COMMUNICATION OF MEDICAL INFORMATION

USING AGENTS

John McGrory, Frank Clarke

Dublin Institute of Technology, Dublin 8, Ireland

Jane Grimson

School of Computing, Trinity College Dublin, Dublin 2, Ireland

Peter Gaffney

Adelaide & Meath Hospital, Incorporating the National Children's Hospital (AMNCH), Tallaght, Dublin 24, Ireland

Keywords: HL7, OpenEHR, CEN-ENV13606, agents, computer, guidelines and protocols.

Abstract: Agents are self-contained software entities which act faithfully and autonomously on behalf of a body of

knowledge. They can operate in a standalone capacity, or as part of a social group collaborating and

coordinating activities with other software agents. To access their knowledge, agents are interfaced with

using message passing communication. The principle behind medical communications is to provide a means

for exchanging information and knowledge from one computerised location to another, whilst preserving its

true meaning and understanding between the listener and sender. Agent communication is similar to medical

communications, but must provide an additional framework element to allow agents to interact at a social

and operational level. Social aspects relate to agents collaborating on shared objectives, and operational

aspects relate to coordination of tasks between the loosely coupled agents working as part of a group.

Medical communications focus on data exchanges specific to the medical domain, while agent

communication was designed for a much broader audience. Therefore, it is essential to verify if agent

communications can support standard medical data exchanges. This paper investigates current forms of

agent based communications and demonstrates they can support medical communication, yet retain their

social and interaction information exchange functionality.

1 INTRODUCTION

An agent is a standalone, self-contained software

application. It contains its own inference engine and

is encoded using goals, plans and beliefs (Rao et al.,

1995). The goals are used to describe its motivation,

the plans are used to describe its intention (e.g.

different types of workflow activities) and its beliefs

are used to describe facts (e.g. weight, height or

gender). Each plan and goal has some triggering

condition using beliefs which must be satisfied

before the plan or goal can be executed. The

inference engine selects a goal and then chooses

plans which can satisfy that goal whose triggering

condition is also satisfied. Using this execution

dynamic the agent reacts to known beliefs and

events by selecting plans and goals. This permits the

agent to be autonomous and self-contained. Using

this principle the authors have developed an agent to

act on behalf of a clinical and laboratory guideline.

However, guidelines are rarely used in isolation and

it is sometimes necessary for clinicians to review a

number of these documents so a customised

healthcare plan can be made for their specific

patient. For example, if the patient was obese with

early signs of renal failure. A guideline focusing on

the obesity may indicate reducing the carbohydrates

and increasing the protein to lower the weight.

However, another guideline focusing on the renal

failure would indicate reduce the protein in order to

preserve the renal state of the patient. Technically

McGrory J., Clarke F., Grimson J. and Gaffney P. (2008).

COMMUNICATION OF MEDICAL INFORMATION USING AGENTS.

In Proceedings of the First International Conference on Health Informatics, pages 81-86

 SciTePress

both guidelines are correct, but provide conflicting

information affecting patient healthcare.

Applying the agency concept to guidelines, it

was stated that an agent can be encoded with all the

goals, plans and beliefs related to a particular

guideline. Therefore, if the two guidelines referred

to above where encoded into separate autonomous

agents, they could communicate, work as a group

and collaborate to provide a solution. Supporting

knowledge from one guideline could be sent as a

message to the other guideline. When this

knowledge is received by the other it would update

its beliefs (say proteins must be lowered) and this

information changes the goals and plans that can be

selected. This is because the goals and plans are

selected in relation to the agent’s beliefs.

The principle aim of medical data

communication is to provide a means for exchanging

health information and knowledge from one

computerised location to another in a complete and

context rich format. Its goal is to preserve the true

meaning and understanding of information when

communicated to another application. To realise this

different medical communication standards exist,

but none contain any facility for communication to

take place at a social or collaborative level.

The thrust of this paper is to illustrate software

agent communications, particularly the Foundation

for Intelligent Physical Agents (FIPA) message

standard is capable of providing a context rich data

transmission, similar to that offered by existing

medical communication standards, and yet retain its

agent social and interaction information exchange

functionality. This would permit the agent

communication approach to provide equivalent data

exchanges as that provided through the medical

communication standards, but yet allow the agent

act as a socialite with a group of loosely coupled

agents which can collaborate and coordinate to solve

problems.

2 MESSAGE CONSTRUCTS

From a medical perspective there are three main

standards used for constructing messages:

1. CEN-ENV13606-4:1999(CEN, 1999)

(currently under revision prEN13606

:2004(E)),

2. HL7, Release 2,

3. OpenEHR, Release 1,

In the current pre-standard documentation

release of CEN prEN13606:2004 (only certain

parts available at this time) there is a synergy

between the three standards, and although not

officially a standard yet, the CEN standard will be

the focus of this paper.

From the agent perspective there are two main

standards used for constructing agent messages:

1. FIPA Agent Communication Language

(ACL) Message Structure.

2. Knowledge Query Manipulation Language

(KQML)

The FIPA standard incorporates and extends

many aspects of KQML, and has been the most

widely used agent communication standard to date

(Luck et al., 2004). Thus for the purposes of this

paper the agent communication will focus on the

FIPA standard.

Both the CEN medical and FIPA agent

communication standards achieve their exchanges

by message passing where they simultaneously

integrate two types of information within a message

(FIPA, 2002) (CEN, 1999):

1. Lower-level information(message payload)

2. Meta-information about the content of the

message (message envelope).

The message payload is the structure of the

message contents, such as XML, schema, or objects.

The message envelope component relates to how the

message is seen at a network level between the

message sender and receiver. Both the message

payload and envelope have different impacts on

message passing communication. In the following

subsections the function, use and meaning behind

the message envelopes and payloads are discussed.

2.1 Message Payload

The CEN pre-standard prEN13606:2004 does not

insist on a particular message payload type, and

accepts formats such as an XML document, schema,

or an object. All of these formats preserve the

message data relationship model when

communicating information from one system to

another. XML permits developers to organise the

structure and ordering of information in a document,

whilst isolating it from the actual technical content.

XML, in combination with other standards, makes it

possible to define the content of a document

separately from its formatting, making it easier to

reuse content in other applications, or for other

presentation environments. Most importantly XML

provides a basic syntax used to share information

between different kinds of computers, different

HEALTHINF 2008 - International Conference on Health Informatics

applications, and different organisations without

needing to pass through many layers of conversion.

In this case the XML file itself is the message

payload. The schema method is like a map or plan,

where the information elements in the message are

stored in a rigid format, and have a relationship to

each other by virtue of their position in the schema.

The receiving party is aware of the schema structure

and can access the information slots to retrieve the

required information. In this case the schema data

file itself is the message payload. The Object method

is where information is transmitted as a software

object. The receiving party can interface with the

object to retrieve the desired information. In this

case the Object itself is the message payload. FIPA

standard allows any Java object to be sent as part of

a message payload. This object can be an XML,

Schema or other type object which can be handled

by the Java object class.

The primary difficulty in exchanging messages

is to ensure the message being sent is understood

and has the same meaning both by the sender and

receiver. One simple natural language expression,

such as “The woman is on the bus” can be used to

illustrate the complexities associated with

communications between different systems. This

statement can be interpreted in several ways e.g.,

“the woman is travelling in the bus”, or “the woman

is painted onto the side of the bus”, or “the woman is

travelling on top of the bus”. This ambiguity is in

addition to the assumption that the observer (the

listener) receiving the message knows what the

“bus” object is (and this interpretation is the same as

the sender) and “on” is a relationship description

used when discussing the object “bus”. The

confusion associated with this bus example stems

from an overlapping of ontologies. Language and

ontologies are two interconnected components

which are used to formalise the meaning of data, and

preserve that meaning when sending and receiving

messages (

FIPA_a, 2002) (Noy et al., 2001). An

ontology is a data model which represents language

of a domain and is used to reason about these

described objects and relationships between objects.

Most people possess the capability to handle more

than one ontology, such as a domestic and a work-

related ontology. Therefore, different ontologies can

co-exist in the one entity, but care must be made to

ensure the message exchanges are filtered to match

the ontology of the other party. Difficulties in

communicating and sharing medical information

between institutes, individuals or groups has

generated a multitude of ontology and language

implementations for example Galen (Rector et al.,

2005) (Stuckenschmidt et al., 2004), Tambis (Baker

et al., 1999), UMLS (Unified Medical Language

System) (NLM, 2006), ONIONS (Gangemi et al.,

1999), HL7 RIM (Beeler, 2001), GENE (Egana,

2005). These ontologies and language

implementations specify various medical domains

through an abstract conceptualised model of the real

world environment. This demonstrates that no

unique “one-stop-shop” ontology for the medical

domain exists. The FIPA message structure

recognises that in the real world different ontologies

are present, and instead of forcing a single ontology,

it allows many exist in the same environment and

includes a framework to define, describe and

manage them.

The FIPA ontology is composed of two parts, a

vocabulary which describes the terminology of

concepts used by agents in their realm of

communication (e.g. dietitian or renal), and the

classification of the relationships between these

concepts, semantics and structures (FIPA_a, 2002).

Exchanging messages using a specific ontology

provides a richer contextual environment in which to

share information between separate software

entities.

In summary, a payload holds (or contains) the

actual context rich medical information to be

exchanged between two or more systems. The

message is formed using specific ontologies to

ensure the message is understood and has the same

meaning between the sender and receiver. Both the

medical CEN and agent FIPA standards allow

similar types of message payloads to be transmitted.

But for communications to work effectively it is

vital that the message gets to the correct destination.

2.2 Message Envelope

To deliver a message payload to a specific

destination it is necessary to wrap or encapsulate the

payload using a message envelope. A message

envelope consists of a number of key parameters

which allows the message sender, receiver and

content to be clearly identified during message

transmission. Agents not only use messages for

communicating information in a context rich form,

but also for social and collaborative interaction so

they transmit more envelope parameters, and

messages in general. It is therefore imperative to

compare the parameters used by medical and agent

message standards to ensure the agent system can

support them. This will identify what parameters (if

any) would have to be added to the agent

COMMUNICATION OF MEDICAL INFORMATION USING AGENTS

communication model in order to support medical

transmissions.

By analysing parameters used by the ENV

13606-4:1999 (prEN13606:2004-Part 5: Exchange

models was not available) medical communication

standard and comparing them to the FIPA messaging

standard it can be shown that six of the twelve CEN

parameters have similar technical meanings. A list

of these parameters is detailed in Table 1.

Table 1: CEN to FIPA message envelope parameters

comparison.

Item CEN ENV 13606-4:1999 FIPA

1 identification of message

by originator

sender, conversation-

2 EHCR source sender

3 EHCR destination receiver

4 EHCR message related

agent

reply-to

5 language language

6 message reference conversation-id

The main purpose of the FIPA Agent

Communication Language (ACL) messaging

structure is to allow agents to communicate

effectively when being utilised by a wide audience

base, and were not designed specifically for a

medical application. However, FIPA

implementations are free to include customised user-

defined message parameters other than the items

specified within the standard itself. The semantics of

these user-defined parameters is not defined by

FIPA, and FIPA compliance does not require any

particular interpretation of these parameters (FIPA,

2002). The prefix “X-” must be used for the names

of these non-FIPA additional items. By reusing the

overlapping parameters as detailed in Table 1, and

adding the remaining six parameters using the prefix

“X-”, the fixed size of the message envelope is

65kbytes. The agent parameters now include the

ontology parameter, so a message’s ontology can be

clearly identified, in addition to language before the

message payload is accessed.

In summary, the FIPA messaging standard can

be adapted with the addition of these user-defined

parameters, to provide a similar message model to

that detailed in ENV 13606-4:1999, and yet retain its

agent communication functionality.

3 UTILISING AGENT

COMMUNICATION

To compare the agent communications to an existing

approach consider an example were four clinical

guidelines are used together. The implementation

chosen for illustration purposes was the evaluation

of a set of Liver Function Tests to determine the

cause of a chronic anaemia in patients.

One approach to managing the activities of the

four guidelines is to decompose the guidelines into

workflow activities and management rules. The

management rules from each guideline are linked

together centrally using an inference engine. This

inference engine is constructed using rules that

logically link the various workflow activities

together. These management rules provide the

motivation for a centralised inference engine to

choose particular workflow activities depending on

the patients known characteristics (e.g. weight,

gender, height). The patient data is retrieved from

the LIS using an XML message payload coupled

with a standard CEN message envelope. As more

guidelines are decomposed and added, the number

of workflow activities and the size and complexity

of the centralised inference engine increases. But all

the decisions on choosing a particular workflow

activity are performed centrally by the inference

engine as no other separate modules exist in this

system. The guidelines no longer exist as separate

entities.

An alternative solution to capturing the

knowledge of the four guidelines is to encode each

as a separate autonomous agent, one agent for each

guideline. This is achieved by encoding the

guidelines beliefs, goals and plans together in the

agent and using each agent’s inference engine to

interpret them. In this example there is no

centralised management resource, therefore each

agent must establish links to the other agents. To

achieve this agents use message passing to share

supportive information and coordinate activities.

The patient data is retrieved from the LIS by each

agent separately using an XML message payload

coupled with a standard CEN message envelope.

When an agent received a message from another

guideline it altered (if necessary) its own execution

based on this belief. If an agent wished to forward

supporting information to the other agents it used

message passing. Therefore, the only

communication between the separate agent modules

to coordinate activities together was via message

passing, not direct linking as in the centralised

solution. In this implementation no centralised

resource was used to coordinate activities between

the separate agents. After the agents completed

deliberations the outcomes were compared to the

documented case histories from which the

evaluation data was taken. Although outcomes

HEALTHINF 2008 - International Conference on Health Informatics

derived from the agent approach matched that

provided by the case histories, it highlighted that the

agent approach transmitted more messages than the

former centralised approach. On average the group

of four agents transmitted a total of 12 messages per

second between them. This is because the separate

agents relied solely on message passing in which to

share information. During this evaluation it was also

found that although the message envelope was a

fixed size the message payload size could vary

dramatically. This was because different types of

information were being sent between the agents.

Small messages where in the form of short

supportive information comments. A longer message

is where a more detailed packet of information was

sent. Depending on the data being transmitted

varying sizes of message payload ranging from

2kbytes to 360kbytes could occur. So how does this

affect the computer system using the agents? If the

agents, because they can be distributed were located

on separate machines and a total of 12 messages was

transmitted between them the network bandwidth

would range from 8kbps for 2kbyte payloads, to

51kbps for the 360kbyte payloads (assuming the

network had an efficiency of 10bits/byte). This is a

substantial network overhead. However, this

network overhead could be reduced if the agents

with high message coupling where located on the

same machine. The main reason for this message

overhead is that the separate agents need to transmit

data, social interaction data and collaboration data in

order to operate. Whereas the centralised approach

used fixed links between the workflow activities

within the one application and so no network traffic

was needed. However the agent approach was

capable of providing distributed processing of the

activities and only relationships between guidelines

needing to collaborate needed to be established,

where as the centralised approach needed to be run

on a single machine.

4 CONCLUSIONS

This research demonstrates the agent communication

approach offers the capability to transmit medical

information, in an equivalent fashion to that

provided by the existing medical communication

standards, yet retains its agent social and interaction

information exchange functionality. The FIPA ACL

standard also allows for the ontology of a message to

be identified so a richer form of data interaction can

occur between agents. However, the concept of

using the agents to represent guidelines highlighted

that as the agents communicated data using

messages, they also used messages to socialise and

coordinate activities which could have a substantial

impact on the network overhead. A summary of the

difference between the two approaches is shown in

Table 2.

Table 2: Summary Agent and centralised approaches.

Aspect Agents Centralised

Workflow

activity links

between

guidelines.

Achieved using

coordination

message passing.

Achieved within

the centralised

inference engine.

Sharing of

information.

Achieved using

message passing.

Achieved by

triggering rules

within the

inference engine.

Access to LIS. Achieved using

message passing.

But each agent

accesses it

separately.

Achieved using

message passing.

But all required

information

accessed via one

message.

Processing of

guideline

knowledge.

Distributed. Centralised.

Guideline

knowledge file

size.

Small as each agent

is self-contained.

Large as

inference engine

must cover all

guidelines.

Adding,

altering or

deleting of

guideline

knowledge.

As each agent is

independent there is

no fixed link

between them. So

guidelines can be

added, altered or

deleted without

impacting any other

resources.

Adding, altering

or deleting

requires the links

in the centralised

inference engine

to be modified.

Therefore, all

other resources

affected.

REFERENCES

Baker P.G., Brass A., Bechhofer S., Goble C., Paton N.,

Stevens R., 1999. ”TAMBIS-Transparent Access to

Multiple Bioinformatics Information Sources”,

Scientific and Statistical Database Management, 1999.

Eleventh International Conference on 28-30 July 1999

Page: 280.

Beeler G.W., 2001. Reference Information Model (RIM)

HL7 Version 3, tutorial in two parts Part I & Part II

accessed via http://www.hl7.org, 2001.

CEN/TC 251 Health Informatics, 1999. ”ENV13606-

4:1999 Health informatics – Electronic Healthcare

record communication - Part 4: messages for the

exchange of information, accessed via

www.centc251.org.

Egana A., “Ontology Design Patters for the Formalisation

of Biological Ontologies”, Degree of Masters of

Philosophy, University of Manchester, 2005 accessed

via http://gong.man.ac.uk/docs/MPhilThesis.pdf

COMMUNICATION OF MEDICAL INFORMATION USING AGENTS

FIPA, 2002(Foundation for Intelligent Physical Agents).

FIPA Agent Message Transport Service Specification,

Document Number SC00067F, Dated 2002/12/03,

accessed via http://www. fipa.org

FIPA_a, 2002(Foundation for Intelligent Physical Agents).

“FIPA Device Ontology Specification”, Document

Number SC00091E, Dated 2002/12/03, accessed via

http://www. fipa.org

Gangemi A., Pisanelli D.M., Steve G., 1999. “An

Overview of the ONIONS Project: Applying

Ontologies to the Integration of Medical

Terminologies”, Data and Knowledge Engineering,

vol. 31, 1999, Pages 183-220, published by Elsevier.

Luck M., McBurney P., Preist C., 2004. Agent

Technology: Enabling Next Generation Computing: A

Roadmap for Agent-Based Computing, accessed via

AgentLink web site

http://www.agentlink.org/roadmap/ 2004 ISBN 0854

327886.

NLM 2006, National Library of Medicine USA Unified

Medical Language System accessed via

http://www.nlm.nih.gov/pubs/factsheets/umls.html,

accessed October 2006.

Noy N.F., McGuinness D.L., 2001. Ontology

Development 101: A Guide to Creating Your First

Ontology.

http://protege.stanford.edu/publications/ontology_deve

lopment/ontology101.pdf. Stanford Knowledge

Systems Laboratory Technical Report KSL-01-05,

March 2001.

Rao A., Georgeff M., 1995. BDI Agents: From Theory to

Practice, proceedings of the first international

conference on multi agent systems (ICMMAS-95) San

Francisco USA 1995.

Rector A., Roger J., 2005. Ontological & Practical Issues

in using a Description Logic to Represent Medical

Concepts: Experience from GALEN, School of

Computer Science, The University of Manchester,

Preprint Series, CSPP-35.

Stuckenschmidt H., van Harmelen F., Serafini L., Bouquet

P., Giunchiglia F., 2004. Using C-OWL for the

Alignment and Merging of Medical Ontologies,

Proceedings of the First International Workshop on

Formal Biomedical Knowledge Representation

KRMed'04", Whistler, BC, Canada, 1 June 2004.

HEALTHINF 2008 - International Conference on Health Informatics