PATIENT-CENTRED LABORATORY VALIDATION

USING SOFTWARE 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: Clinical laboratory validation, patient-centred, patient-focused, agents, computer-interpretable-guideline

(CIG), guidelines and protocols.

Abstract: Guidelines are self-contained documents which healthcare professionals reference to obtain knowledge

about a specific condition or process. They interface with these documents and apply known facts about

specific patients to gain useful supportive information to aid in developing a diagnosis or manage a

condition. To automate this process a series of Standard Operating Procedures (SOP) and workflow

processes are constructed using the contents of these documents in order to manage the validation flow of a

patient sample. These processes decompose the guidelines into workflow plans, which are then called using

condition triggers controlled by a centralised management engine. The software BDI agent offers an

alternative dynamic which more closely matches the modus operandi of narrative based medical guidelines.

An agent’s beliefs capture information attributes, plans capture the deliberative and action attributes, and

desire captures the motivational attributes of the guideline in a self-contained autonomous software module.

Agents acting on behalf of guidelines which overlap and interweave in similar domains can collaborate and

coordinate in a loosely coupled fashion without the need for an all encompassing centralised plan.

1 INTRODUCTION

The primary role of a clinical laboratory is to

support frontline healthcare professionals who are

licensed to deal with patients (McLoughlin, 2006).

Their function is to accept appropriate patient

samples, analyse them, and report their findings

back to the ordering clinician (Marshall et al., 1995).

The reported results are not considered a diagnosis,

but used by clinicians to deliver patient specific care.

In the majority of cases these results are use to aid in

planning treatments, quantifying medication

amounts and monitoring patient responses, which

could all have a detrimental affect on the patient if

incorrect information was used (Witte et al., 1997).

Therefore, the single most important activity

performed by the laboratory technologists is to

ensure their generated results are valid and plausible

for the specific patient from whom the sample was

taken.

Clinical guidelines are condition focused

documents through which domain specific aims,

goals, procedures, plans and normal reference ranges

are disseminated to healthcare professionals. The

purpose of these documents are to guide the reader,

and streamline activities around a particular medical

condition or process using evidence based

supportive information. When a clinical or

laboratory guideline is developed by an expert group

they focus on best practice for the specific condition

or process. They include all relevant knowledge,

logic and motivational aspects they deem necessary

to adequately describe the domain.

274

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

PATIENT-CENTRED LABORATORY VALIDATION USING SOFTWARE AGENTS.

In Proceedings of the First International Conference on Health Informatics, pages 274-279

 SciTePress

Clinicians and laboratory technologists care for

patients not diseases or processes, therefore it is their

responsibility to filter through these guidelines

acting on a patient’s behalf. They must try interface

with these documents, to make use of the maximum

decision-making support for healthcare delivery

based on the known facts about their individual

patient. To automate this process of searching

through guidelines on a patient’s behalf, the

laboratory technologists in association with the

clinicians construct a series of Standard Operating

Procedures (SOP), and workflow processes using the

contents of these documents in order to manage the

validation flow of a patient sample. This is

accomplished by decomposing each guideline in to a

series of separate workflow activity paths. Then

develop a set of centralised management rules to

link these activities based on the presented patient

data. However, these procedures are not truly

patient-centred but process-centred. The guidelines

knowledge, logic and motivation can no longer be

accessed as a standalone resource, but as a series of

workflow triggers managed by a centralised

software package, which no longer resembles the

author’s guideline. This process is fundamentally

different to the true operation of a medical guideline,

where guidelines are used to provide supporting

information based on their holistic view of the

domain, rather than a series of linked activities

relating to a process. So is there an alternative

approach where the process can be distributed and

the guidelines retain logic, knowledge and

motivation as a standalone self-contained unit.

Agent oriented architectures operate on similar

principles to elements found in human decision-

making by combining attributes (beliefs), methods

(plans) and desires (goals). The BDI agent approach

in particular is based on the principle of a belief

capturing the informational attributes, the desire

capturing motivational attributes and the intention

capturing the deliberative attributes of an agent (Rao

et al., 1995). Therefore, agents can be considered

self-contained knowledge sources (KS), with a

social communication interface and have the ability

to act autonomously, or as part of a larger group. In

research completed by the authors it was shown that

a software agent can successfully capture and be

encoded with the knowledge, logic and motivation

of a guideline (McGrory_a et al., 2008). In

additional research completed by the authors it was

shown that although agent communications provide

a facility to transmit data between agents, it is also

used to provide a social and collaborative aspect

(McGrory_b et al., 2008). This allows the separate

agents work in groups and collaborate on shared

goals. This later research also demonstrated that

agent communication was capable of being adapted

to comply with a medical standard for

communication (i.e. CEN ENV 13606-4:1999).

The thrust of this paper is to illustrate that

software agents offer an alternative approach to

reproduce the function of medical guidelines than

the more commonly used centralised approaches.

This paper also presents an operation of a

framework which allows these agents coordinate and

collaborate to validate a patient sample in a

distributed fashion, without the need for a

centralised all encompassing plan.

2 ANALYSIS OF GUIDELINE

REPRESENTATION FORMATS

The traditional approach to combining separate

bodies of knowledge (such as guidelines) together is

to decompose the knowledge and logic into separate

workflow activities and link these activities together

using a centralised inference engine. Three

commonly used techniques are rule base, direct

coupling or blackboard systems.

The rule-base approaches are designed around a

nodal tree, where expert knowledge in the form of a

workflow activity is the branch, and the selecting of

a particular branch at each node is based on patient

information or process data. Although selecting rules

based on presented facts during execution can be

indicative of an illness, the rules which link them

directly to a diagnosis do not reflect anything deeper

than a casual understanding of human physiology.

These systems are centralised and the original

guideline knowledge is now absorbed within a

labyrinth of rules.

Direct coupling architectures are made up of a

group of separate expert knowledge modules. Each

expert knowledge module contains local storage, a

KS and a control switch to link to the other software

modules according to their data-flow requirements

using a direct call or link (Corkill, 2003).

Complications arise when specific modules are

subject to change and/or when the ordering of

module control switching cannot be determined until

run-time (Kavanagh et al., 2002). As the system

expands and evolves the links change and the

process becomes unwieldy and unmanageable. In

addition to the aforementioned issues, the direct

coupling model does not provide a clear

PATIENT-CENTRED LABORATORY VALIDATION USING SOFTWARE AGENTS

275

representation of the overall problem, and there is

nothing more than relationship links used.

The blackboard model is based around three

components: KS, control element and the blackboard

(Turban et al., 2005). The KS is an expert at solving

specific elements of the overall problem. The

blackboard, acts as a central repository for data,

partial solutions and control information. The

blackboard also acts as a communication medium

for the transfer of information, and a KS triggering

mechanism. The control element directs the

problem-solving process by allowing KS’s to

respond to blackboard changes, and it selects the

most appropriate KS to be executed next, as shown

in Figure 1. After completing a task the KS reports

back to the blackboard and returns control to the

control element. KS’s are not aware, and cannot

communicate with other KS’s directly. They know

nothing about the other experts (e.g., what

parameters they use, what processes they perform, or

what services they provide). The blackboard

architecture tends to be a labyrinth of different

configurations, levels of abstraction, and partial

solutions which are orchestrated to provide a flexible

problem solving mechanism. The blackboard system

eliminates the communication issues raised by the

directly coupled monolithic model, and gives a

representation of the problem to be solved to all

participants. But the blackboard does not have the

capacity to indicate how group members can

collaborate to solve a problem, but can only select

from partial solutions it already possesses.

Dietician

Knowledge

Source

Diabetes

Knowledge

Source

Knowledge

Source

Renal

Knowledge

Source

Interaction

Blackboard Control

Plans

& Logic

Local

Blackboard

Manager

Dietician

Knowledge

Source

Diabetes

Knowledge

Source

Knowledge

Source

Renal

Knowledge

Source

Interaction

Blackboard Control

Plans

& Logic

Local

Blackboard

Manager

Figure 1: Blackboard Management Communications.

3 DESIGN OF A GUIDELINE

AGENT

The agent approach is based on the principle that

each agent can represent a single guideline. It

captures all the guidelines knowledge, logic and

motivation. In addition to this the agent has a rich

communication facility where data and social

interaction between separate agents can take place.

But how does each agent know what the other is

doing since there is no centralised all encompassing

plan? If the blackboard is simply a repository of

information, although layered to some degree, there

is no absolute necessity for it to be in a single

location. Therefore, it is possible to replicate a copy

of the blackboard within each Autonomous

Socialising Knowledge agent (ASK-agent) as shown

in Figure 2.

Interaction

Blackboard Control

Knowledge

Source

BDI

Local

Dietician Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

GP Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

Diabetes Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

Renal Guidelines

AGENT

Interaction

Blackboard Control

Knowledge

Source

BDI

Local

Dietician Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

GP Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

Diabetes Guidelines

AGENT

Blackboard Control

Knowledge

Source

BDI

Local

Renal Guidelines

AGENT

Figure 2: Autonomous Socialising Knowledge agent

model.

Each ASK-agent now contains a localised

blackboard, knowledge source, localised beliefs,

localised control and its own inference engine. The

retention of the motivational component of the

guideline within the ASK-agent is fundamentally

different to the centralised approach of other

systems. This allows the ASK-agent to act

autonomously on behalf of the guideline in a self-

contained capacity. When patient specific

information is presented to the individual agents,

they have the ability to apply their encoded

knowledge and logic, and provide a supportive

response based solely on that information. Using this

approach an ASK-agent module can make use of the

maximum supportive response from the other

HEALTHINF 2008 - International Conference on Health Informatics

276

separate ASK-agent’s based on the known facts

about the individual patient. By providing a

framework which allows separate ASK-agent

broadcast supportive communications to each other,

the agent approach offers the opportunity for the

data to be validated in a patient-centred fashion. But

how can these separate, autonomous, self-contained

ASK-agent modules share data, work in groups or

collaborate to solve a problem.

3.1 Agent to Agent based Activity

A theory which can aid issues relating to

collaborating guidelines is Activity Theory (AT)

which emanated through the social sciences. AT

focuses on the collaborative nature of separate

autonomous systems such as individuals (Engestrom

et al., 1999), on which agents are based, and have

the capability to perform certain tasks as part of a

group. Agents synthesise human decision-making

through their goal, plan and belief elements, but do

not explicitly detail how they can socialise or

collaborate. AT in itself does not provide an output

which can be exactly transposed into computer

software, but does provide a useful framework based

around interfacing interaction and collaboration of

software modules. These interfaces can be used to

develop an increased sense of interaction and

collaboration ability in autonomous modules using a

software program independent approach.

The structure of human activity according to

Engestrom can be compartmentalised using rules,

community, subject, object, division of labour and

instruments. AT is an iterative process where an

activity is developed from a simple low level

activity to a higher level activity.

Instruments

Artifacts i.e. laboratory results

Division of

Labour

Using message content

such as LiverDisease

alerts other agents to get involved

Rules

Rules acted by the ALK_P agent

By the community

By the object

Community

Liver experts

GGT agent, ALT agent, Protein Agent

Subject

ALK_P agent

Object

The human body

Outcomes

Instruments

Artifacts i.e. laboratory results

Division of

Labour

Using message content

such as LiverDisease

alerts other agents to get involved

Rules

Rules acted by the ALK_P agent

By the community

By the object

Community

Liver experts

GGT agent, ALT agent, Protein Agent

Subject

ALK_P agent

Object

The human body

Outcomes

Figure 3: Low level activity.

In terms of medical guidelines the iterative

process dynamic exists by virtue of the design of

guidelines and their focus on a condition, disease or

organ. Consider for example, a low level

implementation activity being the guideline behind

the validation of a single analyte result, say Alkaline

Phosphates as shown in Figure 3. A higher level

implementation is where the result is combined with

some other single analyte results, such as Bilirubin

and GGT, to perform and aid in the reporting of a

Liver Function Test as shown in Figure 4. The Liver

Function Test is then part of a higher level suite of

tests for other medical disorder classifications.

Instruments

Artefacts i.e. diagnosis liver dysfunction

or indication LiverDisease

Division of

Labour

Using message content

such as Hepatitis

alerts other agents to get involved

Rules

Rules acted by the Liver Expert

By the community

By the object

Community

Experts associated with the liver

Dietician expert agent, Kidney expert agent

Haematology Expert agent

Subject

Liver Expert agent

Object

The human body

Outcomes

Instruments

Artefacts i.e. diagnosis liver dysfunction

or indication LiverDisease

Division of

Labour

Using message content

such as Hepatitis

alerts other agents to get involved

Rules

Rules acted by the Liver Expert

By the community

By the object

Community

Experts associated with the liver

Dietician expert agent, Kidney expert agent

Haematology Expert agent

Subject

Liver Expert agent

Object

The human body

Outcomes

Figure 4: High level activity.

One guideline does not cover the whole body,

but more specifically focuses on an abstract

conceptualisation of body components (e.g., liver

function in a group of male diabetes patients).

Another guideline relates to the same body

component but from a different abstract

conceptualisation viewpoint (e.g., kidney function in

a group of male diabetes patients). Although the two

guidelines are separate autonomous documents they

are linked by virtue of their domain of discourse.

Therefore, a link between different guidelines

already exists within the guideline document itself.

The overlapping knowledge is provided in two main

forms. The first is in the form of similar domain

knowledge that uses alternative inference

mechanisms in order to derive a result (i.e. both

statistical and rule-based inference engines being

able to validate the same result). The second is in the

form of overlapping knowledge which observes

different viewpoints of the same domain. For

example, the kidney filters toxins from the blood

passing it to the urinary tract. As the kidney is such

an integrated organ in the body there are many

guidelines describing its operation from different

viewpoints such as blood filtering, urinary tract,

autoimmune disorders etc. Using this approach the

organ disease or condition is described from

different viewpoints through various guidelines.

Each guideline describes different knowledge, logic

and motivational aspects associated with the organ.

Therefore supportive information can be exchanged

between these guidelines in order to aid in

PATIENT-CENTRED LABORATORY VALIDATION USING SOFTWARE AGENTS

277

describing the operation of the organ, or in the

validation of a sample result.

3.2 ASK-agent Model

To utilise this overlapping knowledge link a social

structure was developed to manage the interfacing

between agents. This social interface took the form

of a mandatory set of searchable service

descriptions, beliefs and actions. The service

descriptions (i.e. Name, Type, Ownership,

GuidelineReference, InformationNeeded,

ValidationType, EndResultType, Ontology and

Language) permitted each agent to be located within

the agency platform through the Directory

Facilitator (DF) (a feature of the Foundation for

Intelligent Physical Agents (FIPA) standard offering

searchable goldenpages facility to locate agents)

(McGrory_a et al., 2008). The beliefs (i.e.

CurrentlyValidating, PlausibilityScore and localised

blackboard) permit the ASK-agent to interact with

other group members. The actions relate to

automated responses the ASK-agent must return to

other agents when queried (e.g.

CurrentlyValidating), and the sending of

information to other agents it believes should be

reported (e.g. it determined the presence of liver

disease during its deliberation). Therefore, each

ASK-agent only needs to know its overlapping

neighbours, which it can find and interact with using

the agent platforms DF and message passing. With

access to supportive and overlapping knowledge it is

not necessary to have a single all-encompassing rule

set to manage the ASK-agents interaction.

The fundamental concept of the ASK-agent

system proposed in this paper is to allow

components to collaborate and share supportive

information without having to explicitly disclose

their position as part of the large encompassing

community. An ASK-agent does not need to identify

exactly what every other agent is doing; only what

its neighbours (i.e. neighbours it interfaces with) are

doing. To illustrate this point further and

demonstrate some boundaries, consider the example

of a jigsaw with 500 pieces. A jigsaw piece has two

discrete dimensions: the irregular shaped edge

containing four sides, and the image printed on the

face. To solve the puzzle, a person directly matches

individual jigsaw pieces onto the jigsaw image, say

the image shown on the box. Each piece is identified

using the image on its face and placed in the

appropriate position. This method requires a view of

the whole system to be presented before starting, but

involves no greater skill than straightforward pattern

matching. An alternative approach is to use a

combination of the localised image on the face of the

piece and its four corners to match it to a suitable

neighbour (i.e. matching the shape of the pieces

together). Jigsaw assembly using these interfaces do

not require the full picture to be known. Using the

jigsaw example as a solution metaphor, the heart and

lungs image depicted in Figure 5 is a symbolic

representation of the heart and lungs as a whole, not

just the image it represents. The agent was not

intended to be a large all encompassing structure,

but a group of loosely coupled autonomous expert

knowledge sources (represented by each jigsaw

piece) which could be readily and easily interfaced

with as shown in Figure 5. The ASK-agent only

needs to know its neighbours (i.e. the expert it

interfaces with), in a similar way the jigsaw piece

only needs to know another piece with similar edge

profile and compatible image, not the whole picture.

The ASK-agent does not need to know anything

about any other piece of the jigsaw only its

interfacing neighbours. The interface can be

considered the ontology, overlapping facts, common

laboratory results and various viewpoints of the

universe of discourse. This is analogous to the

jigsaw edge shape profile.

Figure 5: Jigsaw metaphor representing agent components

of the heart and lungs.

4 CONCLUSIONS

This research demonstrates the agency approach

offers a facility to manage and interface with

medical guidelines electronically, in a similar modus

operandi to original guideline documents. This is

because of the synergy between the knowledge base,

plans, decisions, action, goals and the self-contained

nature components between guidelines and agents.

HEALTHINF 2008 - International Conference on Health Informatics

278

Table 1: Summary of centralised and ASK-agent

approach.

The agents can be encoded to reproduce the beliefs,

desires and intentions of the narrative guideline and

act accurately, faithfully and autonomously on

behalf of that document. This body of knowledge

and logic can then be interfaced with, whenever that

information needs to be accessed. The addition of

activity theory and in particular the iteration model

concept showed that the guideline documents

already contain aspects that link them together.

Using these links and the developed social

communication the ASK-agents can locate, access,

communicate, collaborate and coordinate activities

between each other. This allows supportive

information exchanges to be completed between

separate expert agents about an individual patient,

without the need for an all encompassing centralised

plan. In cases where there is an inconsistency in held

patient specific information, this agent approach

offers an advanced, robust and efficient patient

centred validation alternative to existing approaches.

However, if overlapping knowledge between

guidelines is not available the links created using

this approach are not present and the separate

guidelines are standalone islands of information. The

guidelines knowledge, logic and motivations are still

accessible as a standalone entity, but other agents

would need to be created to provide the links.

Developing a system using the latter approach still

permits distributed processing to be accomplished,

but not without a source of knowledge to provide the

links. A summary of the differences between the

centralised and ASK-agent approach are given in

Table 1.

REFERENCES

Corkill D., 2003. Collaborating Software Blackboard and

Multi-Agent Systems & the Future. Accessed via

www.cs.umass.edu, Proceedings of the International

Lisp Conference published in October 2003.

Engestrom Y., Miettinen R., Punamaki R. Perspectives on

Activity Theory. Published by Cambridge University

Press January 1999. ISBN13: 9780521437301.

Kavanagh M., Price S., “The quest for a computerized

guideline standard: The process, its history, and an

evaluation of the most common and promising

methods used today”, Capstone Project documentation

2002.

Marshall W.J., Bangert S.K., 1995. Clinical Biochemistry,

Metabolic and Clinical Aspects, ISBN 0 443 043 418,

published in 1995 by Churchill Livingstone.

McGrory_a J., Grimson J., Clarke F., Gaffney P., “Agents

representing medical guidelines”, HEALTHINF 2008

International Conference on Health

informatics, Funchal, Madeira, Portugal, January

2008.

McGrory_b J., Grimson J., Clarke F., Gaffney P.,

“Communication of medical information using

agents”, HEALTHINF 2008 International Conference

on Health informatics, Funchal, Madeira, Portugal,

January 2008.

McLoughlin V., Millar J., Mattke S., Franca M., Jonsson

P., Somekh D. and Bates D., 2006. Selecting

indicators for patient safety at the health system level

in OECD countries. International Journal for Quality

in Health Care 2006 18(Supplement 1):14-20.

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.

Turban E., Aronson J.E., Liang TP., Decision Support

Systems and Intelligent Systems. published in 2005 by

Pearson Prentice Hall, Seven Edition, ISBN 0-13-

123013-1, Chapter 10, pgs 538-570.

Witte D.L., VanNess S., Angstadt D., and Pennell B.J.,

1997. Errors, mistakes, blunders, outliers, or

unacceptable results: how many? Published by

Clinical Chemistry 1997; Vol:43 pgs 1352-1356.

Element Centralised

Approach

ASK-Agent

Multi-ontologies No Yes

Processing Centralised Distributed

Requirement for

overlapping

knowledge

None Required to provide

the links.

Addition,

altering or

removal of

guidelines from

the system

Any changes

require

centralised

inference to be

recompiled.

Each ASK-agent is

independent and

loaded separately.

System

resilience

None Yes, all ASK-agents

have a copy of the

blackboard.

Independent

accessible

knowledge

None. All access

to information

through the

centralised

engine.

Yes, all ASK-agents

are independent.

Information

accessed via

message passing.

Clinician having

access to

specific

guideline

knowledge

A clinician

cannot access

knowledge

directly.

A clinician can

access each ASK-

agent via a message

and directly access

the specific

guideline

knowledge.

Method of

collaboration

Direct links using

the centralised

engine.

Using the

mandatory beliefs,

action and

descriptions in the

agent platforms