A FRAMEWORK FOR DYNAMIC KNOWLEDGE ACQUISITION

Ana Aguilera and Alberto Subero

Facultad de Ciencias y Tecnología, Universidad de Carabobo, Valencia, Venezuela

Keywords: Collaborative Work, Multi-agents, Medical diagnosis.

Abstract: This article proposes a conceptual model based on agents in medical workgroups for the resolution of

clinical cases. Our objective is to provide a suitable conceptual framework for the dynamic acquisition of

expert knowledge from different sources. This framework is sufficiently flexible to allow its transformation

to the symbolic level that finally characterizes the implementation approach; it is a design model prior to

implementation. We considered an interface design to be used for the integration and combination of

different sources of expertise. The subject which interests us is the knowledge produced in the workgroup

processes.

1 INTRODUCTION

There are many applications related to the

workgroup such as decision-making, monitoring,

transmission, planning, idea generation, problem

resolution, resulting discussion, negotiation, conflict

resolution, analysis and system design and

collaborative group activities such as the preparation

and distribution of documents, to mention just a few.

This work style is illustrated by real-life

applications, where their nature is clearly

collaborative. Several of these applications imply

various specialized fields, such as medical diagnosis.

The field of medicine has become so vast that it is

divided into several specialities and many medical

cases are considered by group analysis. Among the

various workgroup activities, complex-problem

resolution is highlighted as being an increasingly

important subject today.

Different information technologies which

consider these group modalities have been

developed in software engineering. Thus, we found

a broader field entitled group support systems or

electronic meeting systems which include other wide

areas like Group Decision Support Systems (GDSS),

Distributed Group Support Systems (DDSS),

Computer Support Collaborative Work (CSCW),

Groupware, where the common denominator of such

technologies covers, but is not limited to, distributed

facilities, computer hardware and software, audio

and video technology, procedures, methodologies,

facilitation, and applicable group data (Turban,

1995).

On the other hand, the evolution of the

application domain in traditional Artificial

Intelligence (AI) to cover other complex and

heterogeneous fields such as aid to decision-making,

form recognition and comprehension, process

control, etc., shows the limits of traditional AI

approaches. All this has contributed to the birth of a

new discipline: Distributed Artificial Intelligence

(DAI), which is interested in the intelligent

behaviours (and their modelling) that are the

outcome of cooperative activity between several

agents. Nevertheless, the passage from individual

behaviour to collective behaviour is not only

considered as an extension but as an enrichment of

AI, as new properties and new behaviours emerge

from it. The purpose of DAI is to provide a remedy

for the shortcomings of the traditional AI approach

by proposing expertise distribution for an agent

group capable of working and acting in a common

environment, and solving possible conflicts. New

concepts in AI have appeared, such as cooperation,

action coordination, negotiation and emergence.

Three fundamental research areas of DAI are multi-

agent systems (MAS) (Weiss, 2000), distributed

problem solving (DPS) and Parallel Artificial

Intelligence (PAI).

2 THE PROBLEM

Everyday medical work (implying inter-consultation

between specialists, case conferences, and hospital

morning rounds) includes exchange and cognitive

Aguilera A. and Subero A. (2008).

A FRAMEWORK FOR DYNAMIC KNOWLEDGE ACQUISITION.

In Proceedings of the International Conference on e-Business, pages 58-63

DOI: 10.5220/0001913800580063

 SciTePress

processes within groups. It is possible to observe

different physicians (specialists and non-specialists)

working together, at the same time or at the different

time, on the same patient (clinical examination often

requires the cooperation of several specialists). The

contribution of these exchanges to solve complex

problems is becoming more and more essential.

These acts of collaboration are important for clinical

decision-making concerning diagnosis and treatment

both for students still in training and recently

qualified physicians, without forgetting the

continuous experience which helps to develop expert

knowledge. There are several examples: cases of

multi-system illnesses where physiopathology and

the nature of origin of the illness make it necessary

to examine the results of several diagnostic

procedures; patients with chronic disorders such as

diabetes mellitus, obstructive pulmonary illnesses,

cardiological illnesses; or patients receiving

palliative care at home.

On the whole, the most frequent scenarios where

it is possible to observe this collaborative pattern

are:

Inter-consultations. This is a process where a

physician needs specialized consultation with

several other physicians. There are two possibilities

for this process: 1) between specialists with the same

speciality (consultation between radiologists who

study an image to decide on a diagnosis) and 2)

between specialists with different specialties (an

obstetrician who refers his/her patient to a

cardiologist for a coronary problem).

Case Conferences and morning rounds. There are

cases where it is possible to find several physicians

(specialist physicians interconnected with general

practitioners) exchanging information on several

medical cases in order to make patient evaluations,

to work on the publication of cases, etc. These are

considered as staff meetings.

A clinical case illustrating this type of work is

shown in figure 1 (A more detailed explication is

found in (Quintero, 2003). In this case, we can

observe the interaction between a generalist, a

radiologist and a pathologist.

3 PROPOSITION

3.1 Architecture

The problem follows a structure on three levels

(Figure 2): the collaborator level or users (human

agents), computer systems agents (software agents)

and the data level (databases and knowledge bases).

This architecture determines the structure on which

the framework will be defined and permits us to

clarify the semantic definition of the problem under

study.

3.2 Methodology

Different methodologies for the development of

knowledge-based systems (where knowledge

acquisition and modelling play a leading role) have

emerged during the past 15 years. We can mention

in particular methodologies such as KADS

(Schreiber, 1993), VITAL (Domingue, 1993),

REFLECT (Reinders, 1991), (van Harmelen, 1992),

ACKnowledge (van Heijst, 1992), CommonKADS

(Schreiber, 2000) and KATEMES-MultiExp

(Ladibi, 1995). KADS and VITAL focus on the

development cycle of a knowledge-based system,

REFLECT proposes theories and architectures for

reflexive systems, ACKnowledge introduces

directive models for knowledge modelling.

CommonKADS offers methodology for knowledge-

intensive system development and for knowledge

management and KATEMES-MultiExp is a

prototype tool for multi-expert knowledge

acquisition.

The approach used is based on an extension of the

KATEMES-MultiExp methodology. This method

integrates a set of models for modelling multi-

expertise and its objective is to help the knowledge

engineer in the collecting and modelling phase of

human behaviours in cooperative problem solving.

These models are: agent, organization, cooperation,

task and communication models (Aguilera, 2003b).

We added a coordination model (Aguilera, 2003a)

because originally KATEMES was designed for

knowledge acquisition before system development

and we considered the dynamic knowledge

acquisition approach. One of the important features

considered is the ability to extract expertise in a non-

intrusive way. This minimizes biased behaviour by

monitoring expert decision makers directly (through

the information system) during their daily tasks,

without changing anything about their choices or

strategies. The approach is based on the concept of

the agent, with the idea of modelling a group of

experts via a community of agents in interaction.

This gives us the additional advantage of multi-agent

systems which provide a way to relax the constraints

of centralized, planned, sequential control and to

make systems that are decentralized, emergent and

concurrent available. We are convinced that, the

modularity of a multi-agent architecture facilitates

knowledge acquisition and the parallel design of

each expertise. In fact, the multi-agent systems are

A FRAMEWORK FOR DYNAMIC KNOWLEDGE ACQUISITION

Circumscribed 2-3 cm Nodule located in the

right upper lobe of the lung with the

resence of

interior calcifications of non-specified type.

Scanner recommended. Heart and rest of study

without obvious modifications.

CAT Lung scanner: It indicates a 2 x 3 cm mass with non-epimacular aspect located in the right upper lobe of

the lung with non-specified calcifications.

There is no affectation of Mediastinum lymphatic ganglia. There are no other masses in the thorax.

Pulmonary biopsy: macro and microscopic analysis of post-operative piece.

Diagnosis: ENDOBRONCHIAL HAMARTHOMA.

The patient leaves the hospital and considering the benign origin of the pathology, the doctor recommends an

annual check up with his general practitioner.

Patient data

Age: 40 years old.

Sex: masculine

Context: He is a non-smoker and he does not present any obvious particular antecedents in his past medical

history.

Symptoms: He went to his general practitioner with a non-productive cough of three months.

Physical Test: normal

Treatment: palliative treatment

Complementary Tests: laboratory and paraclinical tests (Postero-anterior chest x-ray)

Figure 1: A clinical case.

Medical

Past History

Medical

Staff

Physician

(in charge

of the case)

Collaborator Level

Knowledge

base

Distributor

Agent

Planning

Agent

Manager

Agent

Interface

Agent

Data

Agent

Computer Systems Level

Data Level

App. Speciality-

oriented

Figure 2: Problem Structure.

suitable for highly dynamic domains. They belong to

the complex open systems category and make

incremental application development possible.

Our choice is justified because agent modelling

adapts well to the distributed nature of the problem,

as well as, to the preliminary phases of the

knowledge acquisition and to the development of

collaborative activities between several experts.

3.3 The Framework

It is clear that the installation of a knowledge

acquisition environment coming from multiple

experts requires a multi-field approach which, at the

same time, integrates social, cognitive and data-

processing dimensions. It is for this reason that the

objective of this model is to integrate social aspects

ICE-B 2008 - International Conference on e-Business

coming from the human activities, cognitive aspects

held in the databases and knowledge bases of the

system, and the aspects related to computer systems.

This proposal considers important aspects of multi-

agent systems and software engineering.

We consider the importance of the conceptual

model phase in the development of any technology,

since it provides a good definition and a clear

analysis of the problem treated. The model helps the

knowledge engineer to consider the presence of all

significant elements and their correct interrelations

and it is also an instrument which enables him to

exchange opinions with colleagues and experts, etc.

The advantage of this framework is that it offers us a

high level tool of abstraction. This tool enables us to

discuss viewpoints with others and is therefore a

facilitator for conceptual analysis. Additionally, this

framework is a set of models that gives a wide

vision including all actors, their interrelations, their

organization, their communication mechanisms, and

more.

The framework definition is given,

simultaneously in the structuring and

conceptualizing of each model in the extended

KATEMES-multiexpert methodology. At the

conceptual level we can define the following

models:

• Agent model

This model enables us to identify and define all

human and artificial individuals interacting and

taking part in cooperative problem solving. These

agents are (figure 2):

- The medical meeting agent: This is considered

as a high level hierarchical agent. It is a set of agents

and it is defined to identify the group of physicians.

The communication protocol between agents follows

human communication procedure based on personal

conversations; that is, somebody intervenes and the

others remain silent and listen, and when one person

has finished, another intervenes and so on. The

leadership position can be assumed by the doctor

sitting opposite the patient (Physician in charge of

the case).

- Medical speciality agents. They define a

classification based on the different medical

specialities, i.e., human agents such as radiologists,

neurologists, pathologists and others. The

knowledge of agents is heterogeneous. The agents

divide the tasks amongst themselves and share data

about the patient. Each one of these specialists can

observe only one part of the “outside” (i.e., the

patient).

- The patient: A human agent who as an active

being takes part in his/her treatment. He/she requests

medical appointments, takes part in his/her cure,

asks for information. He/she makes it possible to

establish the clinical context.

- The planning agent: An artificial agent which

manages everything related to time in the system:

assigns the tasks with a time context, checks the

time of their activation, execution and ending. It

returns the requests of task assignments if it has not

received an opportune response. It centralizes

information about task executions and their different

states of development.

- The distributor agent: An artificial agent which

locates the agents that will execute the tasks,

according to their specialities and availability for

executing them. For this, it takes into account the

urgency of tasks required. It centralizes information

about group members.

- The manager agent: An artificial agent which

controls the requests for tasks. It sends and receives

results.

- Interface agent: Artificial agent who controls

the security accesses to systems and data. It

adequately adapts the appropriate interface

according to each user with his/her different levels.

It centralizes general access information.

- The data agent: An artificial agent which

manages data in the databases and knowledge bases.

It manages metadata, user view definitions and

authorizations to data access. It also controls

anything related to data integrity and recovery

mechanisms.

Subsequently and for reasons of brevity we will

define only the radiologist agent. This is a human

agent, an expert and a specialist belonging to the

radiology department. He may be or not be

subordinate according to his status. He is part of a

community including radiologists or other

colleagues of different specialities (general

practitioners, pathologists, etc). All of these people

rank equally and take part in the development of a

diagnostic solution. This agent also has a

hierarchical relationship with the department head.

With respect to inter-agent relationships, the

radiologist could influence the reasoning line of an

agent with another speciality, who requires his

service for a diagnostic examination, or he could

directly influence another radiologist who discusses

a particular case with him. In the radiology

department, software systems for image processing

are frequently used. The execution of these systems

can be initiated or stopped at any time by

radiologists. Concerning cooperative relationships,

the agent can assist another radiologist in the

establishment of a diagnostic hypothesis or in other

tasks. In the medical group context, his task does not

interfere with other tasks but sometimes its

A FRAMEWORK FOR DYNAMIC KNOWLEDGE ACQUISITION

accomplishment is important for the continuation of

other tasks. Other inter-agent relationships can be

established, but in order to simplicity, only those

related to the example given (figure 1) will be

mentioned.

• Expertise model

Knowledge based systems are founded on a

separation between the knowledge necessary for

problem solving (domain knowledge) and the

reasoning mechanisms needed to exploit this

knowledge (inferential process knowledge for

problem solving). In the research field, this subject

remains open and there is a great diversity of

analyses and different viewpoints about it. This topic

is related to reasoning mechanisms used by doctors

in diagnostic problem solving. Many variables are

considered: patient, doctor experience, general

medical knowledge. The discussion of this theme is

extended and goes beyond the domain of this article.

We suggest readers refer to specialised literature

(Torasso 2001, Long, 2001), in particular, we refer

to (Park, 2003).

• Organisation model

In the architecture proposed (Figure 2), we find

artificial and human agents. From this, we identify

the medical group as a hierarchical structure with

two levels; headed by the physician treating the

patient, and then the rest of the doctors are

considered to be collaborators. This structure can

change by establishing as group leader another

doctor in any other speciality.

It is also possible to observe other hierarchical

structures, for example in a medical group within the

same department (radiology department, pathology

department, etc.), with the department head at the

highest level.

Within a group of artificial agents, some have

only coordination functions. A hierarchical structure

is present and centralized in the manager agent. It is

possible to observe delegation of tasks. For example,

the manager agent requests that the planning agent

specify the localization of a task and then the latter

asks to the distributor agent which finally assigns the

task.

• Cooperation model

The model presupposes the cooperative

motivation of its participants. In the case of a

medical group, the doctors can be assisted within the

same speciality, when they request consultation at

any step of the diagnostic process (doubts,

confirmations, etc.). In different specialities, no one

interferes directly with the others but their effective

participation, in the total establishment of diagnosis,

is decisive.

• Task model

To arrive at a diagnosis, doctors carry out several

tasks. In particular we mention: hypothesis

generation, selection and test of hypotheses and

diagnostic conclusion. In the case of hypothesis

generation, which is the example presented here, it is

not clear if the same process is followed equally in

every medical speciality. Undoubtedly, there are

different contexts and expertise models can vary

within the same speciality. However, it is not our

objective to present a discussion of the diagnostic

process.

• Communication model

There are several communication mechanisms

between agents given by the type of their

interrelations. For example, in the communication

between human agents, there are communications

based on spoken conversations, written

conversations (chats) or by data (medical file).

Between human and artificial agents, the

communication is established by an interface agent.

Among artificial agents, there are mechanisms based

on message passing, and finally, between artificial

agents and data, through a data agent.

Communication modelling is described by inter-

agent relationships via R-intervention relationships

which do not contradict the organisational structures

described. Some of them are shown in the definition

of the radiologist agent.

• Coordination model

Coordination is supported by artificial agents: the

manager, the planning and the distributor agent.

They centralize everything related to the

coordination tasks required. Coordination will also

depend on the dynamics of the cases exposed, their

characteristics, their emergency requirements, their

resources and the interrelations established between

the active agents..

Therefore services offered by this model depend

on: the session context, the resources provided by

participants and the social behaviour of participants

(actions carried out by human agents) (Ossowski,

1999). The combination of these three elements

creates the dynamics of coordination and the basic

actions for this collaboration; the context fixes the

limits of these dynamics.

3.4 System Design and Implementation

At implementation level, we design a multi-agent

architecture based on the framework proposed. The

different medical specialities are modelled well with

a modular structure. The current project is

progressing. So far, we have designed and

implemented a groupware application that facilitates

the management of clinical cases (Cárdenas, 2001)

and a groupware application for radiological

ICE-B 2008 - International Conference on e-Business

teleconsultations. The conceptual data model of our

application uses an entity relationship model. At the

moment, we are developing a workflow application

for the coordination of medical diagnostic work.

4 CONCLUSIONS

In this article we defined an agent-based conceptual

framework that integrates the social aspects of

human activities, the cognitive aspects structured in

databases and knowledge bases of the system, as

well as, the computer system aspects which are

present in the software applications and information

technologies, in the medical diagnostic context. This

approach is model-based; thus six models are

defined: the agent model, the organization model,

the cooperation model, the task model, the

communication model and the coordination model.

It is possible to instantiate these models for

particular situations, for example, clinic and hospital

frameworks. We have considered the extension of

the metamodel proposed in (Ladibi, 1995), the

annexing of the coordination model and we provided

general templates for medical groupwork. The

selection of the model-based method and the design

of the model were directed towards the dynamic

knowledge acquisition produced in collaborative

medical interactions, which is the next the step in

this research.

ACKNOWLEDGEMENTS

We are grateful to FONACIT, Venezuela, under

Project No G-2005000278, which has sponsored this

work.

REFERENCES

Aguilera A.(a), “An agent model for medical diagnostic

support with multi-experts", Sciences of Electronic,

Technology of Information and Telecommunications,

SETIT 2003, Sousse, Tunisia, Mars 2003, paper

R234A002.

A. Aguilera (b), « Télé-consultation en radiologie

médicale », Département ITI, ENSTB. Colloque Les

Nouvelles Technologies dans la cité, Rennes- France,

Déc. 2003 WWW : http://tic-cite.univ-

rennes1.fr/resume.php?lien=42.txt&mail=ana.aguilera

@enst-bretagne.fr.

F. Cárdenas, J. Pacheco, “Desarrollo del software para un

sistema colaborativo de diagnosis médica”. Thesis in

electrical engineering, Universidad de Carabobo,

2001.

J. Domingue, E. Motta Et S. Watt, “The Emerging VITAL

Workbench”, Lecture Notes in Computer Science,

volume 723, pp. 320-339. Springer-Verlag, Toulouse

and Caylus, France, 7th European Workshop, EKAW,

1993.

S. Ladibi, “Ingénierie de la Connaissance dans le cadre de

Projets Multi-Experts : Méthode, Techniques et

Outil”. Thèse Doctorale, Université de Nice – Sophia

Antipolis, 1995.

W. J. Long, “Medical informatics: reasoning methods”,

Artificial Intelligence in Medicine, Volume 23, Issue

1, August 2001, Pp. 71-87.

S. Ossowski, “Coordination in Artificial Agent Societies.

Social Structure and Its Implications for Autonomous

Problem-Solving Agents”, Number 1202 in LNAI.

Springer Verlag, 1999.

W. Park, “Causalité et Raisonnements dans le

diagnostique médical”, proceedings Setit, Sousse-

Tunisia, 2003.

J. Quintero, “Décision collaborative dans le domaine

médical analyse d’un cas de diagnostic”, proceedings

Setit, Sousse- Tunisia, 2003.

M. Reinders, H. Akkermans, et al., “A conceptual

modelling framework for knowledge-level reflection”,

AI-Communications, 4(2/3) :74-87, 1991.

G. Schreiberakkermans, H., Anjewierden, A., et al.,

“Knowledge Engineering and Management. The

CommonKADS Methodology”. The MIT Press, 2000.

P. Torasso, “Multiple representations and multi-modal

reasoning in medical diagnostic systems”, Artificial

Intelligence in Medicine, Volume 23, Issue 1, August

2001, Pp 49-69.

E. Turban, “Decision support and Expert Systems.

Management Support Systems”, Prentice Hall, 4°

Edition, 1995.

R. Van Harmelen, “Knowledge-level reflection :

Specifications and architectures”, Technical Report

R.2 RFL/UvA/III.2, ESPRIT Basic Research Action

P3178 REFLECT, REFLECT Consortium, 1992.

G. Van Heijst, P Terpstra, et al., “Using generalized

directive models in knowledge acquisition”, In Proc.

of the 6th European Knowledge Acquisition

Workshop, pp. 112-132, Heidelberg and

Kaiserslautern, Germany, Lecture Notes in Artificial

Intelligence, 1992.

G. Weiss, “Multiagent Systems: A Modern Approach to

Distributed Artificial Intelligent”, The MIT Press,

2000.

A FRAMEWORK FOR DYNAMIC KNOWLEDGE ACQUISITION