Using Explicit Ontologies in Agent-based Healthcare Information Systems

Sabina Falasconi, Giordano Lanzola and Mario StefanelliUniversity of Pavia, Dept. of Informatics and Systems Science

Medical Informatics LaboratoryVia Ferrata, 1 - 27100 Pavia - Italy

sabina@ipvaimed6, unipv, it{giordano, mario}@ipvaimedl.unipv.it

Abstract. Within knowledge engineering a new researchparadigm is emerging based on the Multi-Agent System(MAS) architectural framework, allowing human andsoftware agents to interoperate and thus cooperate withincommon application areas. Such a vision entails bridgingthe different "’views of the world" of knowledgeable agentsthrough the commitment to common definitions of theconceptual entities and of the technical terms employed inknowledge/data bases. We claim that within a multi-agentframework ontological libraries and also standardizedterminological repositories should be "’agentified" intoontology and terminology servers providing other agentswith the common semantic foundation required foreffective interoperation, and allowing the configuration ofsuitable application ontologies for distributed applications.We restrict our analysis to the case of an organization ofcognitive agents, which we’ll illustrate with examplesfrom a prototypical health care MAS, that is, a so-calledDistributed Healthcare Information System (D-HIS). Theprototype makes use of an ontological library written in thestandard language Ontolingua.

10ntologies in Agent-based HISs

The management of a patient in a shared-care contextis a knowledge intensive activity, involvingmultimodal elaboration of clinical information data,structured text, images and signals. Health careproviders must be able to exchange information andshare a common understanding of the patient’sclinical evolution. To support collaborative work inmedical care, computer technology should not onlyaugment the capabilities of the individual specialist,but also enhance the ability of collaborators tointeract with each other and with computationalresources.Thus, a major shift is needed from first generationHospital Information Systems (HISs), which weremainly intended as simple centralized informationrepositories, to a distributed environment composedof several interconnected entities which activelycooperate in maintaining a full track of the patientclinical history and supporting care providers in all

the phases of patient management. This involvesassembling multiple, in general heterogeneousknowledge/data based systems into a community of atleast semi-autonomous agents, thus forming what iscalled a Multi-Agent System (MAS) [Werner, 1992]. our notion, the resulting MAS includes also the usersreferred to as human agents, in order to distinguishthem from software/hardware ones. The maindifferences with respect to former approaches lie in theencapsulation of software agents’ capabilities, and in aminimal autonomy requirement, so that they canoperate someway without user intervention andpossess a module controlling their behavior[Wooiridge and Jennings, 1995; Castelfranchi, 1995b].Agents communicate via a language that defines basicinteractions. A suitable Agent CommunicationLanguage (ACL) is thus needed that minimizesdependence on agents’ implementation and can beuniformly interpreted by community members[Genesereth and Ketchpel, 1994; Labrou and Finin,1994].In order that heterogeneous and specialized health careproviders, human or software ones, may collaborateand interoperate, their views of the world must besomehow bridged. A prerequisite is that the providerscommit to explicitly and formally stated specificationsof standardized lexicons and descriptions of theinvolved entities and concepts, or ontologies.Besides, traditional methodologies scarcely take intoaccount the actual scatter of data, concerning the sameentity, across multiple, big-sized, heterogeneous andgeographically distributed repositories. For examplethe very same person may hold records as a worker ina corporate database, as a citizen in a nationaladministrative database and as a patient in a clinicaldatabase. To resolve interoperation problems arisingfrom the semantic mismatches specifically due to thedifferent domains covered by large-scale datarepositories special-purpose ontologies [Wiederhold,1994a; Wiederhold, 1994b] are being proposed as abasic instrument.

189

From: AAAI Technical Report SS-97-06. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.

Explicit ontologies as a mechanism enabling thesharing and reusing of existing data and knowledgebases among diversely specialized computer-basedsystems have been standing in the breach for quite awhile [Gruber, 1993: Gruber and Oisen, 1994], but themost recent research developments and trendsinvolving ontological engineering appear focused onissues regarding the design of multi-componentsystems based on (semi-)autonomous knowledgeableunits competent for different tasks and, in the generalcase, skilled in different domains, like e.g. in [Takedaet al., 1994; Lanzola et al., 1996].Thus, we can say that the knowledge/data engineeringfield seems to be clearly evolving towards a newparadigm in which cognitive agents (endowed withsymbolic representation of knowledge) ofheterogeneous nature, that possess diverse (but ofcourse, at least partially compatible or inter-translatable) conceptual views, or ontologies,modelling both their own expertise and the externalenvironment, make available their informationresources or problem solving abilities for cooperativeprocesses addressing the construction of a new agentor the achievement of some common goal throughcorrelated execution of tasks. The reliance on networkfacilities for providing such agents with the necessarycommunication abilities is also becoming an almostunavoidable choice especially in the case ofknowledge-based MASs used in real-worldorganizations. Furthermore, the need of guaranteeingsharability of data and data-based services is widelyrecognized. Recent projects and applications in thedistributed healthcare area, meeting at least some ofthe mentioned design principles and requirements,are described in [Reddy et al., 1993; Engelmann et al.,1994; Szolovits et al., 1994; Huang et al., 1995].We claim that, to be effectively shared, ontologiesshould become a service offered by an enchargedagent, thereby called ontology server, to a communityof interoperating agents with which it results suitablyintegrated, possibly relying upon a medicalterminology server, in its turn exploiting existing largecomputerized repositories of standard medicalterminologies and classification schemes. Moreprecisely, we restrict our analysis to the case of anorganization of cognitive agents [Castelfranchi,1995a; Conte and Castelfranchi, 1995], which we willillustrate with examples from a prototypical healthcare MAS, that is, a so-called Distributed HealthcareInformation System (D-HIS) [Lanzola et al., 1995;

Lanzola et al., 1996]. Such prototypical D-HIS iscurrently being developed through the agentificationof a set of pre-existing standalone systems thataddress several aspects of knowledge representationand problem solving in medicine [Lanzola andStefanelli, 1992; van Heijst et al., 19941.

2 Modelling Organizations

As already hinted, a main reason for the shift to MASarchitectures can be identified in the inescapable needof effectively integrating, in a scalable way, peopleand computer systems within real, complexworkplaces. Socio-economic and psychologicalliteratures view them as instances of humanorganizations, for which they worked outterminologies and models widely borrowed andadapted by researchers in Distributed ArtificialIntelligence [Gasser, 1991; Werner, 1992], the areathat gave origin to the MAS field.What characterizes an organization? According to[Cont,~ and Castelfranchi, 1995], groups ororganizations can be viewed as social systemsendowed with macro-level goals influencingindividual agent’s behaviors so that they are orbecome adapted to pursue those purposes. Socialactions can occur only in the presence of an oftencomplex network of conditions and relations thataccount for the social system structure. Morespecifically, for the social system to be adequately"implemented", the member agents must be boundeach other through social commitments [Gasser,1991], that is relations expressing the engagement ofan agent to perform some action in behalf of (at least)another one [Castelfranchi, 1995a]. This means thatan organization is "more" than the sum of theproperties and behaviors of its members, which stillenjoy a "socially bounded autonomy" allowing themto pursue individual (micro-level) goals at leastthrough self-sufficient task execution (executiveautonomy [Castelfranchi, 1995b]).But for an organization to show reliable and stablecourses of activity some more inter-agentrelationships are needed, expressing theorganizational roles of member agents in terms ofthe general tasks they result permanently andnormatively committed to. These also identify therelationships of mutual dependence and power/authority (e.g. within a hierarchy), that account for shared "awareness" of the need to cooperate. Suchstructuring links, known by member agents,

190

distinguish organizations, groups and teams fromother forms of "extemporaneous" social systems orcommunities.The need of formalizations expressing the underlyingframework of composite organizations exploitingcomputer-based systems was captured also by otherprojects, like those converging into the EnterpriseProject, aiming at providing a methodology and atoolset for enterprise modelling, thanks to thespecification of structural concepts in a centralEnterprise Ontology [Uschold et al., 1995]. Thisextensive and quite complex ontology, written in thestandard language Ontolingua [Gruber, 1993],includes as a basic component a theory oforganizational structure general enough for adaptingalso to non-business frameworks: for example, itdefines concepts like ORGANIZATIONAL UNIT,PERSON and MACHINE. The same could be said forother subtheories specifying concepts inspired fromthe area of multi-agent (therein called actors)planning, like those of activity, plan and resource. Themain role taken by the ontology is that of a mediatingmechanism between people alone, people andcomputational systems, computational systems aloneand finally enterprises as a whole. The general modelis intended to be further specialized into a set ofmodels to be stored in a library, as advocated also e.g.,n [Dieng et ai., 1995].Within such organizations, various types ofinteractions may become feasible, being function ofthe structural relationships among agents (roles,dependence). In a strongly normative institution, therewill be limited space for task bargaining amongagents. For example, in [Conte and Castelfranchi,1995] an orchestrated cooperation model is devised,in which only the "boss agent" knows and enforcesthe cooperative plan, assigning roles and thus goalsand tasks to other agents. We call this kind ofinteraction, in which a dominant legitimate authorityexists, coordination. Coordination limits the amountof necessary inter-agent communication, tries toeliminate situations of conflict and avoids antagonisticor non cooperative behaviors, like those illustrated byWemer [Werner, ! 992] in his diagram representing theso-called "Spectrum of Cooperation Styles". In thecase of a D-HIS for a hospital, the enforcement ofofficial guidelines, indicating treatment procedures forgiven classes of diagnosed diseases, can be suitablyobtained creating an agent possessing the globalprocedure (the guideline manager) and delegating

steps or single tasks of the latter to other onesaccording to their institutional competence.When instead a substantial margin is left to tasks andservices negotiation among agents, we will speak of"’proper" cooperation. Cooperation requires, ingeneral, a richer ACL (that is, more or morestructured agent communication primitives), may beitself object of negotiation, and may depend on theproblem at hand. For example, it may be establishedin an occasional, temporary or permanent way. In thesame D-HIS, health care providers may set suchcooperation links with each other whenever in anoccasional, short-term or long-term perspective theyneed to consult another specialist or to benefit from aparticular service they are not directly competent orauthorized for.Therefore, besides the role assignment as prescribedwithin the organizational ontology, identifyingstructural relationships in a declarative fashion.agents will necessarily possess bodies of proceduralknowledge implementing rt~ies and routines of socialinteraction, whose models may likewise be describedin ontological theories, to be used as an aid to MASand ACL design and implementation. Such proceduralknowledge chunks will indeed define theconfiguration, interpretation and reaction modes(varying with the interaction style) with respect ACL messages and will operationalize real workflowroutes and tracks. Some interaction rules andprocedures may as well be the result of a negotiationsession where some steps or parameters of thecoordination or cooperation plan still need to beestablished or specified. Thus, the workflowknowledge evolves dynamically, while stayingcommitted to the adopted organizational ontology.

3 Agent and MAS Models

From a software engineering point of view, a softwareagent can be defined as an at least semi-autonomousprocess able to interoperate with other processes(running on the same or on a separate machine)through a suitable ACL [Genesereth, 1992;Genesereth and Ketchpei, 1994]. We assume anintuitive general software agent model, adapted from[Burmeister and Sundermeyer, 1992] and shown inFig. !, in which each "isolated" entity possesses oneor more conveniently represented informationrepositories using which it can perform, withoutinteracting with other agents, processing operationsby means of suitable elaboration and control

191

,-,I

I IElaboration and Control

F-]ComponcnL~ACL Sl~aking

Ctmm|unication Facility

Information Repositories

AGENT AENVIRONMENT

Figure I. A general software agent model.

components. Information flows in an agent peculiarway along its I/O channels facing the environment(e.g. sensors and actuators).Cooperation/coordination with other agents isachieved by extending the representation mechanismsof this isolated entity with ACL communicationprimitives. An essential feature of such primitives isthat they must allow agent integration at theknowledge level [Newell, 1982], that is, in a mannerindependent of implementation related aspects. For thesake of simplicity, the inter-agent communicationfacility is shown as a self-contained module in thefigure. Thus individual agents are essentially viewed as"units of knowledge and interaction" [Gasser, 1991].Among proposed ACLs, the Knowledge Query andManipulation Language (KQML) developed withinthe Knowledge Sharing Effort [Finin et al., 1993] isreceiving more and more consensus both from thetheoretical and from the applicative point of view. Itconsists of a set of communication primitives ofnotification ( t: o 1 l, unt e 11) and of request and reply(ask-if, reply), and allows to linguisticallydecouple the communication modes (expressed inKQML) from the communicated information. Thecontent portion of a message should mention theobjects and relations in a predefined conceptualization,that is in an ontology known both by the sending andby the receiving agents. These design characteristicsappeared suitable for the agentification process weundertook on previously developed prototypicalsystems, and also for the integration of a new agent inan existing community. Therefore we strongly tookinspiration from KQML while developing ourexperimental agent communication protocol [Lanzolaet al., 1995].As already explained, in our framework we considerthe individual agent makes little sense as an isolated

entity and rather we are interested in the roles and thebehaviors it assumes in the context and in behalf of alarger group. Thus, even at an organization-independent design stage we firstly made afundamental distinction among software agents.Server agents provide other agents with informationelements from their repositories, which are oftenstructured as libraries of sharable and reusablecomponents. That is, these components can be used asbuilding blocks for off-line agent construction orexpansion, or can be serviced on-line on request.Instances of this category can be terminology servers,ontology servers, problem solvers servers, dataservers, and so on, each named after the kind ofinformation chunks it is able to offer. Server agentsare thus needed to speed up agent configuration and toallow the distribution and reuse of resources (e.g. ofclinical data and medical knowledge). A fundamentalrole in the specification of an agent’s competence isplayed by ontology and terminology servers: termsand concepts definitions identify the set of entities ~hat"exist" for an agent. Application agents are insteadendowed with specialistic competence; thus they canreply to queries pertaining to their expertise field. Inthis category we can classify systems equipped withapplication knowledge and appropriate reasoningtechniques such as KBSs, specialistic databases withtheir DBMS, systems provided with particularcomputational skills as statistical software packages,devices endowed with peculiar abilities of interactionwith physical environment such as monitoringsystems, active sensors, robots, and so on.Secondly, we devised three basic types of medicalapplication agents, characterized by complementarycompetencies to be integrated in order to create aknowledgeable composite entity, that we namedapplication agency [Lanzola et al., 1995], here shownin Fig. 2. Desktop Agents are essentially more or lesscomplex and customized user interfaces from whichhuman users can access, besides routine services likee-mail, the facilities of the two other agent types: DataManagement Agents (DMAs) and ReasoningManagement Agents (RMAs). While the formermanage local or distributed databases, the latter areendowed with local or remote application knowledgebases and reasoning capabilities in order to performcomplex and highly specialized problem solvingactivities.All the facilities offered by traditional databases,KBSs and user interfaces can be gathered and

192

DesktopAgent

IO’Ma"agemen’lA ,n, IS--gMana ’mentlAge-- JFigure 2. The basic application agency.

simultaneously made distributed by such applicationagencies, that are shaped, starting from a commonexplicit conceptualization or application ontology, tomeet the needs and the patient management view of aparticular category of human users. Actually, theapplication agencies offer a high-level, tailored viewon the services provided by the networked systems.For example, the prototypical agencies implementedwithin a D-HIS were adapted to support thephysician, the nurse and the administrator in somerespectively typical tasks. Therefore, a physicianagency is more biased on the medical side helpingclinicians in assessing the patient’s state andidentifying a suitable therapy, while a manageragency is instead committed to the financial aspects,evaluating the suitability of patient management on acost-effectiveness basis. Finally, nurse applicationskeep track of the evolution of patient states helping incollecting data and detecting the occurrence ofabnormal situations [Lanzola et al., 1996].In summary, the general network of software agentscan be schematized as in Fig. 3, with the shownapplication agencies representing a particularconfiguration of the data-based or knowledge-basedservices offered by the distributed architecture,obtained thanks to the application model andvocabulary. Social or organizational roles are thensuperimposed on these autonomous knowledgeablecollective agents [Conte and Castelfranchi, 1995]directly interfacing with human ones.

4 The Ontological Library

A well known notion of ontology identifies it with thespecification of a conceptualization [Gruber, 1993]that is, with a set of definitions both of the objectspertaining to some area of interest, and of therelationships hypothesized to hold among them. Sucha specification, when performed through a logic-based language endowed with declarative semantics

Ontology [Terminology i

Server iServer 1

~SNetwork of /~xoftware A gents z~A~!nact~On~

\I

fOata-bases ! Solvers

i Server ~ ~ Server

Figure 3. The general network.

and, in general, allowing to associate natural languagedescriptions with formal statements (and thus, humanreadability with machine readability), constitutes special-purpose axiomatization of an application area.The adoption of a standard or conventional ontologyspecification language allows to employ an ontologyrelative to a distributed application as a couplingmechanism among the cooperating parties, allcommitting to the chosen definitions (that is, sharingthe same ontological commitments) while exchanginginformation (data/knowledge) or posing/answeringqueries.As the Ontolingua language [Gruber, 1993] wasprecisely developed according to these design lines,we decided to employ it to construct a preliminarylibrary of ontologies [Falasconi and Stefanelli, 1994;van Heijst et al., 1995]. We claim that the underlyingstructure of the knowledge required for a particular,and in general distributed, application is given in anapplication ontology assembled by including andpossibly refining some theories selected from thelibrary. That is, for the complementary components ofa running application agency to effectively collaboratein supporting a consistent user-oriented set of services,they must rely on the same basic conceptualization,well known also to human operators. As a matter offact, each Desktop Agent of the existing prototypicalagent-based workstations offers a local applicationontology browsing facility enabling the user to exploreand grasp the semantic foundation underlying all theavailable tools and facilities. Indeed the design andconfiguration processes of a new application agencyshould start with the explicitation of the corresponding

193

application ontology.Both library and application ontologies are notintended as task-independent knowledge bases[Guarino, 1995]. but as sets of definitions expressed ina rigorous declarative style, that is, independently ofany procedural detail. In other words, they shouldleave aside features strictly related to the encoding ofknowledge at a certain abstraction level and with acertain representation formalism, in order to allow theRMA to coordinate and exploit heterogeneousproblem solvers. Thus the specification of anapplication ontology represents an attempt to abstractfrom reasoning and elaboration details, and to expressthe assumptions under which some knowledge anddata bases result "true" and can be profitablyemployed.Nevertheless, the task model and the employedrepresentation formalisms may be usefully introducedinto the final application ontology in the form of theroles taken by instances of domain concepts withintask execution in that application using thoseformalisms. Such role assignment is then viewed as ana posteriori procedure within the ontologyconfiguration process, maybe performed through theconsultation of general task and method ontologiesstored in the library, as explained in the following.Thus, we adopt a non-philosophical, utilitarian idea ofontologies to be used in knowledge engineering,remarking that a primary purpose is that of clarifyingall the assumptions that are hardly reconstructiblestarting only from coded knowledge, e.g. from a"’uniform" frame hierarchy.The current library is structured as an inclusion graphof ontological theories, partially shown in Fig. 4. Thegraph is a Directed Acyclic Graph whose nodes aretheory names and whose arcs connect an includingtheory to an included theory. A criterium ofminimization of inclusion links was followed: eachtheory includes, directly or indirectly, the minimumset of other definitions it presumes. Theory namesarise from the principal ontological category aroundwhich the definitions are organized. For example, thetheory Diseases is centered around the notion ofdisease, represented as an Ontolingua class withseveral binary relations defined on its instances, likedisease .manifestation. Some of theserelations are also functions, likedisease.etiology. According to the valuesassumed by these relations, equivalent to frame slots,in an application knowledge base (an instantiation of

GENERIC-CONCEPTS

FUNDAM ENTA I.-M EDICA I.-CON(’EPTS

AN!TOM%

GENERIC-TASKS ~kPIt~ SIOI.OG~

SURGERIES DRUGS FINDINGS

\/ CI.INIC .-S’I’ATY~,\

DISEASES

v tTHERAPIF~ /~F~TS

CIANICAI,-ENVIRONMENT

~IC.PATIENT

FUNDAMENTA I r M EDICA I,-TASKS

Figure 4. Theories inclusion graph of the current librao’of medical ontologies.

an application ontology), actual disease instances canbe classified into opportune taxonomies, whosesubtype relations (e.g. etiological-disease-subtype) are too defined in Diseases.As a library manager and distributor, the ontologyserver should maintain possibly inconsistentontological theories in the library, keeping track of therelations among them. A categorization of ontologiesaccording to the kind and scope of theirconceptualization (e.g. generic versus domain-specificontologies) could be useful to this aim. We identifiedsome tentative categories, not intended as exhaustiveor mutually exclusive, that single out clusters oftheories within the current graph, and could work as"context-markers" for alternative available theories.Representation Ontologies (not shown in Fig. 4) definea set of representation primitives pertaining toknowledge representation paradigms, languages andformalisms. In our framework, they are intended toshow the neutrality, with respect to world entities, thatwas pointed out (and contested) in [Guarino andBoldrin, 1993]. An example is the Frame Ontology[Gruber, 1993], defining frame-oriented concepts(class, subclass-of, etc.) without makingclaims about the world. Being the "raw materials"with which further definitions are to be forged,representation ontologies are situated on top of the

194

REPRESENTATION AND GENERIC ONTOLOGIES

/¢

Organizati~otogy

Health Care Organization Ontology

Health Care Economic Health-relatedPayoff Ontology Payoff Ontology

PayoffI Payoff

Ontology Ontology

Figure 5. Extended theoo" inclusion graph.

inclusion graph and constitute the first ontologicalchoice to be made.Generic Ontologies, that include representationtheories, are theories containing definitions of meta-domain or "top-lever’ concepts (state, event,process, part, etc.), like Generic-concepts, or theories expressing general, domain-independent conceptualizations of inference models,like Generic-tasks, schematizing the generalepistemological model we use, the Select and TestModel (ST-Model) [Ramoni et al., 1992].Domain Ontologies are instead conceptualizations(here, theories from Fundamental-medical-concepts to Generic-patient) that specializetop-level definitions in order to model knowledgedomains. For example, disease is a subclass ofclinical-process (defined in theoryFundamental -medical -concepts), that inits turn is subclass of process (defined in

Generic-concepts). According to thecomplexity of internal structure, reflecting theintended use of the conceptualization, domainontoiogies can be distinguished into terminologicalontologies, which mainly play the role of labels to beassigned in the context of term classification;information ontologies, similar to database conceptualschemata, to be used for data organization or retrievalpurposes; gnosiological or knowledge modelingontologies, apt to constitute the foundation or supportthe development of domain knowledge bases to beelaborated by reasoning processes. The actual medicaldomain theories arise from the basic concepts (e.g.disease, finding, therapy) employed byKBSs constructed with the meta-system M-KAT [vanHeijst et al., 1994], though some theories like Drugsare better classified as information ontologies.Task Ontologies contain definitions modeling tasks:Fundamental-medical-tasks models

195

Theory ORGANIZATION¯ Last modified: Frula~.l 7 Jan~rv 1997¯ Source code: organization.lisp¯ List of other known theories

Theory documentation:

Theories included by Organization:

Economic -Concept ¯$oc£o-Concepts

Theories that include Organization:

Hos i t h -Ca~:e -Orgsn:L sat £on

4 ciasses defined:

Or~anlzatxonal-Un~tDrganxzatlon

Responsxble

Or~an~za~1oNal-RoleOr~an~za~onal-TasK

8 relations defined:

Orgsnizst£on. Phyeical-ResouzceOroan£zation.Reapons£bleOroan£¯at£on.StaffOrganizational-Sub-UnitOzoan£zat£onsl-un£t.Physical-Re¯ourceOrgan£zational-Unit.Respons~bleOzosnlzst~onsl-Unit.Staf~St&~f.Htul~D-ResouzceOrgan£zationsl-C~amktment

1 function defined:

Class ORGANIZATIONAL-UNIT

An organization is a corporate or stmilar institution exploiting human and physical resource~and acting ~-~ a collective agent

Subclass-Of: Collecuve-agent

Slots O/" ln.~’tances.

Organizational-Unit.Budget:Slot.CardituHin’: I

Organizational-Unit.Physical-Resource:Mimmum-Slot-Cardinality: l

Organizational-Unit.Responsible:Minimum-Slot-CardimHity’ 1

Organizational-Unit.Staff:Mintraum-Slot-Cardinalztv" I

¯ Defined in theory: Organization¯ Source code: organization.lisp

No instances defined.

Figure 6. Ontolingua reports on the structure of the theory Organ i za ti on and of the classorgani za ti onal -unit.

diagnosis and therapy planning as instantiations of theST-Model, specifying which inferential roles areplayed by instances of categories defined in medicaldomain ontoiogies. For example, instances ofdisease play the hypothesis role withindiagnosis, and the datum role within therapyplanning. As a matter of fact, within M-KAT the ST-Model is used as a general problem-solving method forthe coordination of heterogeneous problem solvers,each charged with the accomplishment of particularinference steps. Representing the knowledgerequnrements of such problem solvers (that is, ofcomputational formalisms) in domain-independentMethod Ontologies within the library would allow thespecification of fine-grained application ontologies(taking into account domain, task and method models)and thus speed up the configuration of new medicalKBSs with M-KAT. As explained in [van Heijst et al.,

1994], such configuration contemplates thedevelopment in sequence of a knowledge-levelepistemological model, including the domain model,its mapping into the task model and its instantiationwith application knowledge, and of a symbol-levelcomputational model, including the set of problemsolvers designated to handle task inferences. Thus theavailability of method ontologies expressing theassumptions underlying the available problem solverscould greatly help the knowledge engineer in assessingthe epistemoiogical adequacy of the latter (that is, thecompatibility of problem solvers with the knowledge-level model previously established), and introduce level of parallelism in the described configurationprocesses.In order to effectively face issues arising fromcoordination/cooperation sessions, the arrangement ofour ontological library is situated "in between" two

196

common cases that can be found in literature, that is"flat", enumerative, fine-grained sets of definitionsand. on the other side, coarse-grained, rarefied, under-specified "top-level" categorizations. Anotheressentml design feature is that the library is intendedto cover multiple domains, in order to account for thepresence of diversely specialized professionals. Forexample, an ontology of basic socio-economicconcepts was introduced to account for economicevaluation analyses of health care activities anddecisions [Quaglini et al., 1996]. Besides, re-utilizingand generalizing some definitions already present inthe theory Clinical-environment (like that ofclinical-unit) [Falasconi arid Stefanelli, 1994]belonging to the core medical ontology, we undertookthe development of an ontology of health careorganizations to be exploited in every transactionmaking explicit reference to agents’ structural roles.The resulting extended library is shown in Fig. 5,while Fig. 6 visualizes some Ontolingua reports abouttheory Organization, giving an idea of itsstructure.

5 Ontology and Terminology Servers

In a multi-agent environment like the one weenvision, ontologies maintained and distributed byontology servers act as open-ended "dictionaries ofwords" describing common application areas andallowing consistency among the programs that have tocommunicate about those areas. Modularity reasonssuggest to partition large ontoiogies into manytheories related by inclusion links (that is. a theory caninclude the definitions of other ontological theoriesthus forming a larger theory). As already mentioned,there will also be multiple ontologies, alternative oreven incompatible ones, to describe the sameapplication area, for example from a differentperspective or with a different granularity [Falasconiand Stefanelli, 1994]. Thus, the ontology server willstore and manage a true library of ontologies madeavailable in a distributed fashion.While ontology specification is best performed using astandard language, the definitions must be enforced ina system-specific way within the various agents. Forexample, a portion of an application ontology may beused to generate, in an automated or semi-automatedway, the database view of the DMA or the graphicalKnowledge Acquisition Tool needed for theconstruction of a RMA. Translation or mappingfacilities are needed from the language used for

ontology specification in the server to differentknowledge/data structures or user-interface graphicalforms.The ontology server differs from a "’data meta-model"agent containing meta-data about the schemata usedby collaborating databases: like a conceptual schema,an ontology provides a logical description of thestored information, allowing different applicationagents to interoperate independently of internalrepresentation structures; but an ontology is not meantto be complete or exhaustive. On the contrary, itshould try to capture the minimal common conceptualmodel for a set of interoperating agents, allowingthem to instantiate, specialize or extend the specifieddefinitions. The ontology server differs also from aterminology server, in that it isn’t intended as anextensive repository of standardized controlledvocabularies, or an intermediary to such vocabularies,but as a repository of special-purpose, thoughreusable, concept definitions. Anyway, while writingan ontology, the consultation of a standard or agreed-upon domain terminology is recommended at least forcomparative purposes.Some attempts to enable computer-assisted medicalterminological modeling and automated classificationwere undertaken, relying on existing extensivestandard medical terminologies [ICD9-CM, 1991:SNOMED, 1982; MESH, 1994]. They oftensuperimpose a well-defined and enforced generalcategorization to "flat" terms repositories. This is thecase of the MED (Medical Entities Dictionary)[Cimino et al., 1994], that reuses and extends thesemantic types of the UMLS [Lindberg et al., 1993] tobuild a hierarchical semantic network of frames withslots, and of GALEN (Generalised Architecture forLanguages, Encyclopaedias and Nomenclatures inMedicine) [Rector et al, 1995], that provides compositional formalism, GRAIL (GalenRepresentation and Integration Language) allowingthe modeller to specify concepts and relations.These two advanced approaches to medicalterminological services show significant commonfeatures. The most evident is the already mentionedsemantic foundation on a central system of formalconcept definitions, explicitly referred to as ontologyin the GALEN effort. The representational frameworkof such terminological ontologies makes directreference to the semantic network model, and usesformalisms and control tools (e.g. consistencycheckers) not tested outside clinical domains. Radical

197

ACL Comm.FaciUty

ONTOLOGY SERVER

Figure 7. The architecture of an ontology server.Agent components are specialized intofunctional submodules addressing ontologycreation, refinement, storage, translation andcoupling with agents.

changes to the core medical ontology would probablyresult into unpredictable ill-functioning of the systems.In contrast with the single conceptual model approachof the above mentioned terminology servers, anetworked ontology server essentially plays the role ofan ontological library manager and distributor. Besidesproviding means for ontology editing, refining andstoring (as done in a distributed collaborative fashionby the Stanford server [Farquhar et ai., 1995]), andbesides being able to preserve application ontologiesconsistent, the ontology server should maintain eveninconsistent ontological theories in the library, keepingtrack of their inter-relationships, also through theexploitation of the previously described categorizationof ontologies.

Besides, the idea of an ontological library accounts forthe coverage of non-clinical domains. For example, inan agent-based D-HIS the specification of someeconomics in an ontological theory complying with thehealth care administrator’s view of the world becomesnecessary. The already mentioned conceptualizationfor health care economic evaluation analysesrepresents a first step in this direction.The choice of Ontolingua, devised as an interlinguaamong multiple representational paradigms andformalisms, largely copes with the structural

complexity and expressive power required forontology specification. As a consequence, enablingthe enforcement of the stated ontologicalcommitments constitutes a main burden on thetranslation modules that must be carefully designedand tested.Fig. 7 shows the main architectural components anontology server needs to perform the activitiespreviously identified. The schema reproduces thegeneral software agent model (see Fig. I) at a higherlevel of detail, emphasizing some functional modulesand interaction paths. Among the elaboration andcontrol agent components, specialized modulesaddress the management of the library ontoiogies andof the application ontoiogies supported within thenetworked HIS, ensuring the correctness andconsistency of the definitions in a theory.The repositories holding these two kinds of ontoiogiesare placed in distinct areas of the agent’s informationstoring submodule, as the Ontological Library will bemainly manipulated off-line, while applicationontologies, once constructed from library theorieswith the proper editor, are destined for constant useand reference in routinary inter-agent transactions.Thus, they are stored in a separate ApplicationOntology Repository.The "on-line zone" of the information repositoriesmay also provide support for another service thesoftware agent could carry out. In a HIS characterizedby a plurality of application agencies, such as theprototypical ones previously mentioned, the ontologyserver could keep track, in an Ontology-AgentDatabase, of the relations between an applicationontology and the agents committing to it (that is, keeptrack of which portion of an application ontology isreferenced or enforced by which agent). For example,it will record which agent is in charge of diagnosticreasoning activities within a physician agency: in ouragency conceptualization, it will store that particularRMA is able to perform the diagnosis task. In thisway, the ontology server could reply to queriesinvolving ontology-agent pairs, and also help routingthe messages along the network on the basis of acontent-level message analysis. The content portion ofa query message, for example, will mention theobjects and relations in a conceptualization shared byasking and answering agents.Also, the organizational roles played by the running

agents, according to the adopted health careorganization model, included as one of the active

198

ontologies (to which each agent commits) in theApplication Ontologies Repository, are intended to bestored in the Database. so that the server is enabled toanswer queries involving agents’ general competenceand their dependence relations. Such structuralinformation could be of great help for performingcontent-independent message routing (that is, basedonly on the desired receiver’s role) or to reduce the setof possible receivers before analyzing messagecontent.Thus, while the library editor manages therelationships among multiple reusable ontologicalaxiomatizations, the application ontology editormanages the relationships between "active"ontologies and members of application agencies,allowing the ontology server to act as a mediator inthe community of software agents [Takeda et al.,19941. Unlike facilitators [Genesereth, 1992], thatgather the entire message handling activity for asubset of agents, the ontology server could carry ononly a partial, semantically founded mediation, asapplication ontologies don’t necessarily constitute fullconceptual models. That is, when an agent doesn’tknow in advance which its proper partners are andwhere they are exactly located on the network (as itmay occur if the agent needs collaboration outside itsapplication agency), it can seek help from theontology server, that, once provided with a list ofontological definitions, or at least with the domainarea of interest, could find out, thanks to the linksrelating application ontologies to original theories inthe library, the possible recipients of the agent’s query.These could be obtained, for example, by identifying,traversing class hierarchies, a common ancestor fordefinitions in two application ontologies, and thenquerying the database for the associated agents.Also, through the exploitation of this functionality, acouple of ontology servers, sharing the sameontology specification language, could act as "front-ends" for their respective HISs, mediating non-localmessage exchanges.Besides the mentioned information types, the databaseshould record, for each ontology-agent pair, theformalism into which the former must be translated inorder to be enforced within the latter. As alreadystated, ontologies must be converted into targetrepresentation schemes (the so-called Ontolingua"implementations") to acquire actual validity inworking agents. If the HIS shows a limited variabilityin the representational formalisms employed, as will

often be the case, the translation modules can beincorporated into the ontology server, and be engagedto properly translate, exploiting information from thedatabase, portions of an application ontology for thecomponents of a HIS agency. Most times, they willperform only a partial conversion, due to mismatchesin expressive power between Ontolingua and targetlanguages. Of course, the outputs of the translationprocesses must be "wrapped" by ACL expressions(identifying e.g. the receivers) before they are sent outon the network. Obviously, such an information flowis intended as bidirectional: the reconstruction ofinverse routes is conceptually straightforward.

6 Conclusions

The integration of multiple, at least semi-autonomousand in general heterogeneous software agents into aMAS architectural framework allows them tointeroperate and thus cooperate within commonapplication areas. Even users can be seen as fullyintegrated (human) agents, to which the software onesoffer services tailored to their specific professionalrequirements. Such a vision entails bridging thedifferent "views of the world" of knowledgeableagents through the commitment to commondefinitions of the conceptual entities and of thetechnical terms employed in knowledge/data bases.We claim that within a multi-agent framework, highlyempowered with respect to the single-system and thedistributed knowledge/data approaches, ontologicallibraries and also standardized terminologicalrepositories should be agentified into ontology andterminology servers providing other agents with thecommon semantic foundation required for effectiveinteroperation. Application ontoiogies configuredfrom library theories can thus be used to modelapplication agencies shaping a particular view on theservices provided by the MAS, such as those ofmultiple DBMSs and KBSs. Depending on the extentto which application ontologies overlap, or on therelationships that can be established among theirdefinitions starting e.g. from information in a commonoriginating ontological (sub)library, they can also used by the ontology server to mediate (in a partial,semantically founded fashion) transactions amongdifferent application agencies.This research effort is just at its earliest developmentstages, as there are many issues still left open.Actually one limitation shown by the applicationagencies implemented [Lanzola et al., 1995; Lanzola

199

et al., 1996] is represented by the adoption of a fixedschema for connecting their component agents. So far,each agent knows in advance which are the other onescooperating with itself and their exact location on thenetwork. So the whole agency fails if any componentfor some reason becomes unavailable at a given time,The institution, within the ontology server, of adatabase storing information on ontology-agent pairswill ease the search of alternative candidates for thedelivery of the requested service.

Acknowledgments. This work is part of the project CASISsupported by the National Council of Research, Italy. It isalso supported by a MURST grant.

References

[Burmeister and Sundermeyer, 1992] Burmeister, B., andSundermeyer, K. (1992). Cooperative Problem-SolvingGuided by Intentions and Perception. In Werner, E. andDemazeau, Y. (Eds.), Decentralized A.I. -3,Proceedings of the Third European Workshop onModelling Autonomous Agents in a Multi-AgentWorld, pages 77-92, Elsevier, Amsterdam.

[Castel franchi, 1995a] Castelfranchi, C. (1995).Commitments: from Individual Intentions to Groupsand Organizations. In Proceedings of the FirstInternational Conference on Multi-Agent Systems(ICMAS95), AAAI-MIT Press, S. Francisco.

[Castelfranchi, 1995b] Castelfranchi, C. (1995). Guaranteesfor Autonomy in Cognitive Agent Architecture. InWoolridge, M., and Jennings, N.R. (Eds.), IntelligentAgents: Theories. Architectures, and Languages.Lecture Notes in Artificial Intelligence, vol. 890, pages56-70, Springer Verlag.

[Cimino et al., 1994] Cimino, J. J., Clayton, P. D.,Hripcsack, G.. and Johnson, S. B. (1994). Knowledge-Based Approaches to the Maintenance of a LargeControlled Medical Terminology. Journal of theAmerican Medical Informatics Association, i:35-50.

[Come and Castelfranchi, 1995] Conte, R.. andCastelfranchi, C. (1995). Cognitive and Social Action,UCL Press.

[de Greef and Breuker, 1992] de Greef, P., and Breuker. J.(1992), Analysing System-User Cooperation.Knowledge Acquisition, 4:89-108.

[Dieng et al., 1995] Dieng, R., Corby, O. and Labidi, S.(1995). Agent-Based Knowledge Acquisition. Steels, L., Schreiber, G., and van de Velde, W. (Eds.),Proceedings of the EKAW94 Workshop A Future forKnowledge Acquisition, Lecture Notes in ArtificialIntelligence, vol. 867, pages 63-82, Springer-Verlag.

[Engelmann et al., 1994] Engelmann. U., Jean, E C. andDegoulet, P. (Eds.) (1994). The HELIOS Software

Engineering Environment. Computer Methods andPrograms in Biomedicine, vol. 45. Supplement.

[Falasconi and Stefanelli, 1994] Falasconi. S., andStefanelli. M. (1994). A Library of Medical Ontologies.In Mars, N. J. I. (Ed.), Proceedings of the ECAI94Workshop Comparison of Implemented Ontologies,pages 81-9 I, Amsterdam.

[Falasconi et al., 1996] Falasconi, S., Lanzola, G., andStefanelli, M. (1996). Using Ontologies in Multi-AgentSystems. In Proceedings of the 10th KAW (KnowledgeAcquisition for Knowledge-Based Systems Workshop),Banff, Canada, 9-14 November 1996.

[Farquhar et al.. 1995] Farquhar, A., Fikes, R., Pratt, W., andRice, J. (1995). Collaborative Ontology Constructionfor Information Integration. Technical Report KSL-95-63, Knowledge Systems Laboratory, StanfordUniversity.

[Finin et al., 1993] Finin. T., Weber, J., Wiederhold, G.,Genesereth, M. R., Fritzson, R.. McKay, D., McGuire,J., Pelavin, P., Shapiro, S. and Beck, C. (1993).Specification of the KQML Agent CommunicationLanguage. Technical Report EIT 92-04, EnterpriseIntegration Technologies. Palo Alto, CA.

[Gasser. 1991] Gasser, L. (1991). Social Conceptions Knowledge and Action: DAI Foundations and OpenSystems Semantics. Artificial Intelligence. 47:107-138.

[Genesereth. 1992] Genesereth. M. R. (1992). An Agent-Based Approach to Software lnteroperability. InProceedings of the DARPA Software TechnologyConference, pages 359- 366, Meridian Corporation,Arlington,VA.

[Genesereth and Ketchpel, 1994] Genesereth, M. R., andKetchpel, S. P. (1994). Software Agents.Communications of the ACM, 7(37):48-53.

[Gruber. 1993] Gruber. T. R. (1993). A TranslationApproach to Portable Ontology Specifications.Knowledge Acquisition, 5:199-220.

[Gruber and Olsen, 1994] Gruber, T. R., and Oisen, G.(1994). An Ontology for Engineering Mathematics. Doyle, J., Torasso, P., and Sandewall, E. (Eds.),Proceedings of the Fourth International Conference onPrinciples of Knowledge Representation andReasoning, Morgan Kaufmann.

[Guarino and Boidrin, 1993] Guarino, N., and Boldrin, L.(1993). Ontological Requirements for KnowledgeSharing. Paper presented at the IJCAI93 Workshop onKnowledge Sharing and Information Interchange,Chambery, France.

[Guarino, 1995] Formal Ontology, Conceptual Analysis andKnowledge Representation. International Journal ofHuman-Computer Studies, 43:625-640.

[Huang et al., 1995] Huang, J., Jennings, N. R. and Fox, J.

200

(1995). An Agent-Based Approach to Health CareManagement. Applied Artificial Intelligence: AnInternational Journal, 9(4):401-420.

[ICD9-CM, 1991 ] ( 1991 ). The International Classificationof Diseases, Ninth Revision, with ClinicalModifications. Fourth Edition. Health Care FinancingAdministration, Washington, DC.

[Labrou and Finin, 1994] Labrou. Y., and Finin, T. (1994). Semantics Approach for KQML - a General PurposeCommunication Language for Software Agents. InProceedings of the Third International Conference onInformation and Knowledge Management (CIKM94),ACM Press.

[Lanzola and Stefaneili, 1992] Lanzola, G., and Stefanelli,M. (1992). A Specialized Framework for MedicalDiagnostic Knowledge-Based Systems. Computers andBiomedical Research, 25:351-365.

[Lanzola et al., 1995] Lanzola, G., Falasconi, S.. andStefanelli, M. (1995). Cooperative Software Agents forPatient Management. In Barahona, P., Stefanelli, M..and Wyatt, J. (Eds.), Proceedings of the 5th Conferenceon Artificial Intelligence in Medicine Europe(AIME95). Lecture Notes in Artificial Intelligence, vol.934, pages i 73-184, Springer-Verlag.

[Lanzola et al., 1996] Lanzola, G., Falasconi, S., andStefanelli, M. (1996). Cooperating AgentsImplementing Distributed Patient Management. InProceedings of the Seventh European Workshop onModelling Autonomous Agents in a Multi-AgentWorld, Lecture Notes in Artificial Intelligence, vol.1038, pages 218-232, Springer-Verlag.

[Lindberg et al., 1993] Lindberg, D., Humphreys, B. andMcCray, A. (1993). The Unified Medical LanguageSystem. In van Bemmel, J. (Ed.), 1993 Yearbook Medical Informatics, pages 41-53, InternationalMedical Informatics Association, Amsterdam.

[MESH, 19941 (1994). Medical Subject Headings. NationalLibrary of Medicine. Bethesda, MD.

[Newell, 1982] Newell, A. (1982). The Knowledge Level.Artificial Intelligence, 18:87-127.

[Quaglini et al., 1996] Quaglini, S., Dazzi, L., Falasconi, S.,Marchetti, M., Stefanelli, M., and Barosi, G. (1996).An Ontology-Based Framework for Building EconomicEvaluation Models in Health Care. In Brender, J.,Christensen, J. P., Scherrer, J.-R., and McNair, P. (Eds.),Proceedings of Medical Informatics Europe ’96(MIE96). Part B, lOS Press, Amsterdam.

[Ramoni et al., 1992] Ramoni, M., Stefanelli, M., Magnani,L., and Barosi, G. (1992). An EpistemologicalFramework for Medical Knowledge-Based Systems.IEEE Transactions on Systems, Man, and Cybernetics,6(22): 1361-1375.

[Rector et al., 1995] Rector, A. L., Solomon. W. D., Nowlan.W. A. and Rush, T. W. (1995). A Terminology Serverfor Medical Language and Medical InformationSystems. Methods of Information in Medicine, 34: 147-157.

[Reddy et al., 1993] Reddy, R.. Jagannathan, V.. Srinivas,K., Karinthi, R., Reddy, S. M., Gollapudy, C., andFriedman, S. (1993). ARTEMIS: a CollaborativeFramework for Health Care. CERC Technical ReportCERC-TR-TM-93-005.

[SNOMED, 1982] Cote, R.A. (Ed.) (1982). SystematizedNomenclature of Medicine, Second edition, College ofAmerican Pathologists, Skokie, IL.

[Szolovits et al., 1994] Szolovits, P., Doyle. J., Long, W. J..Kohane I., and Pauker, S. G. (1994). Guardian Angel:Patient-Centered Health Information Systems.Technical Report, Laboratory for Computer Science,Massachusetts Institute of Technology.

[Takeda et al., 1994] Takeda, H., Iino, K.. and Nishida, T.(1994). Agent Communication with MultipleOntologies. Proceedings of FGCS94 Workshop onHeterogeneous Cooperative Knowledge Bases, pagesI I 1 - 124, ICOT.

[Uschold et al., 1995] Uschold, M., King, M., Moralee, S.,and Zorgios, Y. (1995). The Enterprise Ontology.Technical Report, Artificial Intelligence ApplicationsInstitute, University of Edinburgh, UK.

[van Heijst et al., 1994] van Heijst, G., Lanzola, G.,Schreiber, G., and Stefanelli, M. (1994). Foundationsfor a Methodology for Medical KBS Development.Knowledge Acquisition, 6:395-434.

[vanHeijst et ai., 1995] van Heijst, G., Falasconi, S., Abu-Hanna, A., Schreiber, G., and Stefanelli, M. (1995). Case Study in Ontology Library Construction. ArtificialIntelligence in Medicine, 7:227-255.

[Werner, 1992] Werner, E. (1992). The Design of Multi-Agent Systems. In Werner, E., and Demazeau, Y. (Eds.),Decentralized A.I. -3, Proceedings of the ThirdEuropean Workshop on Modelling Autonomous Agentsin a Multi-Agent World, pages 3-28, Elsevier,Amsterdam.

[Wiederhold, 1994a] Wiederhold, G. (1994). lnteroperation,Mediation, and Ontologies. In Proceedings of theFGCSOB94 Workshop Heterogeneous CooperativeKnowledge Bases, pages 33-48, Tokyo.

[Wiederhold, 1994b] Wiederhold, G. (1994). An Algebra forOntology Composition. In Proceedings of 1994Monterey Workshop on Formal Methods.

[Wooiridge and Jennings, 1995] Woolridge, M., andJennings, N. R. (1995). Intelligent Agents: Theory andPractice. The Knowledge Engineering Review,10(2):115-152.

201