From: AAAI Technical Report SS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. The GAMES-IIMethodology for medical KBSdevelopment Giordano Lanzola+ Gertjan van Heijst $ Mario Stefanelli + $Guus Schreiber +University of Pavia, Medical Informatics Via Abbiategrasso209, 1-27100, Pavia, Italy Tel: +39 382 39 1350; Fax: +39 382 42 2881 E-mail mstefa@ipvstefa.unipv.it, giordano@ipvaimedl.unipv.it $University of Amsterdam,Social Science Informatics Roetersstraat 15, NL- 1018 WB,Amsterdam,The Netherlands Tel: +31 20 525 6789; Fax: +31 20 525 6896 E-mail{ gertj an, schreiber } @swi.psy.uva.nl 1 Introduction GAMES-II is an european AIMproject aiming at developing a general methodologyfor the construction of medical knowledge based systems (KBS) based on a general modelof medical reasoning. Such a modelsplits medical reasoning in two subsequent steps. Whilethe first oneis aimedat selecting a set of hypotheses representing possible solutions of the problemat hand, the second one is aimedat testing each of them. Becauseof the paradigmon whichit is rooted, the modelhas been dubbedselect and test model(STModel).It can represent the three generic medical tasks: diagnosis, therapy planning and patient monitoring, by defining the knowledgeroles played by the domainentities in each one of them. The GAMES-IImethodology is based on three principles that haveemergedduring the last decadeof AI research: (i) knowledgelevel modeling, (ii) reusability of both task and domainknowledge, (iii) integration of multiple reasoningtechniques. The knowledgemodeling (Newell, 1982) principle states that knowledgeshould be modeledon a higher level than that of exploited knowledgerepresentation formalisms, to avoid premature design decisions, and to facilitate communicationwith domainexperts. Thesecondprinciple implies that the complexity of KBSdevelopmentmust, just as with any other engineeringactivity, be tackled by the construction of libraries of reusable components (Puerta et al., 1992). Thethird principle is basedon the consideration that the "weakmethods"are too weak, and therefore that KBSshould use multiple specialised 78 reasoning techniques for the different steps in the problemsolving process. (Wielingaet al., 1992) GAMES-IIviews the KBSdevelopment process as composed of two phases. In the first phasean epistemological modelis constructed representing the conceptual modelof the KBSwewant to build. In the secondphase this modelis transformedinto a computational model, a model of a computer implementation to performthat task, preservingthe structure of the epistemological modelas muchas possible. The construction of this modelis a four-step process; (i) construction of a task model, (ii) configuration of a domainontology, (iii) mappingthe task modelonto the domainontology (Ramoniet al., 1992) and, (iv) instantiating the domainontology. Theavailability of libraries of generic task instantiations and libraries of generic ontologiesmakeseasier the construction. Within GAMES-IItools supporting all of these activities have been developed. Weidentify three steps in the construction of the computational model: (i) selection/formulation of the meta rules, (ii) selection of problemsolvers, (iii) representation of the domainknowledgethrough the formalisms used by the selected problem solvers. The availability of the epistemological model ensures that all the relevant knowledgeis present whenthe computational model is constructed. The construction of the computationalmodelis supported by the availability of a blackboardshell. Oneof the main theses of GAMES-II, is that one single modelof reasoning(the STModel) is sufficient to describe all kinds of reasoning found in medical practice. The STModelviews reasoning as a cyclic process of ABSTRACTION of relevant problem features from data, ABDUCTION of potential solutions (hypothesis) from these relevant problemfeatures, subsequent DEDUCTION of expected observations if the hypotheses would hold, and eliminative INDUCTION of hypotheses whose deduced predictions were false. In data rich environmentssuch as clinical environmentsabstraction is the key to successful problemsolving. The GAMES-IIapproach is being used in nine clinical pilot projects, and provedto be successfulin managingpatients integrating several types of problem solvers and exploiting time reasoning. 2 The Computational Model The computational architecture we adopted is extensible. This design choice is required to allow a smoothintegration of different problemsolvers into the same frameworkduring the process of KBSconstruction. Addingnewproblemsolvers accounts for the possibility of using each time the most suitable formalisms for knowledgerepresentation matching the ontological structure of the givendomainat best. Amongall the computational architectures found in the literature, the "BlackboardArchitecture"(BA) Central Module Control Knowledge Sources CmtralBlackboard ~/ Control I]B~ Q KnowledgeSources Problem Solver1 Fig. 1 Problem Solvea" 2, The Distributed Control BlackboardArchitecture. 79 seemsthe mostsuitable in fitting those requirements. BAimposesas its only requirementthe existence of a "Blackboard Structure", simply referred as blackboard, and several" KnowledgeSources". While the blackboardacts as a central repository for the information where the state-of-the-art concerning the problem solution is kept, the knowledgesources incrementally add new information on the blackboard, therefore cooperatingin the process of building a set of plausible solutions for a given case. No control componentis taken into account in the BA since that frameworkis just aimed at providing an abstract conceptual wayof structuring and organizing information within a KBS. The architecture adopted within the GAMES-II project is based on a slightly modified implementation of the Control Blackboard Architecture (CBA) proposed by B. Hayes-Roth(1985). Oneof the major goals of the project is indeedto represent the dynamic flow of reasoning, makingexplicit howdifferent generic medicaltasks need to be solved invokingall the available problemsolvers. Sucha goal is achieved by describing each task through a suitable taskdecomposition hierarchy. A careful description is given about the conditions which must be met for each task to be invoked,as well as aboutthe sub-tasks whichmust be executedto performit. This process is iterated on each subtask until a completedescription of the internal functionality of eachproblemsolver is achieved. Thepresence of a control moduleis therefore essential for achievingthis goal. Moreover,since a completedescription of a task at somepoint of the task decomposition hierarchy will also tackle the internal arrangementof a problemsolver, a modular framework must be adopted which allows an easy integration of different problemsolvers. Accordingto this scheme, which has been dubbed Distributed Control Blackboard Architecture (DCBA),each problemsolver is considered as a separate modulebased on CBA,as shownin Figure 1. It is therefore provided with several local knowledge sources implementing specialized formalisms for representing and operationalizing knowledge.Each moduleis also providedwith a local blackboardable to interface with its knowledgesources. Finally, a local control moduleis used by establishing the mappings betweenthe tasks modeland the ontologyat the problemsolver level, therefore taking advantage of the particular formalismsimplementedby it. As shownin Figure 1 a central moduleis also provided which is always included in each KBS.This module implements the "Frame Ontology" and somehigh level task representation primitives. The choice of forcing the frameontologyto be includedin each KBShas been motivatedby its generality. Actually the need to represent object as classes linked either by class-subclass or class-memberrelationships is alwaysa basic requirementfor representing medical knowledge. The central module also provides the system blackboard, whichis rooted on the frame ontology. This meansthat the basic construct for representing objects within our frameworkis given by the class. Classes are described by a set of features, and features mayhave several attributes, each one able to accept multiple values characterizing it. Locatingthe systemblackboardin the central module results in having it module independent, therefore allowing an easy wayfor modulesto communicate and exchange data amongeach other. 3 Problem Solvers Problemsolvers are used for implementingdifferent representational formalismscombinedwith a reasoning technique. Fromthe point of view of the computational architecture they maybe considered as black boxes able to inspect the blackboard. As a result of their invocationthey mayreturn suitable values to be stored back on the blackboard. Theconfiguration of problemsolvers into an executablesystemis supported by M-KAT (Medical KnowledgeAcquisition Tool), a frameworkimplementingand operationalizing D-CBA.Oneof the key issues of the whole process is the cooperation amongdifferent problem solvers. This meansthat the results providedby one of themwill be used by another one and so on. Sucha cooperation is madeavailable through the uniform formalism for data representation implementedby M-KAT. In a very general scenario, different problemsolvers should be available for performingthe sametask. Users should be allowedto specify the one to be used according to somecriteria concerning computational complexity,explanation, representation capabilities, and so on. Obviouslyfor this to worka set of translators should be implemented within each problem solver so that eachone of themis able to translate the same application knowledgefrom a generic representation formalism into its own representation 8O schema.Several research efforts are currently being addressed to this topic, whosemost important are probably represented by Ontolingua (Gruber, 1993) and KIF(Geneserethet al., 1992)projects. However, despite these efforts, the availability of suchtools is still to come.Dueto this problemin the current implementation each problem solver mayalso replicate part of the domainknowlegdefor its ownpurposes. This is neededfor representing knowledgein a more appropriate wayfor the specific formalismsprovided by the problemsolver itself. Several problem solvers have been implemented so far, and a short descriptions for each one of them will be provided in the following. Qualitative Abstractionis one of the simplestavailable. Its aimis that of extracting froma large set of data a lowernumber of problemfeatures from whicha plausible solution can be inferred. Starting from the values of several findings by meansof lookuptables the system is able to comeup with a symbolicvalue expressing the same information content in a more synthetic way. NumericalAbstraction maybe considered as a simple formula evaluator, used wheneveran algebraic expression must be evaluated starting from the numerical values taken by some component evidences. Episodic Abstraction is aimed instead at identifying a set of episodes starting from a time sequenceof data and accordingto someuser specified criteria. The available criteria, which obviously dependon the data type of the componentevidence, include: increasing, decreasing, constant, range, pattern of values,set of valuesetc... Production Rules is one of the most important problem solvers available. Besides being used for modeling domainknowledge,it has been chosen for implementing control knowledge. Rules are structured as usual, with an IF part including a statement whichmustbe satisfied for their THEN part to be activated. Oneof the mainfeatures providedby this problem solver is represented by the possibility of assembling complex clauses starting from elementary statements, therefore simplifying the wholerule editing process. This has been accomplishedthrough the identification of a newformalism based on the idea of operator. Eachoperator maybe described as an entity able to convert a statement expressed through a humanunderstandablesyntax into the corresponding computer executable piece of code, so that there is always an exact matchingbetweenthe ~.--~T -SHOW-DOSAGE /THERAPEUTIC-INTERPRFTATION<~...... T -SHOW -JUSTIFICATION fINAL -INTERPRETATION’~... ~"DIAGNOSTIC-INTERPRETATION D-SHOW -EVOLUTION /MONITORING -ROOT // //., / CORTISON -RESPONSE T-COLLECT-LEAVES /T-STRUCTURING-SOLUTION -SPACE //THERAPV -ROOT~-’--~T -ABSTRACTION jT-EP,SOD,C-AHSTRAGT,ON T -CVCLIC-ABSTRACTiON~’~.-"--T-TABULAR -ABSTRACTION ~.--~T -FOCUSED -ABDUCTION--.._ ~"~T -FORMULA-ABSTRACTION ~ "T-ABDUCTION~..,._ ~’~ ~’T-UNFOCUSED-AHDUCTION ~T-PURSUE-FOCUS PATIENT-MANAGER~ ~ \ .~-COLLECT -LEAVE5 YTRuCTUHE-SOLUTION-SPACE~:I:PV::~:::;ULT‘ "EQUEST-EVIDENCES<’’REQUEST-ADSTRACTION"’flEQUEST-FI.DING ’AONOS’S-ROOTk-----_AHDUOT,ON --OCUSED-ADDUCT’ON -"UNFOCUSED-ABDUCTION "~ ~ADSTRACTION PURSUE -FOCUS _.,..,--’TABULAR-ABSTRACTION CVCLIC - ABSTRACT ION’C’---........f ORMU LA -ABSTRACTION Fig.2 The task network in a KBSdeveloped through M-KAT. declarative statement madeby the user and the system interpretation. The attempt wepursued has been to providethe user with the very basic buildingblocks neededto express the ill-shaped knowledgeinvolved in medical reasoning. Each rule is therefore composed of several clauses, whereeach clause refers to a particular value of a feature attribute, checksfor relationships existing in the ontologyor related to the inference model,invokesthe executionof a task etc... Additionally, a clause mayjoin toghether several subclauses into a more complexstructure. For implementing each clause a different operator has been devisedable to supplythe systemwith all the relevant information about the syntax and semantics of that clause. As a first attempt to integrate external problem solvers into the architecture we used an Influence Diagram developed with the GAMEES environment (Bellazzi et al., 1991). This problemsolver allows the definition of a networkcomposedof different nodes connectedthroughinference relations. Threetypes of nodes have been implementedso far: decision nodes, chancenodesand utility nodes. Additionalresearch is being carried on concerning several other problemsolvers. It is worthwilementioning QCMF (Ironi et al., 1993),a tool that supports the construction of qualitative modelsof pathophysiological systems. The underlying ontology is based on the notion of compartments,and the dynamicsof the system are modeledthrough flows representing transfers between compartments. QCMF is associated with a simulationalgorithmable to support rapid prototyping, and therefore it can be used as a standalone tool for the construction of knowledgebased systems. 4 Modeling Tasks with Metarules Within the computational model developed, metarules were adopted as the basic formalism for implementing inferential knowledge (Lanzola and Stefanelli, 1993). Metarulesprovedto be very useful in associating control situations with actions to perform. Moreover, this formalism showeda semantics whichcould be easily madeoperational into the system. By meansof a special parser each clause can be directly translated into the matchingpiece of code executable by the system. Thelayout of a rule is structured as usual. Eachone of themis composed both of a left-hand side and of a right-hand side. Theleft handside consists of a conjunction of inquiries over the control blackboardor over the domainblackboard dependingon the particular kind of metarule, while the right handside triggers the activation of a possible action upon the satisfaction of its premise. Morespecifically two kinds of metarulescan be distinguished. Thefirst kind ofmetarulesdeals with the sequencing of the several inference types involved in the 81 inference modelof the system. Theyrepresent possible strategies to adopt as a consequence of situations which mayoccur within the control blackboard. Theirleft parts therefore includetests abouteither the executionstatus or possible results achievedby previously executed tasks and substasks, while their right parts are able, uponsatisfaction of their premises, to trigger the activation of a newtask. Their purposeis to give rise to a task-subtask decomposition hierarchy starting fromthe top level task, whose goal is to find a solution for the given medicalproblem, and going on each time with simpler sub-tasks whoseexecution is required in accomplishing the super-ones. Metarules are therefore grouped into classes, whereeach class is linked to others, so forming a network. Figure 2 showsthe rule class network available in a KBSfor managingpatients affected by Acute Myeloid Leukemia developed with M-KAT. Each class includes all the rules able to performa giventask. Thesecondkind of metarulesis usually located at the bottom level of a task-subtask decomposition hierarchy and deals with the process of establishing a mappingbetweenthe inference modeland the ontology. Those rules are used to provide a detailed description of howa particular task can be implementedwith references to the activation of domain specific knowledge sources. Their left sides therefore deal with either inferential or ontologicalelements,in order to successfullyprovideinstances for their right sides whichwill trigger the activation of particular areas of the ontology useful to accomplish the selected task. Rulesbelongingto this class proveto be very useful in giving the user the possibility of describing a task in a highly detailed way, allowing him to exactly specify which ontological entities should be exploited to achievea goal. Whena newsystemis being built fromscratch, the basic set of metarulesprovidedby the systemis simply aimedat implementinga generic instance of the STModel.This metarule set will be able to implement all the functionalities providedby the problemsolvers selected duringthe initial configurationstep without exhibiting any particular domain dependent strategy. Theuser is then left the job of analyzingand customizing these rules according to the specific application needs. 82 5 Conclusions This paper addresses one of the most exciting issues in AI, that is knowledgeacquisition based on the idea of knowledgesharing and reusability. The approachwe adopted wasaimed at identifying a commonlanguage as the basis for defining conceptual modelsof problemsolving methods. That formalism provesto be successfulin givingthe user the possibility of modelinghis strategies within the inference modelof a KBSin a straightforward way. The immediate contribution of this research is therefore a methodfor providing KAtools also with powerful capabilities allowing an end user to modelhis own strategies within a KBSin a context free situation. 6 References Bellazzi R. Quaglini S. Berzuini C. and Stefanelli M. (3amees: a probabilistic environment for expert systems. Computer Methods and Programsin Biomedicine, 35:177-191,1991. Hayes-RothB. A Blackboardarchitecture for control. Artificial Intelligence, 26:251-321, 1985 (3enesereth M.R. and Fikes R. E. KnowledgeInterchange Format, Version 3.0 Reference Manual. Technical Report Logic 92-1, ComputerScience Department, Stanford University, 1992 (3ruber T. A Translation Approachto Portable OntologySpecifications. KnowledgeAcquisition, 5:199-220, 1993 Ironi L. Cattaneo A. and Stefanelli M. A tool for pathophysiological knowledgeacquisition. Proc. of the 4th European Conferenceon AI in medicine, 13-31 (IOS Press), 1993. Lanzola (3. and Stefanelli M. Inferential knowledgeacquisition. Artificial Intelligence in Medicine,5:253-268,1993 Newell A. The KnowledgeLevel. AI Magazine, 3:1-19,1982 Puerta A.R Egar LW. Tu S.W. and Musen M.A. A multiplemethodknowledge-acquisitionshell for the automatic generation of knowledge-acquisition tools. KnowledgeAcquisition, 4:171-196, 1992 RamoniM. Stefanelli M. Barosi (3. and MagnaniL. An epistemological framework for medical knowledge based systems. IEEE Transactions on systems, Man and Cybernetics, 22(6): 1361-1375,1992. Wielinga B.J. Schreiber A.Th. and Breuker J.A. KADS:a modeling approach to knowledgeengineering. Knowledge Acquisition, 4:5-53, 1992. Acknowledgements This work is part of the AIM project A2034 entitled "GAMES-H: a General Architecture for Medical Expert Systems", and was supported by the European Commission.It is also supported by a MURST grant.