“Expert” Knowledge Module
2 Hour Tutorial for Biomedical Computing Interest Group (BCIG)
Biomedical Informatics Tutorial (BCIG-BITs) December 2002Gary Berg-Cross
Knowledge [email protected]
We need Medical Expertise?Can we represent expert knowledge?
Expert System Structure
UserInterface
EnvironmentLanguage/Shell
ExplanationFacility
InferenceEngine
KnowledgeBase
Blackboard
Separating the KB & IE was quite significant, perhaps the most significant early contribution of expert systems.
In theory we could take the knowledge base & use different inference mechanisms, or take the inference mechanism & use different KBs.
One Definition of Expert System• A computing system capable of representing and
reasoning about some knowledge rich domain, which usually requires a human expert, with a view toward solving problems and/or giving advice. – the level of performance makes it “expert”– Some also require it to be capable of explaining its
reasoning.– Does not have a psychological model of how the
expert thinks, but a model of the expert’s model of the domain.
Categories of Expert SystemsCategory Problem Addressed Prediction Inferring likely consequences of given situations
Diagnosis Inferring system malfunctions from observations, a type of interpretation
Design Configuring objects under constraints, such as med orders
Planning Developing plans to achieve goals (care plans) Monitoring Comparing observations to plans, flagging
exceptions Debugging Prescribing remedies for malfunctions
(treatment) Repair Administer a prescribed remedy Instruction Diagnosing, debugging, and correcting student
performance Control Interpreting, predicting, repairing, and
monitoring system behavior
Knowledge in a Knowledge Base• Knowledge specific to the domain + facts specific to the problem
being solved• A medical KB is defined in HANDBOOK of MEDICAL INFORMATICS as:
– “a systematically organized collection of medical knowledge that is accessible electronically and interpretable by the computer.”
• They note “a medical KB usually:– includes a lexicon (vocabulary of allowed terms) and – specifies relationships between terms in the lexicon. “
• For example, in a diagnostic KB, terms might include:– patient findings (e.g., fever or pleural friction rub),– disease names (e.g., nephrolithiasis or lupus cerebritis) and – diagnostic procedure names (e.g., abdominal auscultation or chest computed
tomography).• Knowledge Representation is the key issue
– Aim is usually to present the knowledge in as "declarative" a fashion as possible
Traditional Feature Comparisons: E/KBS versus ANN
E/KBS• Symbolic• Logical• Mechanical• Serial• Rule Based• Needs “Rules”• Much Programming• Requires Reprogramming• Needs an Expert
Neural Networks• Numeric• Associative• Biological• Parallel• Example Based• Finds “Rules”• Little Programming• Adaptive System• Needs a Database
But much of this too simple, KBS are not really “logical” and can use examples etc.
Processing ComparisonsINPUT(Type)INPUT(Type)
OUTPUTOUTPUTConventional Processing
Procedures applied to data
(Procedural)Data Input
1, 2, 3, etc.Data
Output
Expert/KB System
Inference Engine & Knowledge (Rules)
(Logical)Facts
(A is True)Decision
Recommend
Neural NetworksNetwork Algorithm Identifies
Patterns
(Statistical)Patterns Pattern
recognized
Medical Expert and KB systems• are designed to give expert-level, problem-specific advice in
the areas of :– medical data interpretation, – patient monitoring,– disease diagnosis, – treatment selection,– prognosis, and – patient management.
• Research in medical expert and knowledge-based systems and the development of such systems has been most significant to the broad realm of quality assurance and cost containment in medicine.
One Distinction Between an Expert System and a Knowledge-Based System
• To be classified as an ‘expert system’ the system must be able to explain the reasoning process.
• This is often accomplished by displaying the rules that were applied to reach a conclusion.
Some Basic Concepts• Knowledge representation deals with the formal modeling of
expert knowledge in a computer program.– Important questions in this respect concern the given degree of
structuralization of the medical domain under consideration, the necessity to include vagueness of medical terms and uncertainty of medical conclusions into the chosen formal representation, as well as the extent and completion of the respective knowledge domain.
• Reasoning mechanisms are inference methods which draw medical conclusions from given patient data by means of the stored medical knowledge. – Most important is the selection of the appropriate formal approach with
respect to the given medical domain.– One differentiates methods to infer logical conclusions (e.g., propositional
and predicate logic, three-valued logic, fuzzy logic, non-monotonic logic) and to combine medical evidence (e.g., Bayes theorem, certainty factors, Dempster-Shafer theory of evidence).
Assertional Knowledge• It might be a detailed description of a complex domain
like a disease, a linguistic structure, etc. • This type of knowledge is used to describe a given
clinical situation usually in an object structure.• This is done by associating the different elements or
objects characterizing the context inside the same framework with the consideration of the relationships between these objects.– Example: an exhaustive description of a specific disease
organized following: the set of its symptoms, its possible treatments, medicines, etc.
Alternative KB Approaches
• Rule-based approach– Events trigger firing of rules (condition/action pattern)– e.g. Arden Syntax and Medical Logic Modules (MLM)
• Case & Model-based approach– Create a model (template) of clinical guidelines– e.g. PRODIGY, EON, PROforma, GLIF
But AI is a broad field - a tree representation
Knowledge-base may really include many things
Knowledge-base
HeuristicsHypothesis Rules
Facts
Processes
EventsDefinitions Relationships
Attributes
Objects
user
KBS Editor
Inference Engine
Explanation System
General Knowledge-Base
Case Specific Knowledge-Base
User Interface
may employ:
question & answer
menu-driven
natural language, or
GUI styles
KBS architecture and components
Knowledge AcquisitionModule
Knowledge representation formalisms& Inference
KR Inference* Logic Resolution principle* Production rules backward (top-down, goal directed)
forward (bottom-up, data-driven)* Semantic nets &
Frames Inheritance & advanced reasoning
A Representation: First-Order Logic• Constants: Mr_Smith, Dr._Jones, anemia• Variables: X, Y• Functions: Address(X), Age(Y)• Predicates: Diagnosis(X, anemia); Male(Y); Patient(Z)• Negation: ¬Male(X); ¬Name(X, Smith)• Connectors:
– Conjunction (AND): Patient(X) ∧ Male(X)– Disjunction (OR): Doctor(X) ∨ Nurse(X)– Logical implication: Female(X) ⇒ ¬Male(X)
• Quantifiers:– Universal quantifier: ∀ X (Patient(X) ∨ Doctor(X))– Existential quantifier: ∃ Y (Patient(Y) ∧ Name(Y, Jones))
From Yuval Shahar, “Frame-Based Representations and Description Logics”Temporal Reasoning and Planning in Medicine
Alternatives Ways of Modeling
• X has Diabetes:– Diabetes (x)– Has_Diagnosis (x, “Diabetes”)– Has (x, “Diagnosis”, “Diabetes”)
• Trade off between efficiency and expressiveness– Has (x, y, “Diabetes”)
Relationship Of this K to a DB
• Representing patient X has Diabetes in a table:– Diabetes (x)
• A table called Diabetes with column (s) identifying patient x and a column of the value of Diabetes (x)
– Has_Diagnosis (x, “Diabetes”)• A table called Diagnosis with column (s) identifying
patient x, and diagnosis y and a column of the value of Has_Diagnosis (x, y)
– Has (x, “Diagnosis”, “Diabetes”)• A table called observation with column (s) identifying
patient x, observation type y and observation value z and a column of the value of Has (x, y, z)
Experts typically form sets of rules to apply to a given problemSet of rules reflects the skill of the expert on a topic; use different rule sets to reflect problem-solving competence of expertNeed a strategy to know when to apply them ie use meta rulesRule sets often represented in a tree-like structure with most general, strategic rules at the top of the tree; most specific rules at leaf nodesAdopts a top-down approach to problem-solving, where rule sets only used when appropriate;
reflects human approach divide and conquereases modular developmenteach module may use different representation and reasoning techniques (say for body system)
“Production” Rule sets
21
Rules & Decision Tree ExampleQ1: Test is
Q3: theCost is
Q2: the Panel cost is
Q3: thecost is
Q3: thecost is
Q3: the panel cost is
c1: 30%chanceRule 1
c2: 70% chanceRule 2
c3: 10%chanceRule 7
c2: 70% chanceRule 8
c1: 30%chanceRule5
c2: 70% chanceRule6
c1: 30%chanceRule3
c2: 70% chanceRule 4
included
2K 3K 4K
<y
<5K>=6K
>=x<Easy to find a recipe to turn this into a rule representation.
Depth First not included
Examples of Rule Based Expert SystemsMYCIN - begun in 1972 • Consultation system assist internists in diagnosis and treatment of
infectious diseases: meningitis & bacterial septicemia• When patient shows signs of infectious disease, culture of blood and
urine set to lab (>24hrs) to determine bacterial species• Classified as a "production- rule" system, depth-first, backward
chaining. • Given patient data (incomplete & inaccurate) MYCIN gives interim
indication of organisms that are most likely cause of infection & drugs to control disease– Uses certainty factors to handle incomplete and uncertain information, included
the "how" and "why" capabilities that are now considered essential, defining characteristics of Expert Systems.
• Drug interactions & already prescribed drugs taken into account• Able to provide explanation of diagnosis (limited)
– Thoroughly documented in Buchanan and Shortliffe Rule Based Expert Systems, Addison- Wesley, Reading, Mass., 1984.
Top-level goal rule
IF there is an organism which requires therapy, and consideration has been given to the possibility of additional organisms requiring therapy
THEN compile a list of possible therapies, and determine the best therapy in this list.
THERAPY ruleIF the identity of the organism
is PseudomonasTHEN I recommend therapy from
among the following drugs:1 - COLISTIN (.98)2 - POLYMYXIN (.96)3 - GENTAMICIN(.96)4 - CARBENICILLIN (.65)5 - SULFISOXAZOLE (.64)
THERAPY rule
• The number with each drug is the akin to the probability that a Pseudomonas will be sensitive to the named drug.
• To select the actual therapy, the drugs on the list are screened for contra-indications and to minimize the number of drugs administered, while maximizing sensitivity.
Typical RB Exercise:Write Rules by Diagnosis
• Write rules for patients with the following diagnoses (one at a time):– diabetes mellitus– heart failure– myocardial infarction– benign prostatic hyperplasia
K Engineer compares notes and leads discussion on integration.
Evaluation of MYCIN
• In 1974, an initial study of MYCIN was conducted where five experts approved 72% of MYCIN'srecommendations on 15 actual cases.
• The system was improved and in 1979 MYCIN was again compared to experts.
MYCIN’s Performance Compared to Human Experts
MYCIN 52 Actual Therapy
46
Faculty-1 50 Faculty-4 44 Faculty-2 48 Resident 36 Inf. Dis fellow
48 Faculty-5 34
Faculty-3 46 Student 24
Ratings by 8 experts on 10 cases
Perfect score = 80
MYCIN is not currently in use:
• Knowledge base is incomplete, does not cover a full spectrum of infectious diseases.
• computing power was not available in most hospital wards.
• MYCIN's development lead to the development of "EMYCIN" - for "Empty MYCIN". – To demonstrate this capability, they developed "EMYCIN",
the first shell. – The developers of MYCIN believed that the programming
approaches they used in MYCIN could be applied to other domains.
1 The need justifies cost.2 The (human) expertise* is not available in all situations where it is needed.3 The problem may be solved using symbolic reasoning techniques.4 The domain is well structured and does not require common sense reasoning.5 The problem may not be (better) solved using other (traditional) computing methods.6 Cooperative and articulate experts exist.7 The problem is of proper size and scope. This is relative to resources and evolving technology.
What makes an ES feasiblefeasible ?
Life Cycle for Developing Expert Systems
• Problem Definition• Knowledge Acquisition• Knowledge Representation• Prototype system• Operational system• Knowledge base maintenance
Problem Definition• The essential problem is selecting an appropriate
domain:– the problem must require some type of specialized
knowledge, if there are human "experts" this criteria is probably satisfied
– must not be overly large: define the problem fairly narrowly.
– in business organizations, it should a problem that is handled often enough that an investment is expected to have some payoff: the once every 5 years sort of problem going to payoff.
Knowledge Acquisition
• " the transfer and transformation of potential problem-solving expertise from some knowledge source to a program.”
- Buchanan 1983.
• machine learning - building capabilities into the system that allow it to learn from what it is doing.– the problem of induction - how many instances must be
observed before it can be added to the knowledge base as "true"
• knowledge elicitation - extract the knowledge from the human expert, through some means – direct - interaction with the human expert
interviews, protocol analysis, direct observation, etc.
– indirect - utilize statistical techniques to analyze of data and draw conclusions about the structure of the data.
Knowledge Acquisition (cont.)
Knowledge Representation
• A method to represent the knowledge you are eliciting and/or learning.
• Several major methods –rules, bayes nets, frames • Strengths and weaknesses for each. • None is completely dominant.• Trent is to build heterogeneous systems, that ‘s what
experts are.
Knowledge Representation• A method to represent the knowledge about the
domain• Three major symbolic methods:
– rules– semantic objects– logic
• Although a shell contains a way to represent knowledge, shell selection should be influenced by the matching the representation to the knowledge in the domain.
• Knowledge must be coordinated, so that the knowledge base is consistent.
Prototype system• Typically use an "incremental" development
approach to an expert system. – Build an initial prototype and adjust and expand– Allow the expert to interact with the prototype to
get feedback• Reevaluate if the project should be continued,
if major redesign (knowledge representation) is necessary, or to go ahead.
Build Operational System & Knowledge base maintenance
• Once The actual system is built– New rules can be continually added and old ones
refined/ removed. • This is a tricky process, but there does not
seem to be much literature on it. • One characteristic of an Expert system should
be maintainability, so the ability to add/change/delete rules is essential.
Medical Knowledge (Adjusting to Situations)
Biochemical lab rulesGo from simple, modular to confusing complications
From “Toward Situated Knowledge Acquisition”Tim Menzies,Int. J of Human Computer Studies, 1998
Disadvantages of Production KnowledgeDifficult to maintain for Very Large-KB- One reason is addition of new, contradictory knowledge. Consider
Rule 1. IF it is rainingTHEN not (weather is sunny)
Rule 2. IF location is FloridaTHEN not (weather is cloudy)
Rule 3. IF it is late afternoonTHEN weather is sunny or weather is cloudy
FACTS: it is late afternoon location is FloridaConclude?????Maintenance is to ADD Rule 4. IF it is late afternoon AND location is
FloridaTHEN it is raining
Some observe that RB development never ends….KE is a continuous process…..
KBS as real-world problem solvers
- Problem-solving power does not lie with smart reasoning techniques nor clever search algorithms butdomain dependent real-world knowledge
- Real-world problems do not have a well-defined solutions in literature
- Expertise not laid down in algorithms but are domain dependent rules-of-thumb or heuristics (cause-and-effect)
- KBS allow this knowledge to be represented in computer & solution explained• These are not “logical”
A Semantic Network – beyond the ERA model for real world problems
• A directed graph of vertices (V)and edges (E)where Vi are concepts and Ei,j are relations
Jim
PersonIS-A
Disease
5 Days
Mumps
Has
Duration
DiagnosisPatient
27 years
Age
MamalAKA
Focus is on•categories of objects•relations between those obje
Semantic Networks:Arity of Relations
• Unary relations– Person(Jim): IS-A link
• Binary relations– Age(Jim, 27 years): Age link
• N-ary relations– Disease(Jim, Mumps, 5 days): By creating a reified
disease-relation object with several cases (patient, diagnosis, duration)
Frames (Minksy, 1975)
• A type of Semantic network– Both can be used to represent logic systems– Used to graphically represent taxonomies of
objects and their properties• Concepts have roles, or properties, (also
known in OOLs as slots), such as age• Frames encapsulate more meaningful
chunks of knowledge (e.g., birthday party)
Representing Knowledge in Frames1. Frame Architecture
- A record-like data structure for representing stereotypical knowledge about some concept or object (or a class of objects)
- A frame name represents a stereotypical situation/object/process- Attributes or properties of the object also called slot- Values for attributes called fillers, facets provide additional
control over fillers.
Frame Name:
Class:
Properties: Property 1 Value 1
Object 1
Object 2
Property 2 Value 2
… ...
(1) Class Frame- Represents general characteristics of common objects- Define properties that are common to all objects within class- Static & dynamic property
Static: describes an object feature whose value does not changeDynamic: feature whose value is likely to change during operation
Frame Name:
Class:
Properties: Color Unknown
Bird
Animal
Eats WormsNo._Wings 2
Flies TrueHungry UnknownActivity Unknown
Types of Frames
Subclass Frame- Represents subsets of higher level classes or categories- Creates complex frame structures- Class relationships
Bird
Robins Canaries Sparrows
Bird1 Bird2 Tweety Bird3 Bird4
Class
Subclass
Instance
Anything
AbstractObjects Events
Sets Numbers RepresentationalObjects
Intervals
Places
PhysicalObjects
Processes
Categories
Sentences Measurements
Moments
Times Weights
Things Stuff
Animals Agents
Humans
A Quick Ontological View
Medical Entities Dictionary (MED) Structure
MedicalEntity
Substance LaboratorySpecimen Event
LaboratoryTest
LaboratoryProcedure
CHEM-7PlasmaGlucose
PlasmaSpecimen
AnatomicSubstance
BioactiveSubstance
Glucose
Plasma
Chemical
Carbo-hydrate
Substance
Sampled
Part of
Has Specimen
Substance Measured
DiagnosticProcedure
Multiple hierarchy SynonymsTranslationsSemantic linksAttributes60,000 concepts
1. Generalizations ---- “Kind of” relationship
Bird
Robins Canaries Sparrows
“Kind of” links
2. Aggregation ---- “Part of” relationship
Bird
Wings Feather Eyes
“Part of” links
3. Association ---- “Semantic” relationship
Bird owns
Nest Food
“Semantic” links
(3) Instance Frame- Represents specific instance of a class frame- Inherits properties & values from the class- Able to change values of properties & add new properties
Frame Name:
Class:
Properties: Colour Yellow
Tweety
Bird
Eats WormsNo._Wings 1
Flies FalseHungry UnknownActivity UnknownLives Cage
3. Frame Inheritance
Color Unknown
Bird
Class Animal
Eats WormsNo._Wings 2
Flies TrueHungry UnknownActivity Unknown
Color Black-white
Penguin
Class Bird
Eats FishNo._Wings 2
Flies FalseHungry UnknownActivity Unknown
Lives South_pole
Color Unknown
Canary
Class Bird
Eats WormsNo._Wings 2
Flies TrueHungry UnknownActivity Unknown
- Instance frame inherits information from its subclass frame and also its class.
- Inheritance of behavior, facet-Ease coding & modification Of information
A Frame Representation-types and instances and defaults….
Mammals
Humans
Jim
AKA
IS-ALegs: 2
Age:27
Legs: 4
JohnAge:16
IS-A
Lions
AKA
Bats
AKA
Legs: 2
Bibi
IS-A
Implications of Inheritance• Determination of properties of instances
involves a search of the semantic-network graph
• Default reasoning is enabled– high-level nodes can have values that are
inherited by many lower-level nodes unless these values are overridden
– Exceptions imply a nonmonotonic logic• Multiple inheritance is possible, but might be
ambiguous when conflicts occur
- Exception handling- Frame has property value unique to itself must be explicitly encoded
-Multiple inheritance- It is natural to discuss objects as they relate to different worlds- An instance can inherit information from different parent- Frame structure takes form of a network
MenAge UnknownWeight Unknown
EmployeePhone UnknownSalary Unknown
JackAge 30Weight 78kgPhone 123456Salary 12345
4. Facets- Provide additional control over property values & operation of
the system- Constraint on property values
limit a numeric property value to a rangerestrict data type to Boolean, string or numeric
- Instruction to a property how to obtain value or make reaction to changed value
- Types of facets- Type: defines the type of value that can be associated with the property- Default: defines a default value- Documentation: provides a documentation of the property- Constraint: defines the allowable values- Minimum cardinality: establishes the minimum number of values a
property can have- Maximum cardinality:- If-needed: specifies action to be taken if the property’s value is needed- If-changed: ……... changed
6. Rule interaction- Hybrid system: combine frames and rules for KR- Pattern matching, variables are used for locating matching
conditions among all frames, ?X, ?Age
Humans
JackLegs 1
Age 35Sex MaleResidence BelfastSports SwimLikes Unknown
LucyLegs 2
Age 30Sex FemaleResidence BelfastSports HikingLikes Unknown
BobLegs 2
Age 33Sex MaleResidence DublinSports HikingLikes Unknown
Frame ?Xinstance-of HUMANS
WITH Residence = BelfastWITH Age = ?Age
Frame: JACKinstance-of HUMANSResidence = BelfastAge = 35
Frame: Lucyinstance-of HUMANSResidence = BelfastAge = 30
Populating a frame :Example
• Frame: patient– Attribute1: Patient Name.– Associated action: if the name is unknown
then create a new folder if not, take the already existing folder.
– Attribute2: current date.– Associated action: if the patient is known
then calculate the time interval from the last visit.
• Etc.
Advantages of Frames• Classes and instances organize a flat
knowledge base (unlike FOL) by introducing structure on an epistemological level– E.g., specialization of subclasses through
restriction of a range of values for a property• Simple; easy to understand• Inheritance is captured in a natural, modular
fashion• Efficient inference (e.g., for validation) by
following links, compared to standard logics
Problems with Frames
• Negation cannot be represented– Jim does not have pneumonia
• Disjunction cannot be represented naturally– Jim has Mumps or Rubella
• Qualification is not a part of the language– All of Jim’s diseases are infectious
=> Thus, procedural attachments are often added• The semantics of the links are often not well
defined [“What’s in a Link,” Woods, 1975]
- Disadvantages:Departures from prototypesAccommodation of new situationsDetailing heuristic knowledge
Rule-based Frame-basedRule 1 Frame - BoilerIF Boiler pressure < 50 TemperatureAND Boiler water level < 3 Water levelTHEN Add water to boiler Condition
Rule 2 IF Boiler:Temperature>300IF Boiler temperature > 300 AND Boiler:Water_level >5AND Boiler water level > 5 THEN Boiler:Condition = normalTHEN Boiler condition normal
7. Summarizing Advantages & Disadvantages(Frames vs. Rules)
Features Rule-based Frame-basedOrganization of facts scattered in KB related facts collected and knowledge (but easy to add) represented within a single frame
Inheritance no inheritance Yes-a frame trade-mark
Inference process general rules & PM general rules & pattern matchingcan be slow PM fast
Objects Facets & Message-Passing communication
The Advanced Course: Description Logics
• A subset of FOL designed to focus on categories and their definitions in terms of existing relations
• More expressive than semantic networks• Major inference tasks:
– Subsumption (is category C1 a subset of C2?)– Classification (Does Object O belong to C?)
Critical Factors in M-DDS Failures(after Berner, Luger and Stubblefield)
– Impossibility of developing an adequate database– Lack of an effective set of decision rules (no end)– Lack of “deep” (causal) knowledge of the domain (i.e.
systems do not understand physiology)- Lack of robustness & flexibility. If the knowledge base is
unable to deal with a problem or query not contained within it, it is unable to resolve or adapt a strategy.
- Unable to provide deep explanations- Problems in verification- DSS, in general (unless allied in a hybrid to a CBR, classifier
or neural networks) do not learn from their experience
Arden Syntax and Medical Logic Module
• EMYCIN has been used successfully to develop other systems. But has been overtaken by other approaches such as Medical Logic Modules.– An industry standard maintained by Health Level 7 Org
• Organize decision knowledge as a collection of procedural rules (MLMs) that can be triggered by events
• Each MLM designed to model knowledge required to make a single medical decision such as:– Contraindication alerts, management suggestions, data interpretations,
treatment protocols, and diagnosis scores
• Complex things such as guidelines represented as a collection ofMLMs
Summary
• There are multiple representation formalisms• Frames are a type of semantic networks• A fundamental tradeoff exists in all
formalisms [Levesque and Brachman, 1984], between:– 1. Expressive power of a representation language– 2. computational tractability of inference with it
References• Yuval Shahar, “Frame-Based Representations and Description Logics” Temporal Reasoning and Planning in Medicine
http://www.ise.bgu.ac.il/courses/trp/1,
