1 mycin cs538 spring 2004 jason walonoski. 2 presentation outline ► history and overview ► mycin...

11

MYCINMYCIN

cs538 Spring 2004cs538 Spring 2004

Jason WalonoskiJason Walonoski

2

Presentation OutlinePresentation Outline

►History and OverviewHistory and Overview►MYCIN ArchitectureMYCIN Architecture►Consultation SystemConsultation System

Knowledge Representation & ReasoningKnowledge Representation & Reasoning

►Explanation SystemExplanation System►Knowledge AcquisitionKnowledge Acquisition►Results, ConclusionsResults, Conclusions

3

HistoryHistory

►Thesis Project by Shortliffe @ StanfordThesis Project by Shortliffe @ Stanford►Davis, Buchanan, van Melle, and othersDavis, Buchanan, van Melle, and others

Stanford Heuristic Programming ProjectStanford Heuristic Programming Project Infectious Disease Group, Stanford MedicalInfectious Disease Group, Stanford Medical

►Project Spans a DecadeProject Spans a Decade Research started in 1972Research started in 1972 Original implementation completed 1976Original implementation completed 1976 Research continues into the 80’sResearch continues into the 80’s

4

Tasks and DomainTasks and Domain

►Disease DIAGNOSIS and Therapy Disease DIAGNOSIS and Therapy SELECTIONSELECTION

►Advice for non-expert physicians with Advice for non-expert physicians with time considerations and time considerations and inincomplete complete evidence on:evidence on: Bacterial infections of the bloodBacterial infections of the blood Expanded to meningitis and other Expanded to meningitis and other

ailmentsailments

5

System GoalsSystem Goals

►UtilityUtility Be useful, to attract assistance of expertsBe useful, to attract assistance of experts Demonstrate competenceDemonstrate competence Fulfill domain need (i.e. penicillin)Fulfill domain need (i.e. penicillin)

►FlexibilityFlexibility Domain is complex, variety of knowledge Domain is complex, variety of knowledge

typestypes Medical knowledge rapidly evolves, must Medical knowledge rapidly evolves, must

be easy to maintain K.B.be easy to maintain K.B.

6

System Goals (continued)System Goals (continued)

► Interactive DialogueInteractive Dialogue Provide coherent explanations (symbolic Provide coherent explanations (symbolic

reasoning paradigm)reasoning paradigm) Allow for real-time K.B. updates by Allow for real-time K.B. updates by

expertsexperts

►Fast and EasyFast and Easy Meet time constraints of the medical fieldMeet time constraints of the medical field

7

MYCIN ArchitectureMYCIN Architecture

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DBPatient DataContext TreeDynamic Data

Static DBRules

Parameter PropertiesContext Type Properties

Tables, Lists

Physician

Expert

8

Consultation SystemConsultation System

► Performs Diagnosis and Performs Diagnosis and Therapy SelectionTherapy Selection

► Control Structure reads Control Structure reads Static DB (rules) and Static DB (rules) and read/writes to Dynamic read/writes to Dynamic DB (patient, context)DB (patient, context)

► Linked to ExplanationsLinked to Explanations► Terminal interface to Terminal interface to

PhysicianPhysician

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DB Static DB

Physician

Expert

9

Consultation SystemConsultation System

►User-Friendly Features:User-Friendly Features: Users can request rephrasing of questionsUsers can request rephrasing of questions Synonym dictionary allows latitude of user Synonym dictionary allows latitude of user

responsesresponses User typos are automatically fixedUser typos are automatically fixed

►Questions are asked when more data Questions are asked when more data is neededis needed If data cannot be provided, system If data cannot be provided, system

ignores relevant rulesignores relevant rules

10

Consultation “Control Consultation “Control Structure”Structure”

► Goal-directed Backward-chaining Goal-directed Backward-chaining Depth-first Tree SearchDepth-first Tree Search

► High-level Algorithm:High-level Algorithm:1.1. Determine if Patient has significant Determine if Patient has significant

infectioninfection

2.2. Determine likely identity of significant Determine likely identity of significant organismsorganisms

3.3. Decide which drugs are potentially usefulDecide which drugs are potentially useful

4.4. Select best drug or coverage of drugsSelect best drug or coverage of drugs

11

Static DatabaseStatic Database

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DB Static DB

Physician

Expert

► RulesRules►Meta-RulesMeta-Rules► TemplatesTemplates► Rule PropertiesRule Properties► Context PropertiesContext Properties► Fed from Knowledge Fed from Knowledge

Acquisition SystemAcquisition System

12

Production RulesProduction Rules

►Represent Domain-specific KnowledgeRepresent Domain-specific Knowledge►Over 450 rules in MYCINOver 450 rules in MYCIN►Premise-Action (If-Then) Form:Premise-Action (If-Then) Form:

<predicate <predicate function>function><object><attrib><value><object><attrib><value>

►Each rule is completely modular, all Each rule is completely modular, all relevant context is contained in the rule relevant context is contained in the rule with explicitly stated premiseswith explicitly stated premises

13

MYCIN P.R. AssumptionsMYCIN P.R. Assumptions

►Not every domain can be represented, Not every domain can be represented, requires formalization (EMYCIN)requires formalization (EMYCIN)

►Only small number of simultaneous Only small number of simultaneous factors (more than 6 was thought to factors (more than 6 was thought to be unwieldy)be unwieldy)

► IF-THEN formalism is suitable for IF-THEN formalism is suitable for Expert Knowledge Acquisition and Expert Knowledge Acquisition and Explanation sub-systemsExplanation sub-systems

14

Judgmental KnowledgeJudgmental Knowledge

► Inexact Reasoning with Certainty Inexact Reasoning with Certainty Factors (CF)Factors (CF)

►CF are not Probability!CF are not Probability!►Truth of a Hypothesis is measured by a Truth of a Hypothesis is measured by a

sum of the CFssum of the CFs Premises and Rules added togetherPremises and Rules added together Positive sum is confirming evidencePositive sum is confirming evidence Negative sum is disconfirming evidenceNegative sum is disconfirming evidence

15

Sub-goalsSub-goals

►At any given time MYCIN is establishing At any given time MYCIN is establishing the value of some parameter by sub-the value of some parameter by sub-goalinggoaling

►Unity Paths: a method to bypass sub-Unity Paths: a method to bypass sub-goals by following a path whose goals by following a path whose certainty is known (CF==1) to make a certainty is known (CF==1) to make a definite conclusiondefinite conclusion

►Won’t search a sub-goal if it can be Won’t search a sub-goal if it can be obtained from a user first (i.e. lab data)obtained from a user first (i.e. lab data)

16

Preview MechanismPreview Mechanism

► Interpreter reads rules before invoking Interpreter reads rules before invoking themthem

►Avoids unnecessary deductive work if Avoids unnecessary deductive work if the sub-goal has already been the sub-goal has already been tested/determinedtested/determined

►Ensures self-referencing sub-goals do Ensures self-referencing sub-goals do not enter recursive infinite loopsnot enter recursive infinite loops

17

Meta-RulesMeta-Rules

►Alternative to exhaustive invocation of Alternative to exhaustive invocation of all rulesall rules

►Strategy rules to suggest an approach Strategy rules to suggest an approach for a given sub-goalfor a given sub-goal Ordering rules to try first, effectively Ordering rules to try first, effectively

pruning the search treepruning the search tree

►Creates a search-space with Creates a search-space with embedded information on which embedded information on which branch is best to takebranch is best to take

18

Meta-Rules (continued)Meta-Rules (continued)

►High-order Meta-Rules (i.e. Meta-Rules High-order Meta-Rules (i.e. Meta-Rules for Meta-Rules)for Meta-Rules) Powerful, but used limitedly in practicePowerful, but used limitedly in practice

► Impact to the Explanation System:Impact to the Explanation System: (+) Encode Knowledge formerly in the (+) Encode Knowledge formerly in the

Control StructureControl Structure (-) Sometimes create “murky” (-) Sometimes create “murky”

explanationsexplanations

19

TemplatesTemplates

►The Production Rules are all based on The Production Rules are all based on Template structuresTemplate structures

►This aids Knowledge-base expansion, This aids Knowledge-base expansion, because the system can “understand” because the system can “understand” its own representationsits own representations

►Templates are updated by the system Templates are updated by the system when a new rule is enteredwhen a new rule is entered

20

Dynamic DatabaseDynamic Database

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DB Static DB

Physician

Expert

► Patient DataPatient Data► Laboratory DataLaboratory Data► Context TreeContext Tree► Built by Built by

Consultation Consultation SystemSystem

► Used by Explanation Used by Explanation SystemSystem

21

Context TreeContext TreePatient-1(person)

Culture-1(curculs)

Culture-2(curculs)

Organism-1(curorgs)

Organism-2(curorgs)

Organism-3(curorgs)

Therapy-1(possther)

Operation-1(opers)

Drug-1(curdrgs)

Drug-2(curdrgs)

Drug-4(opdrgs)

22

Therapy SelectionTherapy Selection

► Plan-Generate-and-Test ProcessPlan-Generate-and-Test Process► Therapy List Creation Therapy List Creation

Set of specific rules recommend Set of specific rules recommend treatments based on the treatments based on the probabilityprobability (not (not CF) of organism sensitivityCF) of organism sensitivity

Probabilities based on laboratory dataProbabilities based on laboratory data One therapy rule for every organismOne therapy rule for every organism

23

Therapy SelectionTherapy Selection

►Assigning Item NumbersAssigning Item Numbers Only hypothesis with organisms deemed Only hypothesis with organisms deemed

“significantly likely” (CF) are considered“significantly likely” (CF) are considered Then the most likely (CF) identity of the Then the most likely (CF) identity of the

organisms themselves are determined organisms themselves are determined and assigned an Item Number and assigned an Item Number

Each item is assigned a probability of Each item is assigned a probability of likelihood and probability of sensitivity to likelihood and probability of sensitivity to drugdrug

24

Therapy SelectionTherapy Selection

►Final Selection based on:Final Selection based on: SensitivitySensitivity Contraindication ScreeningContraindication Screening Using the minimal number of drugs and Using the minimal number of drugs and

maximizing the coverage of organismsmaximizing the coverage of organisms

►Experts can ask for alternate treatmentsExperts can ask for alternate treatments Therapy selection is repeated with Therapy selection is repeated with

previously recommended drugs removed previously recommended drugs removed from the listfrom the list

25

Explanation SystemExplanation System

► Provides reasoning Provides reasoning why a conclusion why a conclusion has been made, or has been made, or why a question is why a question is being askedbeing asked

►Q-A ModuleQ-A Module► Reasoning Status Reasoning Status

CheckerChecker

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DB Static DB

Physician

Expert

26

Explanation SystemExplanation System

►Uses a trace of the Production Rules Uses a trace of the Production Rules for a basis, and the Context Tree, to for a basis, and the Context Tree, to provide contextprovide context Ignores Definitional Rules (CF == 1)Ignores Definitional Rules (CF == 1)

►Two ModulesTwo Modules Q-A ModuleQ-A Module Reasoning Status CheckerReasoning Status Checker

27

Q-A ModuleQ-A Module

► Symbolic Production Rules are readableSymbolic Production Rules are readable► Each Each <predicate function><predicate function> has an has an

associated translation pattern:associated translation pattern:GRIDGRID (THE (2) ASSOCIATED WITH (1) IS KNOWN)(THE (2) ASSOCIATED WITH (1) IS KNOWN)

VALVAL (((2 1)))(((2 1)))

PORTALPORTAL (THE PORTAL OF ENTRY OF *)(THE PORTAL OF ENTRY OF *)

PATH-FLORAPATH-FLORA (LIST OF LIKELY PATHOGENS)(LIST OF LIKELY PATHOGENS)

i.e.i.e. (GRID (VAL CNTXT PORTAL) PATH-FLORA)(GRID (VAL CNTXT PORTAL) PATH-FLORA) becomes: becomes:

““The list of likely pathogens associated with the The list of likely pathogens associated with the portal of entry of the organism is known.”portal of entry of the organism is known.”

28

Reasoning Status CheckerReasoning Status Checker

►Explanation is a tree traversal of the Explanation is a tree traversal of the traced rules:traced rules: WHY – moves up the treeWHY – moves up the tree HOW – moves down (possibly to untried HOW – moves down (possibly to untried

areas)areas)

►Question is rephrased, and the rule Question is rephrased, and the rule being applied is explained with the being applied is explained with the translation patternstranslation patterns

29

Reasoning Status Checker Reasoning Status Checker (Example)(Example)

32) Was penicillinase added to this blood culture 32) Was penicillinase added to this blood culture (CULTURE-1)?(CULTURE-1)?

**WHY**WHY[i.e. WHY is it important to determine whether [i.e. WHY is it important to determine whether

penicillinase was added to CULTURE-1?]penicillinase was added to CULTURE-1?]

[3.0] This will aid in determining whether ORGANISM-1 [3.0] This will aid in determining whether ORGANISM-1 is a contaminant. It has already been established is a contaminant. It has already been established thatthat

[3.1] the site of CULTURE-1 is blood, and[3.1] the site of CULTURE-1 is blood, and[3.2] the gram stain of ORGANISM-1 is grampos[3.2] the gram stain of ORGANISM-1 is grampos

Therefore, ifTherefore, if[3.3] penicillinase was added to this blood [3.3] penicillinase was added to this blood

culture then there is weakly suggestive evidence...culture then there is weakly suggestive evidence...

30

Knowledge Acquisition Knowledge Acquisition SystemSystem

ConsultationSystem

ExplanationSystem

KnowledgeAcquisition

System

Q-A System

Dynamic DB Static DB

Physician

Expert

► Extends Static DB via Extends Static DB via Dialogue with ExpertsDialogue with Experts

►Dialogue Driven by Dialogue Driven by SystemSystem

► Requires minimal Requires minimal training for Expertstraining for Experts

► Allows for Incremental Allows for Incremental Competence, NOT an Competence, NOT an All-or-Nothing modelAll-or-Nothing model

31

Knowledge AcquisitionKnowledge Acquisition

► IF-THEN Symbolic logic was found to IF-THEN Symbolic logic was found to be easy for experts to learn, and be easy for experts to learn, and required little training by the MYCIN required little training by the MYCIN teamteam

► When faced with a rule, the expert When faced with a rule, the expert must either except it or be forced to must either except it or be forced to update it using the education processupdate it using the education process

32

Education ProcessEducation Process

1.1. Bug is uncovered, usually by Bug is uncovered, usually by Explanation processExplanation process

2.2. Add/Modify rules using Add/Modify rules using subset of subset of EnglishEnglish by experts by experts

3.3. Integrating new knowledge into KBIntegrating new knowledge into KB Found to be difficult in practice, requires Found to be difficult in practice, requires

detection of contradictions, and complex detection of contradictions, and complex concepts become difficult to expressconcepts become difficult to express

33

ResultsResults

►Never implemented for routine clinical Never implemented for routine clinical useuse

►Shown to be competent by panels of Shown to be competent by panels of experts, even in cases where experts experts, even in cases where experts themselves disagreed on conclusionsthemselves disagreed on conclusions

►Key Contributions:Key Contributions: Reuse of Production Rules (explanation, Reuse of Production Rules (explanation,

knowledge acquisition models)knowledge acquisition models) Meta-Level Knowledge UseMeta-Level Knowledge Use

34

ReferencesReferences

► Davis, Buchanan, Shortliffe. Production Rules as a Davis, Buchanan, Shortliffe. Production Rules as a Representation for a Knowledge-Based Consultation Representation for a Knowledge-Based Consultation System. System. Artificial IntelligenceArtificial Intelligence, 1979., 1979.

► William van Melle. The Structure of the MYCIN William van Melle. The Structure of the MYCIN System. System. International Journal of Man-Machine International Journal of Man-Machine StudiesStudies, 1978., 1978.

► Shortliffe. Details of the Consultation System. Shortliffe. Details of the Consultation System. Computer-Based Medical Consultations: MYCINComputer-Based Medical Consultations: MYCIN, , 1976.1976.

► Jadzia Cendrowska, Max Bramer. Jadzia Cendrowska, Max Bramer. Chapter 15?Chapter 15?► ““Major Lessons From this Work” Major Lessons From this Work” ► William J. Clancey.William J. Clancey. Details of the Revised Therapy Details of the Revised Therapy

Algorithm. Algorithm. 19771977