Fig. 5.
The applied NN consists of an input and an output layer and
two hidden layers. The number of input-output nodes is
defined by the number of fuzzy sets, while the hidden layers
have 1000 and 100 nodes respectively. Tan-sigmoidal and
log-sigmoidal functions were chosen as appropriate transfer
functions. Supervised learning of the NN is performed by
comparing the membership degrees of the training sets
against the NN output.

Fig. 5. NN architecture (Where IW(n,j) and LW(n,j) is the input (n) and
layer weight respectively, b(j) is the bias and j is the neuron number).
III. EVALUATION AND RESULTS
The NN was trained for 1000 epochs, with the
corresponding input-output membership degrees of the
training set values. The resulted NN, performed with a mean
square error (mse) of 0.0012. The mse refers to the deviation
from the actual membership degrees.
The developed NN-FRBS system was applied to the
evaluation set, in order to identify the deviation of the
suggested ventilator settings from the recorded clinicians’
decisions. The mean square error of the tidal volume and the
respiration rate was 0.222 ml/Kgr and 1.21 bpm
respectively. Large deviations of suggested settings against
recorded settings could be attributed to the failure of evenly
distributing acquired data into training and evaluation sets
due to random allocation. Fluctuations of the systems’
output, suggests that the proposed system is responsive to
fluctuations of the input parameters. Figures 6a and 6b show
an instance of the suggested (solid) and the recorded (dash)
values of VT and RR.
Fig. 6. Top (6a) suggested (solid) vs. recorded (dash) VT. Bottom (6b)
suggested (solid) vs. recorded (dash) RR.
IV. CONCLUDING REMARKS
The developed method and more specifically the outcome
of the Neural- Fuzzy (FUN) MatLab (Mathworks®) toolbox
for automatically generating FRBS from available input(s) –
output(s) data in Microsoft Excel ® format, seems to give
very good results after this first case study that was based on
the data accumulated during forty three hours of non-
invasively monitored physiology parameters and ventilator
settings, from four different COPD patients ventilated in
control mode.
Additionally, there are three more possible near-future
applications that are presently under investigation.
3768
First, the employment of the FUN toolbox for the
development of ventilation advisor systems, for patients
with lung pathologies other than COPD, such as Acute Lung
Injury (ALI) and Acute Respiratory Distress Syndrome
(ARDS). Second, the employment of the system for the
introductory training of ICU medical and nursing personnel.
Finally, the employment of the described tool-box for the
development of a decision supporting system, related to
home ventilation, since it is gradually accepted that is safer
and less costly to treat stable, chronic ventilator-dependent
patients in their homes rather, than in an ICU bed, and that
the patient in home-ventilation has a better quality of life,
enhanced social relationships, and less likelihood of
nosocomial infections.
ACKNOWLEDGMENT
Authors would like to thank the ICU staff of the
Veteran’s General Hospital of Athens (NIMTS), and the
University Hospital of Heraclion, Crete (PEPAGNI), for
their contribution to the presented research.
REFERENCES
[1] D.R.Hess, R.M.Kacmarek, “Essentials of mechanical ventilation”,
2002, 2nd edition, McGraw-Hill companies, ISBN 0-07-135229-5.
[2] E.Cox, “The Fuzzy Systems Handbook”, 1994, AP Professional,
ISBN 0-12-194270-8.
[3] E.H.Mandani, S.Assilian, “An experiment in linguistic synthesis with
a fuzzy logic controller”, 1975, International Journal of Man-Machine
Studies, vol 7, pp 1-13.
[4] P.Picton, “Neural Networks”, 2000, 2nd edition, Palgrave, ISBN 0-
333-80287-X.
[5] A.Tzavaras, P.R.Weller, B.Spyropoulos, “A Comparative Study of the
appropriateness of Neural Networks and Genetic Algorithms for the
Automatic Generation of Fuzzy Rule Based Systems from available
Training Data Sets”, 2007, 11th Int Conf on Cognitive & Neural Syst,
May 16-19, Boston Univ.
[6] L.H.Tsoukalas, R.E.Uhrig, “Fuzzy and Neural approaches in
engineering”, 1997, John Willey & Sons Inc., ISBN 0-471-16003-2.
[7] H.Tagaki, I.Hayashi, “NN-Driven fuzzy reasoning”, 1992,
International Journal of Approx. Reasoning, vol 5 (3), pp191-212.
[8] J.S.Jang, N.Gulley, “Fuzzy Logic toolbox for use with MATLAB”,
1995, The MathWorks Inc.
[9] X.Z. Wang, B.H. Chen, S.H. Yang, C. McGreavy, M.L. Lu, “Fuzzy
Rule Generation from Data for Process Operational Decision
Support”, 1997, Computational Chemical Engineering. 21 pp 661-
666.
[10] H.F.Kwok, D.A.Linkens, M.Mahfouf, G.H.Mills, “Rule-base
derivation for intensive care ventilator control using ANFIS”, 2003,
Artificial Intelligence In Medicine, Vol 29, pp 185-201.
[11] A.Wang, M.Mahfouf, G.H.Mills, “A blood gas hybrid model for
ventilated patients in ICU with new formulations for dead space and
tidal volume”, 2006, Proc 24th IASTED Int. Multi-Conf. Biomedical
Engineering., Innsburg, Austria (519-094 ACTA Press).
[12] F Liu, G.S. Ng, C.Quek, T.F.Loh, “Artificial Ventilation Modeling
using Neuro-Fuzzy Hybrid System“, 2006, Int Joint Conf on Neural
Networks, Vancouver, BC, Canada, Jul 16-21, 2006, pp 5166-5171.
[13] Nortis: http://www.nortis.net
[14] Siemens-Draeger: http://www.Draeger.com