=Paper=
{{Paper
|id=Vol-1867/w14
|storemode=property
|title= A Radial Basis Neural Network Based Agent Module Exploiting ECG Signals to Prevent Heart
Diseases
|pdfUrl=https://ceur-ws.org/Vol-1867/w14.pdf
|volume=Vol-1867
|authors=Salvatore Calcagno,Fabio La Foresta
|dblpUrl=https://dblp.org/rec/conf/woa/CalcagnoF17
}}
== A Radial Basis Neural Network Based Agent Module Exploiting ECG Signals to Prevent Heart
Diseases==
https://ceur-ws.org/Vol-1867/w14.pdf
78
A Radial Basis Neural Network Based Agent
Module Exploiting ECG Signals to Prevent Heart
Diseases
Salvatore Calcagno and Fabio La Foresta
DICEAM, University Mediterranea
89122 Reggio Calabria, Italy
Email: calcagno@unirc.it
fabio.laforesta@unirc.it
Abstract—Today, Electro-Cardiogram (ECG) is considered to the imbalance in sodium and potassium concentration,
the most important diagnostic tool in cardiology, because its the heart muscle behaves like an electric dipole so that it
extremely accuracy to reveal potential pathologic heart activities. can be described by a vector changing its orientation and
In the context of a multi-agent system, where agents provide
to monitor the health of patients in a personalized manner on amplitude when the depolarization/depolarization activities
the bases of different embedded modules, we propose a module take place [15]–[18]. Therefore, during the heart activities,
developed with the aim to prevent possible hearth diseases. It is some abnormalities may occur distorting the signal so that
based on a Radial Basis Neural Network (RBNN) able to analyze many algorithms have been developed in order to analyze
the ECG signals and to evaluate the impact of some specific and interpret the ECG [19]–[22]. The scientific community is
parameters for preventing heart diseases.
Index Terms—ECG, Soft Computing, Multi-agent System, heavily involved in research on two different aspects: the first
Radial Basis Neural Network, Cardiac Diseases. one deals with morphological aspects that affect the cardiac
cycle while the second one deals with the timing of events and
I. I NTRODUCTION variations in patters over many beats. Obviously, to investigate
and interpret cardiac pathophysiology, we need of models
Nowadays, progresses in computer science, communication
easily implementable [23]–[25] because it is impossible to
and sensor technologies, as well as in the medical researches,
distort the cardiac activity for studying the effect on ECGs2 .
have made possible to monitor many aspects of human health.
Moreover, these models help us to mathematically characterize
In this context, a prominent role is played by the agent
cardiac activity by simulating possibly potentially fatal cardiac
technology 1 which has been applied with success in this
diseases. Taking into account the above considerations, the
field (see for example [8]–[14]). In such a scenario, we are
proposed Agent Module was based on a parametric dynamic
studying to realize a multi-agent system where each personal
model for simulating ECG signals relative to single-channel
software agent is associated with a patient and collaborates
ECGs. Furthermore, with the aim to obtain a flexible model
with a central agency devoted to support the first aid activ-
able to give trustworthy results a Computational Intelligence
ities suggesting to the health care operators how to use at
approach has been adopted. As a matter of fact, during
the best the human and technical resources. More in detail,
the last decades Computational Intelligence (Artificial Neural
we are supposing to realize the software agent components
Networks, Fuzzy Logic, Support Vector Regression Machines,
with a modular approach so that the agent, on the basis
genetic algorithms, particle swarm techniques) have been
of the different modules that it embeds, can monitor differ-
employed with success to solve a wide variety of problems
ent pathologies (e.g., cardiac troubles, diabetes and so on).
ranging from microwave engineering [26], [27], [28], non
Specifically, in this paper, we present a software prototype
destructive testing evaluation (NDT) [29], [30], transportation
of the module monitoring the cardiac activity that should be
engineering [31], to biomedical engineering [32]. Among the
embedded into a personal software agent. In particular, this
various methodologies available in this field, in the present
module exploits the past studies on the Electrocardiogram
work, an approach based on the exploitation of a Radial Basis
(ECG). ECG, discovered by Waller and Einthoven in 1903, and
Neural Network (RBNN), is proposed. The main reason of this
consisting of a not invasive reproduction of the heart electrical
choice is due to the the ability of the RBNN to handle pattern
activity in relation with the body surface potentials, is the
estimation problems [32]–[35]. More in detail, by means of
most exploited instrument in heart diagnostics especially to
a RBNN, very useful for pattern estimation problems [26],
investigate fibrillation, ischemia, arrhythmia and many other
[32]–[34], a parameterized model for evaluating the impact
heart diseases. During the heart ventricular contractions, due
on the ECG trends is built. Starting from the evaluation of
1 The interested reader might refer to an overwhelming number of surveys
on the matter among which [1]–[7]. 2 Even though the problem could be avoid by studying the ECG database.
79
TABLE I: Empirical Parameters
P Q R S T
time (sec) -0.25 -0.025 0 0.025 0.25
φj (rad) −π/3 −π/12 0 π/12 π/3
hj 1.25 -5 30 -8 1
kj 0.25 0.1 0.1 0.1 0.5
the RBNN performances by means of the variation of its
spread factor related to each feature carried out by the signal,
the spread factor taking into account the trade-off between
performances and neural net complexity is chosen. At the
end, a suitable neural net based on radial basic function is
trained testing the performances of the model. Finally, the
trained RBNN represents the module exploited by each agent
to monitor the cardiac activity of its associated patient. The
paper is organized as follows: Section II gives the basics on
the proposed model for reproducing the ECG signal. Section
III presents the experimental dataset used to build our RBNN
Fig. 1: Trend of an ECG in a healthy patient.
model. Section IV describes the RBNN based Agent module.
Section V describes the proposed multi agent architecture for
health care. Finally, conclusions are given in section VI. in which ω is a parameter related to the heart rate while ξ =
1 − (H 2 + σ 2 )1/2 obtaining φ = tan−1 {2(σH)}. From a
II. T HE P ROPOSED ECG M ODEL
geometric point of view, fixing the angular velocity ω, φ can
ECG is the graphic reproduction of the electrical activity of be computed by moving amount the H − σ limit line.
the heart during its operation, recorded on the surface of the In addition, to simulate cardiac diseases, it is sufficient in (1)
body. to set differently the parameters (hj , kj , φj ) at the expense
Pulses in the myocardium generate potential differences of the model flexibility. In such a context, the proposed
time and space-varying that can be recorded by electrodes approach parametrizes the H − σ limit line modifying (2)
on the surface of the human body. ECG, with a characteristic adapting the model to the PQRST complex for a timely control
trend only varied in the presence of problems, is characterized generalizing the limit path elliptically by both rotation and
by several positive and negative traits called ”waves” which linear transformation as follows:
repeat at each heart cycle (Figure 1). P-wave, small in size,
corresponds to the atrial depolarization. The QRS complex, a H̄ cosΦ sinΦ a 0 H
= (3)
set of three consecutive waves, corresponds to ventricular de- σ̄ −sinΦ cosΦ 0 b σ
polarization. T-wave represents the ventricular re-polarization which allows to compute φ as tan−1 {2(σ̄ H̄)}. It is worth
[5]. watching that (2) and (3) are equivalent if a = b = 1 and
Globally, the distortion of PQRST complex denotes heart Φ = 0. The advantage of the proposed procedure liens in the
diseases so that the analysis of ECG reveals the presence of fact that, maintaining the same computational complexity, it
major cardiac pathologies. is possible to characterize the distortion of the complex by
As showed in [5], ECG signal can be carried out by solving means of (a, b, Φ).
a system of three coupled ordinary differential equations. As an example of ECG alteration we show in Figure (2)
Formally, an ECG signal f (t), indicating the module 2π some modification of ECG by setting different values of
operation and the empirical parameters by means of | |2π parameters according to the proposed model.
and (hj , kj , φj ) respectively (Table I), can be reproduced as
III. T HE E XPERIMENTAL DATASET
follows:
∆φ2
With the goal of creating a neural learning model, it is
X − j
imperative to create a training database. So, by using the
f˙(t) = hj |φ − φj |2π e 2k2
j − f (t) (1)
model developed and proposed in the previous Section, we
j∈{P,Q,R,S,T }
have built an artificial database of almost 800 ECG signals.
where φ is an angular dynamic parameter which determines In particular, if f is the sampling frequency, we fixed the
the ECG deflection. In order to evaluate φ, in [5] (1) was time step f −1 = ∆t so that we could apply a 4- Runge-
coupled with the dynamic system: Kutta procedure [36], [37] to obtain a good number of ECG
( signals whose peculiar characteristics, according to [23], [38],
Ḣ = ξH − ωσ [39], are listed in Table II. In addition, the time intervals have
(2)
σ̇ = ξH + ωσ been computed between two consecutive beats (the so-called
80
TABLE II: Values assigned to parameters for the construction
of the artificial ECG signals database.
Parameter Amplitude
ECG f 256
Internal sampling frequency 512
Mean heart rate 70
Number of beats 256
Low frequency/High frequency (LF/HF) ratio 0.6
Std of heart rate 1.2 beat/minute
a (no inconsistent ECG) ≥ 0.52
b (no inconsistent ECG) ≤1
rotation −π/2 ≤ Φ ≤ π/2
mens of learning procedures, are able to create the nonlinear
and multi-variable inputs-outputs mapping [29], [31].
Specifically, RBNNs, considered as excellent functions ap-
proximators, are usually constituted by means of three layers
of nodes. As shown in Fig. 3, starting from left side (inputs
layer) to right one (outputs layer), just one hidden layer,
Fig. 2: An example of ECG signal carried out by the proposed where radial basic functions are exploited, takes place. The
model. The continuous line is referred to a normal ECG; the parameters showed in Fig. 3 are detailed in Table III.
other lines are related to two pathologies whose parameters In addition, as displayed in Fig. 3, the input vector p and the
are (1, 0.7, π/4) and (1, π/6, 0.1) respectively. matrix IW1,1 are exploited respectively as input and weight
matrix for ||dist|| box producing a vector, with cardinality is
N 1 , whose elements are the distances computed between p
RR-interval); the amplitude of RS-intervals was considered and the rows of IW1,1 .
in terms of Volts. Finally, we have also considered the QT- The input for the radial basic function is the vector distance
intervals. computed as before mentioned multiplied by the bias (element
In order to test the performance of the proposed model, allowing the sensitivity of the radbas neuron to be tuned).
another database formed by 200 ECG signals has been carried It is worth to observe that the radial basic function (in our
out by means the same procedure as above specified. 2
case, e−n ), for input equal to zero, is the maximum value
Finally, it is worth to emphasize that the training system equal to unity and, in addition, if the output decreases, ||dist||
consists of three inputs, (a, b, Φ), and three outputs, (RR- decreases: in this way a radbas neuron works as a detector
intervals, RS-amplitude, QT-intervals). The choice of con- producing unity if p is equal to its weight vector.
sidering the RR-intervals as an output of the procedure is In this work, if we consider the Spread Factor SF, which
motivated by the fact that they provide peculiar information represents a sort of measurement of the function smoothing,
on the physiological state of the patient characterizing both the performance of RBFNN depends on the variation of SF
bradycardia and tachycardia. Since the ratio of the power so that the larger is SF, the smoother is the function approx-
in both the low and high frequencies components can be imation. However, SF, since it is too large, could generate
considered for the sympathovagal balance [10]. In this paper, numerical problems so that it is necessary to guarantee the
this ratio is defined as [0.015, 0.15]/[0.15, 0.4]. overlapping among active input regions of the radbas neurons
Regarding the choice of considering the RS-amplitudes as ensuring that a lot of radbas neurons have large outputs at any
input, it is motivated because it measures the voltage gap of the instant.
second section of QRS-complex which causes the ventricles This procedure is able to simultaneously guarantee both a
depolarization. In addition, RR-intervals and RS-amplitude are function smoother and a good generalization of the network.
strongly correlated: if RR-interval increases, then the path has About the setting of the weight for the first layer, see Table IV.
more time to be pushed into the sequence R and S waves. Regarding the second-layer, indicating by LW2,1 and b2 the
Finally, QT-intervals cannot be missing among the inputs second-layer weights and the bias respectively, LW2,1 is
because they represent the time-gaps between the ventricular achieved by solving the following linear formulation
depolarizations and the end of ventricular repolarizations: if
QT-intervals are too large the death of the patient is incoming. A1
T = LW2,1 b2 (4)
I
IV. T HE RBNN BASED AGENT M ODULE TO PREVENT
where A1 , I and T indicate the output of the first-layer, the
HEART DISEASES .
identity matrix and the output matrix respectively.
In the Computational Intelligence context, the Artificial In order to evaluate the performance of the proposed RBNN
Neural Networks (NNs) are consolidated systems that, by model, a set of standard statistic indexes have been taken
81
Fig. 3: Simplified representation of the RBNN.
into account. In particular, as showed in (5), Root mean TABLE III: Parameters of the RBNN model.
Squared Error (RMSE), Root Relative Squared Error (RRSE), Parameter Specification
Mean Absolute Error (MAE), Relative Absolute Error (RAE) M cardinality of the input vector
and Willmott’s Index of Agreement (WIA) [40] have been N1 cardinality of the set of neurons related to layer
considered to evaluate the performance by varying the spread 1
N2 cardinality of the set of neurons related to layer
from 0.25 to 20 with a a step equal to 0.25. 2
P 1 Radial Basic Function activation function of N 1
1 m 2 2 (radbas)
RM SE = j=1 (yj − ȳj ) ;
m Linear Function activation function of N 2
Pm 2 1
(ȳj −yj ) 2 (purelin)
RRSE = Pj=1 m ;
Pm j=1 (yj −ȳ)
2
|ȳ −y |
M AE = j=1 m j j ; TABLE IV: Setting of weights and biases
Pm
Pj=1
|ȳj −yj | (5)
RAE = m
|yj −ȳ| ;
j=1 Pm Specification
2
j=1 (yj −ȳj )
W IA = 1 − Pm (|y 2;
first-layer weights transpose of input matrix
j −ȳ|+|ȳj −ȳ|)
first-layer biases cardinality of the set of neurons related to layer
j=1
ȳ=predicted feature 1
radial basic function cross 0.5 at weighted inputs of ±SF
m=cardinality of the testing patters.
The proposed procedure has been implemented on Intel Core
2 CPU 1.47 GHz using MatLab R2013a selecting an SF value user. Moreover, each personal agent is thought to be a light
equal to 0.77 carried out by means a trade-off between the software component able to run on a personal devices, in
complexity of the net and the same performances; for this order to monitor its user throughout his/her daily activities.
computation, the time elapsed was almost equal to 50 minutes. Nowadays, it is possible because mobile devices have suitable
The training procedure has been carried out for a 29-hidden- computational, storage and communication capabilities to host
neurons RBNN whose time elapsed was equal to 0.96 seconds. and support such personal agents in all their tasks. At the same
The obtained results have been collected in Fig. 4, where time, more and more powerful sensors, provided of wireless
RMSE, for each feature, is displayed vs the SF variation and communication capabilities, are currently available. More in
in Table V where the comparison among the values of the detail, the agent gathers the data coming via wireless from
statistical parameters is evident putting out a good quality the sensors both directly applied on the associated user (e.g.,
of the tuning achieved by means of the proposed RBNN
procedure.
V. T HE H EALTH C ARE M ULTIAGENT A RCHITECTURE TABLE V: Performance of the RBNN exploiting statistical
evaluators
The multi-agent system that we are designing consists of
a central Agency and a set of personal software agents, Inputs RMSE RRSE MAE RAE WIA
each one associated with a user (i.e., patient) to monitor. In RR-intervals 0.002 0.37 0.004 0.31 0.94
particular, each personal agent can be configured with specific RS-amplitudes 0.03 0.239 0.031 0.196 0.891
QT-intervals 0.006 0.324 0.005 0.387 0.922
modules on the basis of the pathologies affecting its associated
82
Fig. 4: Performance of the model in terms of RMSE for each input vs the spread.
ECG or glycemia sensors) and into his/her environment where VI. C ONCLUSIONS
he/she lives (e.g., movement sensors). These data acquired by In this paper we have presented a study to realize a module
the agent are analyzed by the corresponding agent module in for monitoring patients with heart troubles, as a part of a
order to evaluate the health state of the monitored user. When software agent equipment in the context of a more complex
the agent detects a potential health risk then it provides to and ambitious multi-agent system oriented to monitor the
inform its Agency. The aim of the Agency is that of supporting patients’ health.
the assigned operators in managing the first aid activities by To realize this module, in order both to solve complex
suggesting the activities to carry out and the best use of the regression problems and develop an alert system, a RBNN
available human and technical resources (i.e., medical doctors, architecture has been exploited. In such a context, starting
health-care assistants, ambulances and so on). from both a biomedical problem where the estimation of
features for ECGs recording is required and a self-modified
A. Preventing Heart Disease dynamic model, the RBNN performances have been estimated
to put out pathological ECGs when heart diseases appear.
In the aforementioned multi-agent scenario, we describe the
The proposed agent module has provided reliable results
case of a patient affected by heart troubles that is monitored
to evaluate the parameters of McSharry’s modified model
by a personal software agent running on his/her mobile device
carrying out pathological events in ECG.
(e.g., a smartphone), equipped with a heart module embedding
As future works, we are considering the advantages of
the trained RBFANN previously described. The agent provides
including in the design of the proposed system (i) a control
to analyze the data coming from the patient’s ECG sensors
on the reliability of the sensor data in order to avoid the risks
and required in input by the RBFANN. If the agent detects
due to wrong or missed alarms and, to this purpose, trust and
the symptoms of a potential health diseases then it provides
reputation system appear as potential candidates for realizing
to alert both its patient and its remote Agency by exploiting
it also on the basis of recent models and applications [41]–
the device communication capabilities.
[43] and (ii) an evolutionary strategy [44], [45] for improving
When the Agency receives the agent alert then it, on the
in time the Agency performance.
basis of both the potential heart risk and the health resources
available at that time, provides a list of suggestions to the R EFERENCES
designated operators in order to help at the best the unfortunate [1] C. A. Iglesias, M. Garijo, and J. C. González, “A survey of agent-
user. oriented methodologies,” in International Workshop on Agent Theories,
83
Architectures, and Languages. Springer, 1998, pp. 317–330. Signal,” EURASIP Journal on Advances in Signal Processing, vol.
[2] M. Pĕchouček and V. Mařı́k, “Industrial deployment of multi-agent 43407, pp. 1–14, 2007.
technologies: review and selected case studies,” Autonomous Agents and [25] B. Azzerboni et al., “ Spatio-temporal analysis of surface Electromyo-
Multi-Agent Systems, vol. 17, no. 3, pp. 397–431, 2008. graphy signals by Independent Component and Time-Scale Analysis,”
[3] W. Shen, L. Wang, and Q. Hao, “Agent-based distributed manufactur- Proceedings of The Second Joint Meeting of the IEEE Engineering in
ing process planning and scheduling: a state-of-the-art survey,” IEEE Medicine and Biology and Biomedical Engineering Society, vol. 1, pp.
Transactions on Systems, Man, and Cybernetics, Part C (Applications 112–113, 2002.
and Reviews), vol. 36, no. 4, pp. 563–577, 2006. [26] G. Angiulli, “Design of Square Substrate Waveguide Cavity Resonators:
[4] Z. Zhou, W. K. V. Chan, and J. H. Chow, “Agent-based simulation of Compensation of Modelling Errors by Support Vector Regression Ma-
electricity markets: a survey of tools,” Artificial Intelligence Review, chines,” American Journal of Applied Sciences, vol. 9, no. 11, pp. 1872–
vol. 28, no. 4, pp. 305–342, 2007. 1875, 2012.
[5] M. Metzger and G. Polakow, “A survey on applications of agent tech- [27] G. Amendola, G. Angiulli, E. Arnieri, L. Boccia, and D. D. Carlo,
nology in industrial process control,” IEEE Transactions on Industrial “Characterization of lossy SIW resonators based on multilayer per-
Informatics, vol. 7, no. 4, pp. 570–581, 2011. ceptron neural networks on graphics processing unit,” Progress in
[6] A. L. Bazzan and F. Klügl, “A review on agent-based technology for Electromagnetics Research C, vol. 42, pp. 1–11, 2013.
traffic and transportation,” The Knowledge Engineering Review, vol. 29, [28] G. Angiulli and M. Versaci, “A Neuro-Fuzzy Network for the design
no. 03, pp. 375–403, 2014. of circular and triangular equilateral microstrip antennas,” International
[7] M. N. Postorino and G. M. L. Sarné, “Agents meet traffic simulation, Journal of Infrared and Millimeter Waves, vol. 23, no. 10, pp. 1513–
control and management: A review of selected recent contributions,” 1520, 2002.
in Proceedings of the 17th Workshop “from Objects to Agents”, WOA [29] M. Cacciola, F. L. Foresta, F. Morabito, and M. Versaci, “Advanced use
2016, ser. CEUR Workshop Proceedings, vol. 1664. CEUR-WS.org, of soft computing and eddy current test to evaluate mechanical integrity
2016. of metallic plates,” NDT and E International, vol. 40, no. 2, pp. 357–
[8] J. Nealon and A. Moreno, “Agent-based applications in health care,” in 362, 2007.
Applications of software agent technology in the health care domain. [30] M. Cacciola, F. Morabito, D. Polimeni, and M. Versaci, “Fuzzy charac-
Springer, 2003, pp. 3–18. terization of flawed metallic plates with eddy current tests,” Progress in
Electromagnetics Research, vol. 72, pp. 241–252, 2007.
[9] J. M. Corchado, J. Bajo, Y. De Paz, and D. I. Tapia, “Intelligent
[31] M. Postorino and M. Versaci, “A neuro-fuzzy approach to simulate
environment for monitoring alzheimer patients, agent technology for
the user mode choice behaviour in a travel decision framework,”
health care,” Decision Support Systems, vol. 44, no. 2, pp. 382–396,
International Journal of Modelling and Simulation, vol. 28, no. 1, pp.
2008.
64–71, 2008.
[10] V. Chan, P. Ray, and N. Parameswaran, “Mobile e-health monitoring: an
[32] N. Mammone et al., “Clustering of Entropy Topography in Epileptic
agent-based approach,” IET communications, vol. 2, no. 2, pp. 223–230,
Electroencephalography,” Neural Computing & Applications, vol. 20,
2008.
no. 6, pp. 825–833, 2011.
[11] M. T. Nguyen, P. Fuhrer, and J. Pasquier-Rocha, “Enhancing e-health [33] M. Buhmann, Radial Basis Functions: Theory and Implementations.
information systems with agent technology,” International journal of Cambridge University Press, 2003.
telemedicine and applications, vol. 2009, p. 1, 2009. [34] S.Chen, C. Cowan, and P. Grant, “Ortogonal Least Squares Learning
[12] F. Bergenti and A. Poggi, “Multi-agent systems for e-health: Recent Algorithm for Radial Basis Function Network,” IEEE Transaction on
projects and initiatives,” in 10th Int. Workshop on Objects and Agents, Neural Networks, vol. 2, no. 2, pp. 302–309, 1991.
2009. [35] F. Morabito et al., “Deep learning representation from electroen-
[13] S. Iqbal, W. Altaf, M. Aslam, W. Mahmood, and M. U. G. Khan, cephalography of early-stage creutzfeldt-jakob disease and features for
“Application of intelligent agents in health-care: review,” Artificial differentiation from rapidly progressive dementia,” International Journal
Intelligence Review, vol. 46, no. 1, pp. 83–112, 2016. of Neural Systems, vol. 27, no. 2, 2017.
[14] S. Montagna, A. Omicini, F. Degli Angeli, and M. Donati, “Towards [36] J. Lambert, Numerical Methods for Ordinary Differential Systems. John
the adoption of agent-based modelling and simulation in mobile health Wiley and Sons, 1991.
systems for the self-management of chronic diseases,” vol. 1664, 2016. [37] D. Griffiths and D. Higham, Numerical Methods for Ordinary Differen-
[15] J. Malmivuo and R. Plonsey, Bioelectromagnetism, Principles and tial Equations. JSpringer-Verlag London, 2010.
Application of Bioelecric and Biomagnetic Fields. New York: Oxford [38] M. Malik, “Heart Variability,” European Heart Journal, vol. 17, pp.
University Press, 1995. 354–381, 1996.
[16] O. Dossel, “Inverse Problem of Electro- and Magnetocardiography: [39] P. Schwartz and S. Wolf, “QT Interval as Predictor of Sudden Death
Review and Recent Progress,” International Journal of Bioelecromag- in Patients with Myocardial Infarction,” Circulation, vol. 57, no. 6, pp.
netism, vol. 2, no. 2, 2000. 1074–1077, 1978.
[17] A. van Oosterom, Beyond the Dipole, Modeling the Genesis of the [40] L.Wasserman, All of Statistics. Springer-Verlag, 2004.
Electrocardiogram. New York: 100 Years Einthoven, 2002. [41] D. Rosaci, G. M. L. Sarné, and S. Garruzzo, “Integrating trust measures
[18] D. Geselowitz, “Use of Time Integrals of the ECG to Solve the Inverse in multiagent systems,” International Journal of Intelligent Systems,
Problem,” IEEE Transactions on Biomedical Engineering, vol. 31, pp. vol. 27, no. 1, pp. 1–15, 2012.
73–75, 1985. [42] M. N. Postorino and G. M. L. Sarné, “An agent-based sensor grid
[19] G. Clifford, F. Azuaje, and P. McSharry, Advanced Methods and Tools to monitor urban traffic,” in Proceedings of the 15th Workshop from
for ECG Data Analysis. London, UK: Artech House Publishers, 2006. “Objects to Agents”, WOA 2014, ser. CEUR Workshop Proceedings,
[20] G. Clifford et al., “ An artificial vector model for generating abnormal vol. 1260. CEUR-WS.org, 2014.
electrocardiographic rhythms,” Physiol. Meas., vol. 31, pp. 595–609, [43] D. Rosaci, G. M. L. Sarné, and S. Garruzzo, “TRR: An integrated
2010. reliability-reputation model for agent societies,” in Proceedings of the
[21] B. Azzerboni et al., “ PCA and ICA for the Extraction of EEG Dominant 12th Workshop from “Objects to Agents”, WOA 2014, ser. CEUR
Components in Cerebral Death Assessment,” Proceedings of The 2005 Workshop Proceedings, vol. 741. CEUR-WS.org, 2011.
International Joint Conference on Neural Networks, vol. 4, no. 1556301, [44] S. M. Manson, “Bounded rationality in agent-based models: experiments
pp. 2532–2537, 2005. with evolutionary programs,” International Journal of Geographical
[22] F. L. Foresta et al., “ PCA-ICA for automatic identification of critical Information Science, vol. 20, no. 9, pp. 991–1012, 2006.
events in continuous coma-EEG monitoring,” Biomedical Signal Pro- [45] D. Rosaci and G. M. L. Sarné, “Cloning mechanisms to improve agent
cessing and Control, vol. 4, pp. 229–235, 2009. performances,” Journal of Network and Computer Applications, vol. 36,
[23] P. McSharry, G. Clifford, L. Tarassenko, and L. Smith, “A Dynami- no. 1, pp. 402–408, 2013.
cal Model for Generating Synthetic Electrocardiogram Signals,” IEEE
Transactions on Biomedical Engineering, vol. 50, no. 3, pp. 289–294,
2003.
[24] R. Sameni, C. Clifford, C. Jutten, and M. Shamsollahi, “Multichannel
ECG and Noise Modeling: Application to Maternal and Fetal ECG