Mobile Agents using Data mining for Diagnosis Support in Ubiquitous Healthcare

John

Abstract.  Recent  research  topics  in  healthcare  including  intelligent  decision support  services,  expert  medical  services  and  autonomous  management  are based  on  multi-agent  systems.  The  cooperation  of  these  software  agents  provides efficient monitoring, analyzing, and managing the data  of patient  where abnormal patterns are detected to have an advance treatment and prevent loss of life. In this paper, a framework for ubiquitous healthcare based on multi-agent is presented. This paper proposes a mobile agent for diagnosis support in ubiquitous healthcare. The expert mobile agent (EMA) classifies the data of patient by  using  neuro-fuzzy  algorithm  for  consultation  report.  A  pre-processing method based on the profile of an expert is used to filter the data from the history of patient.  Result of  neuro-fuzzy from cross-validation  test shows a high accurate classification in data compared to other highly accurate classifiers.

1  Introduction
Agent-based  healthcare  system  addresses  the  importance  of  intelligent  programs  to
substitute  the  real  person’s  functions  in  healthcare  services  and  management.  This
technique benefits most of individuals through their decision making and automation
of their tasks. Most of these implementations are used for decision support system [1].
The use of agent-based intelligent support systems is important in medical industries
because  it  allows  doctors  and  nurses  gather  quick  information.  This  information  is
processed  in  various  ways  to  assist  with  making  diagnosis  and  treatment  decision.
Also, because software agents deals with distributed systems, it assist in diversity of
storing and retrieving medical records, analysis of real-time data  gathered  and  other
necessary information retrieval in distributed environment.  
Techniques and algorithms are integrated to the agent-based healthcare system for
medical diagnosis. Neural network is a common technique for medical diagnosis [2, 3].
Successful application examples show that neural diagnostic  systems are  better than
human diagnostic capabilities. Moreover, neural network are used to analyze medical
images  [4,  5].  These  research  articles  survey  various  approaches  and  techniques  to
improved diagnosis in medical images, including mammography, ultrasound and mag-

* This work is supported by the Korea Research Foundation Grant funded by the Korean
   Government (MOEHRD) (KRF-2006-521-D00372).
netic resonance imaging. Neural network-based agents are used for discovering rules
in medical database [6]. Medical databases, which consist of patient histories, special-
ist's  conclusions,  laboratory  results,  etc.,  are  typically  distributed  set  of  semi-
structured data, and because agent technology is well-suited approach to develop the
medicine decision supporting systems, the integration of neural network in the agent is
necessary. Medical databases are dynamically changed because the structures of fea-
tures  characterizing  the  diseases  are  continuously updated.  The  features  of  diseases
depend on tools and technologies those doctors and specialists currently use to diag-
nose  and  treat  the  patients.  Even  though  the  integration  of  neural  networks  within
agents is well-researched, it still needs intensive research on using hybrid systems [20]
because of the vast changes of information and the classical methods may not solve
the problem of classification.
In this paper, we propose an expert mobile agent using data mining to support the
diagnosis of the  patient in ubiquitous  healthcare. Moreover,  a  framework of ubiqui-
tous healthcare based on multi-agent is presented. The framework supports the mobil-
ity of the mobile agent which executes classification algorithm to the data of patient.
The paper investigated efficient classifiers on data  mining to integrate with the pro-
posed  expert  mobile  agent.  The  proposed  expert  mobile  agent  (EMA)  uses  neuro-
fuzzy  classification  for  consultation  of  patient.  On  first  phase,  the  fuzzy  system  of
EMA  is  trained  from  the  previous  data  of  other  patients.  A  pre-processing  method
based on the profile of an expert is used to filter the relevant data from its expertise.
After the training, the EMA are deployed to execute classification of data. Result from
cross-validation  test  shows  that the neuro-fuzzy classification provides  a  high  accu-
racy in classifying the data compared to other highly accurate classifiers.
2  Related Works
Intelligent agent for healthcare plays a crucial role on giving correct information for
diagnosis and  providing immediate medical services. Home healthcare  services pro-
vide information to a consumer of the necessary diagnosis and continuous monitoring
of patient to acquire immediate response and  save lives  in case of abnormal indica-
tions.  Agent-based  intelligent  decision  support  is  proposed  for  the  home  healthcare
environment [7]. The multi-agent platform is combined with artificial neural network
for the intelligent decision support system in a group of medical specialists collaborat-
ing in the pervasive management of care for a patient. Mobile agents are used to serve
the collaboration of services for mobile users [8]. An agent is an autonomous, social,
reactive and proactive entity, sometimes also mobile. Since telemedicine is grounded
on  communication and sharing of resources,  agents  are  suitable  for  its  analysis  and
implementation, and these are adopted for developing a prototype telemedical agent.
Data mining aims to extract interesting information from large databases is used for
decision  support  in  the  field  of  medicine.  In  order  to  have  mobility,  data  mining
framework for mobile environment are proposed by researchers [9, 10, 11]. A context-
awareness on data mining is used to maximize the adaptive capacity of data mining [9].
The  use  of  decision  support  PDA  supported  by  data  mining  facility can  be  a  great



asset to the medical professionals while working on an emergency or while rushing to
attend an emergency. Data mining in mobile environment using mobile agents is found
in the work of Lee, et. al. [10]. This is done by sending a mobile agent to the LBS and
then it performs the classification mining in  the  database.  In  the  HCARD model  of
Gerardo, et. al. [11], proposed an Integrator agent to perform knowledge discovery in
the heterogeneous server in the distributed environment. Data mining are essential in
extracting rules from databases and provide decision support knowledge in healthcare
environment.
2.1  Neuro-fuzzy Classification
Fuzzy systems are used to handle uncertainty from the data that cannot be handled by
classical methods. It  uses the fuzzy set to  represent a  suitable  mathematical tool for
modeling of imprecision and vagueness [12]. The pattern classification of fuzzy clas-
sifiers  provides  a  means  to  extract  fuzzy  rules  for  information  mining  that  leads  to
comprehensible method  for  knowledge extraction from various information sources.
The  fuzzy algorithm is also a popular  tool for  information retrieval.  Fuzzy c-means
classifier (FCM) uses an iterative procedure that starts with an initial random alloca-
tion of the objects to be classified to c clusters. Neuro-fuzzy systems are the hybrid of
artificial neural networks and fuzzy systems. The algorithm borrows the learning abil-
ity of neural networks to determine the membership values. It is among the most popu-
lar data mining techniques used in recent research [13, 14]. There are many types of
neuro-fuzzy  rule  generation  algorithm  [15].  FuNE-I  is  a  neuro-fuzzy  model  that  is
based on the architecture of feed-forward neural network with five layers which uses
only rules with one or two variables in antecedents [16]. A Sugeno-Type neuro-fuzzy
system is used for a scheme to construct an n-link robot manipulator to achieve high-
precision position tracking [17]. A neuro-fuzzy classification (NEFCLASS) is a fuzzy
classifier that creates fuzzy rule from data by a single run through the data set [14].



Fig 1. A NEFCLASS system with two inputs, five rules and two output classes



3  Framework of Ubiquitous Healthcare based on Multi-agents
In  this study,  we propose  a  framework  for  the  ubiquitous  healthcare.  The  proposed
framework consists of multi-agents managing the hospital shown in Figure 2. A ubiq-
uitous healthcare in [18] proposes a method of accessing healthcare services by indi-
vidual consumers applying to mobile computing device. The OnkoNet Mobile Agents
Architecture  was  developed  and  consists  of  cooperation  protocols,  inference  model
and  health  ontology  to  provide  efficient  ubiquitous  healthcare  environment.  In  our
framework, mobility support for expert mobile agent (EMA) and data mining support
for the ubiquitous healthcare are considered. Figure 2 illustrates the proposed architec-
ture based on multi-agent system. Each doctors and specialist have their own EMA. In
Figure 2, there  are three different rooms consists of monitor agent (MA) to monitor
the readings of the sensors in the patient and triggers the room manager (RM) to com-
municate  with  the  hospital  manager  (HM)  for  necessary  diagnosis  or  actions  to  be
taken by the doctors. The movement of EMA from room 1 to room 2 requires com-
munication with RM  to  initialize the interaction  to  the agents inside  the  room.  This
procedure considers verification of accessing data for security.



Fig. 2. Framework of ubiquitous healthcare based on multi-agent
3.1  Components of the Multi-agents
Hospital Manager The framework of ubiquitous healthcare in Figure 2 is consists of
multi-agents to have an efficient service through the ubiquitous healthcare system. The
main software  agent in  the framework is the hospital  manager (HM). It  concerns  in
managing the services in the hospital supporting the decision making and management.
Moreover,  it  communicates to  other  agent  component through facilitator  agent. The
deployments of the services in rooms are done by the hospital manager.


Mobile
Agent
Mobile
Agent
RM
RM


Room 1
Room 2
Room 3
RM



Patients
Patients
Patients
Hospital
Manager



MA
MA
MA
Facilitator
Agent



Services
Hospital Admin



Facilitator Agent The main function of the facilitator agent (FA) is a broker between
the  HM  and  room  manager  (RM).  All  negotiations  of  requesting  services  from  the
room manager are done with the FA before the HM deploys its service in the room.
The  confirmation  of  deploying  the  services  are  received  by  FA  and  the  RM  is  in-
formed if the request of service is possible or not. Also, FA receives the alert message
from the room manager if there are needs of attention with the patient.

Room Manager The framework of the ubiquitous healthcare is consisted of physical
room for the patients shown in Figure 2. A room manager (RM) coordinates the task
of agents within the room. The software agents communicate to RM for every event
that needs attention of the individuals inside the hospital. RM also request for services
needed by the patients to  hospital manager via FA. After the negotiation of FA, the
service is deployed and adds to the RM.

Monitor Agent Healthcare sensors and other sensors used for monitoring the patient
are  handled  by  monitor  agents.  These  are  programmed  to  detect  abnormal  patterns
from readings of the patient. This study assumes that these sensors are used for moni-
toring of patient and send the signal for analysis.

Service  Modules In our proposed  system, the services modules are used  to  support
the diagnosis of patient, decision making and management of the hospital. These are
managed by the HM. FA negotiates the request from the RM before HM deploys the
service in the RM.

Expert Mobile Agent The expert mobile agent (EMA) uses the proposed framework.
Doctors and specialist uses their PDA as the host of EMA. The main function of the
EMA is to help on the diagnosis of a patient by checking the current data and proc-
essed it with the data mining tool. EMA moves to all allowed patient for the service.
Before deploying, the EMA request verification to RM for security reason so that the
data will not be altered by malicious attack.



Fig. 3. Proposed mobile agent middleware
Heterogeneous location based services
User virtual environment  Mobility virtual terminal
Communication  Migration  Naming  Security  Interoperation  Data mining
Support
Mobile agent core services
Persistency
Mobility middleware



Virtual resource management
Java virtual machine



3.2  Mobile Agent Middleware
Mobile agent-based middleware is one of the issues of research for providing an ad-
vanced infrastructure that integrates protocols, mechanism, and tools to permit com-
munication to mobile agents. SOMA in [19] discusses more issues of the mobile agent
middleware. In our research, the design of mobile agent middleware is a Java-based
platform. The infrastructure is divided in layers of service for designing, implementing,
and deploying mobile agent-based applications. As shown in Figure 3, our proposed
middleware consists of four layers. We focused more on the last component which is
the data mining support. This additional service is provided to operate the data mining
of the mobile agent on the data of patient.
4  Data Mining Model for Diagnosis Support
The expert mobile agent or EMA performs the consultation to the patients for advance
diagnosis which is based on the proposed data mining model. Our data mining model
has  two  phases  shown  in  Figure  4.  The  first  phase  includes  the  training  of  EMA’s
fuzzy system based on data pre-processing by selecting relevant information from the
profile of an expert. Also, the training phase provides EMA to have an accurate classi-
fication based on the expert profile of the doctor or specialist. The second phase proc-
esses the data of the patient to neuro-fuzzy classification. The procedure is done by
deploying the EMA from the PDA of the doctor in the room and classifies the data of
patient. The security configuration of deployment is also considered  in this  process.
After the process, the results are returned display the result of consultation.


Fig. 4. Data mining model using neuro-fuzzy for support of diagnosing the patient
History Data
of Patient
Set the
Property of
the Mobile
Agent based
on   
Train the
Fuzzy
Classifier of
Mobile Agent
Phase 1: Training
Expert
Mobile
Agent

Deploy and
read data of
Patient
Classify
using the
trained
EMA  

Result of
Classifica-
tion  

Return to
the Mobile
device
EMA

Patient’s
Data
Phase 1: Classifying



The proposed data mining approach considers the profile of an expert in the mobile
device as the  basis  of extracting the relevant information from Phase  1.  Now let us
consider the set of profiles that will be used in the preprocessing data mining: P = {p
1
,
p
2
… p
x
}. After collecting the profiles, the mobile agent uses these features to select
the relevant attributes of C where it is the raw data from the patient history database.
Let D as the set of the selected tuples from C. Equation 1 represent the pre-processing
algorithm.  The following are the phases of our proposed algorithm.

)( )(  where
},...,,{
xn
1
21
valuepvalueattributec
cccCD
n
i
n
=
=
∑
=



(1)



Let us say the EMA is a cardiologist then the cases related to heart disease are gath-
ered by D. D is used to train the fuzzy system of EMA. The structure of the neuro-
fuzzy system consists of three layered perceptron.  The  1st  layer  is  for  inputs  (U
1
  =
{x
1
,…, x
n
}), 2nd layer is for generating rules (U
2
= {R
1
,…,R
k
}), and 3rd layer is an
output layer (U
3
= {c
1
,…,c
m
}). The system also contains weights from the input layer
(U
1
) to rule layer (U
2
) and from rule layer (U
2
) to the output layer (U
3
). Each connec-
tion between units x
i
∈ U
1
and R
k
∈ U
2
is labeled with a linguistic term A A
jr
(i)
(j
r
∈
{1,…,qi}). The values from the input layer are mapped through the fuzzy sets of the
weights. W(R, c) ∈ {0, 1} holds for all, R ∈ U
2
, c ∈ U
3
. The values from the input
and rule layer are evaluated in the connection of the hidden and output layer. For all
output units, c ∈ U
3
the net input net
c
is calculated Equation 2.

∑
∑

∈
∈
⋅
=
2
2
),(
),(
UR
UR
R
c
cRW
ocRW
net



(2)



To  train the fuzzy sets from the input,  Equation  3  is used. After  the training, the
EMA is ready to classify the data from the patient. A Java codes is shown in Figure 5.

ccRWoo
Uc
RRR
δδ
∑

∈
−=
3
),()1(

(3)





Fig. 5. Neuro-fuzzy classification algorithm integrated in EMA
INPUT:   profile, preprocessdata  
OUTPUT:  NeurofuzzyClassification(preprocessdata)
public class ExpertMobileAgent extends Aglets {
public void Preprocess(String[] profile, String[] val) {
  while(rsData.next())  
  {
if rsData.getObject(profile)=val; AddInfo(rowset)
}
NeurofuzzyClassification(preprocessdata);
}
public void ClassifyData(int in1, double[] pattern) {
}
}






5  Simulation Result
The  proposed  framework  of  multi-agents  was  simulated  using  the  JADE  platform.
Neuro-fuzzy algorithm was coded in Java and embedded it to the expert mobile agents.
The  environment OS platform used here are Windows OS, Red Hat Linux and  Sun
Solaris  8  to  simulate the  heterogeneity  of  system.  To  test  the  performance  of  algo-
rithms, we used data mining tools which are the NEFCLASS and Weka data mining.
We chose the data of heart disease from UCI machine learning repository used by the
machine learning community for the empirical analysis of machine learning algorithms.
5.1  Classification Accuracy
Precision and recall are two typical measures for evaluating the performance of infor-
mation  retrieval  systems.  Given  a  discovered  cluster  γ  and  the  associated  reference
cluster  Γ,  precision  (PγΓ)  and  recall  (RγΓ)  applied  to  evaluate  the  performance  of
clustering  algorithms.  In  classifier  algorithm,  recall  and  precision  is  performed  by
cross-validation  test  of  the  classified  instances.  To  evaluate  the  performance of  the
algorithms,  these  measurements were  used.  This  is  done  by  calculating  the  average
precisions  in  Equation  4  where  AvgP  is  the  summation  of  precision  (P
n
)  of  classes
divided by the number of classes. Average of recall is computed in Equation 5 where
AvgR is the summation of recall (R
n
) of classes divided by the number of classes. The
number of correctly classified instances was used to determine accuracy. The process-
ing time of modeling of the algorithm and cross-validation of the classifier were ob-
served to determine the time constraint and classification accuracy respectively. The
classical methods used for comparison are simple logistic (SL), multi-layered percep-
tron (MLP) and classifier decision tree (J48) [9, 10] which are highly accurate classi-
fication methods.

n

P
AvgP
n
i
n
∑
=
=
1


(4)

n

R
AvgR
n
i
n
∑
=
=
1


(5)
5.2  Result
Comparison  of  classical  methods  for  performance  is  shown  in  Figure  4.  The  bar
graphs present the comparison of processing time and accuracy of neuro-fuzzy classi-
fication and other classical methods. In Figure 4a, the processing time of neuro-fuzzy
is much faster than the MLP while SL and J48 classifier has less processing time. In
accuracy, we can justify the performance of neuro-fuzzy is better than the other classi-
cal methods in the sense that even though it has a high processing time than the SL
and J48, it is more accurate of classifying patterns shown in Figure 6b.




Fig. 6. Bar graphs showing the processing time and accuracy of each algorithm

The result of precision and recall are presented in Table 1 and 2, respectively. It is
important that the classifier has a high accurate in classification to be used in consulta-
tion procedure of EMA. Neuro-fuzzy has the highest precision which has an average
of 0.91 and recall which has an average of 0.91 compared to MLP (0.81, 0.83), and
SL (0.38, 0.35),  and J48  (0.77, 0.77).  Most  of these classical  methods were able to
predict testing data with the number of misclassified patterns between 51 to 63 while
neuro-fuzzy has only 25 misclassified patterns out of 270 tuples.
Table 1. Precision
Classes  NF  MLP  SL  J48
present  0.89  0.828  0.843  0.793
absent  0.92  0.79  0.821  0.742
Average
0.91  0.809  0.832  0.768
Table 2. Recall
Classes  NF  MLP  SL  J48
present  0.90  0.833  0.86  0.793
absent  0.91  0.783  0.8  0.742
Average
0.91  0.808  0.83  0.768


6      Conclusion

Ubiquitous healthcare shows more researchable topics and it includes the integration
of multi-agent systems. In this paper, we present the framework of ubiquitous health-
care based on multi-agent which supports the mobility and data mining of the mobile
agent. We propose the expert mobile agent (EMA) that performs data mining to sup-
port the diagnosis of a patient. The EMA uses the neuro-fuzzy to process the consulta-
tion function. Also, a pre-processing of the relevant data based on the expert profile is
shown to train the fuzzy system more efficiently. Result from simulations shows that
neuro-fuzzy outperformed other high accurate classifiers. Future work will be more on
the functionality of the proposed multi-agent framework in ubiquitous healthcare.
(b)
(a)


A c c u r a c y  

0
1 0
2 0
3 0
4 0
5 0
6 0
7 0
8 0
9 0
1 0 0
N F M L P S L J 4 8
Percent %
Processing Time
0
10
20
30
40
50
60
70
80
90
100
NF MLP SL J48
Time in Seconds

Romeo Mark A. Mateo, Louie F. Cervantes, Hae-Kwon Yang and Jaewan Lee
School of Electronic and Information Engineering, Kunsan National University
68 Miryong-dong, Kunsan, Chonbuk 573-701, South Korea
{rmmateo, lfcervantes, hkyang, jwlee}@kunsan.ac.kr







References
1.  Foster D., et al.: A Survey of Agent-Based Intelligent Decision Support Systems to Sup-
port Clinical Management and Research. 1st Intl. Workshop on Multi-Agent Systems for
Medicine, Computational Biology, and Bioinformatics, Utrecht, Netherlands (2005)
2.  Brause, R.: Medical Analysis and Diagnosis by Neural Networks. Medical Data Analysis.
Springer-Verlag, Berlin Heidelberg (2001) pp. 1-13
3.  Joo, S., Moon, W. K., and Kim, H. C.: Computer-aided diagnosis of solid breast nodules
on ultrasound with digital image processing and artificial neural network. Engineering in
Medicine and Biology Society, Vol. 1 (2004) pp. 1397-1400
4.  Giger, M.L.: Computer-aided diagnosis of breast lesions in medical images. IEEE Compu-
tational Science and Engineering, Vol. 2, Issue 5 (2000) pp. 39-45
5.  Verma, B.; Zakos, J.: A computer-aided diagnosis system for digital mammograms based
on  fuzzy-neural  and  feature  extraction  techniques.  IEEE  Transactions  on  Information
Technology in Biomedicine, Vol. 5,  Issue 1, (2001) pp. 46-54
6.  Schetinin, V.: Neural Network based Agent for Discovering Rules in Medical Databases: a
vailable at http://citeseer.ist.psu.edu/update/605607
7.  Cervantes, L., Lee, Y. S., Yang, H, Ko, S. H., and Lee, J.: Agent-based Intelligent Deci-
sion  Support  for  the  Home  Healthcare  Environment.  ICHIT  2006,  Jeju,  South  Korea
(2006)
8.  Della Mea, V.: Agents Acting and Moving in Healthcare scenario: A Paradigm of Tele-
medical Collaboration. IEEE Transaction on Information technology in Biomedicine, Vol.
5, Issue 1 (2001) pp. 10-15
9.  Vajirkar, P., Singh, S., and Lee, Y.: Context-Aware Data Mining Framework for Wireless
Medical Application. Database  and Expert  Systems Applications,  LNCS 2736 Springer-
Verlag (2003) pp. 381-391
10.  Lee, J., Mateo, R. M., Gerardo, B. D., and Go, S. H.: Location-Aware Agent Using Data
Mining for the Distributed Location-Based Services. ICCSA 2006,  Part V,  LNCS 3984,
Springer-Verlag (2006) pp. 867-876
11.  Gerardo, B. D., Lee, J. W. and Joo, S.: The HCARD Model using an Agent for Knowl-
edge Discovery. International Journal of KAIS, Vol. 14, Special Issue (2005) pp. 53-58
12.  Zadeh, L. A.: Fuzzy Sets. Information and Control (1965) pp. 338-353
13.  Klose,  A.,  Nürnberger,  A.,  Nauck  ,  D.,  and  Kruse  R.:  Data  Mining  with  Neuro-Fuzzy
Models. Data Mining and Computational Intelligence, Springer-Verlag (2001) pp. 1-36
14.  Nauck, D., and Kruse, R.: NEFCLASS - A Neuro-Fuzzy Approach for the Classification
of Data. In Proceedings of ACM Symposium on Applied Computing, Nashville (1995)
15.  Mitra,  S.,  and  Hayashi,  Y.:  Neuro-Fuzzy  Rule  Generation:  Survey  in  Soft  Computing
Framework. IEEE Trans. Neural Networks, Vol. 11 (2000) pp. 748-768
16.  Halgamuge,  S.  K.,  and  Glesner,  M.:  Neural  Networks  in  Designing  Fuzzy  Systems  for
Real World Applications. Fuzzy Sets and Systems, Vol. 65, No. 1 (1994) pp. 1-12
17.  Wai, R. J., and Chen, P. C.: Intelligent Tracking Control for Robot Manipulator Including
Actuator Dynamics via TSK-type Fuzzy Neural Network. IEEE Trans. Fuzzy Systems Vol.
12 (2004) pp. 552-560
18.  Kirn,  S.:  Ubiquitous  Healthcare:  The  OnkoNet  Mobile  Agents  Architecture.  NetObject-
Days (2002) pp. 105-118
19.  Bellavista, P., Corradi, A., and Stenfalli, S.: Mobile Agent Middleware for Mobile Com-
puting. Computer Journal. (March 2001) pp. 73-81
20.  Hardala, F., Kiri, A., Özdemir, H., Yilmaz, T., and Güler, I.: A Neurofuzzy Classification
System  for  the  Effects  of  Diabetes  Mellitus  on  Ophthalmic  Artery.  Journal  of  Medical
Systems, Vol. 28,  Issue 2  (2004) pp. 167-176

DOI: 10.1007/978-3-540-72830-6_83 · Source: DBLP
Conference: Agent and Multi-Agent Systems: Technologies and Applications, First KES International Symposium, KES-AMSTA 2007, Wroclaw, Poland, May 31- June 1, 2007, Proceedings

来自圈子: 东游西逛