A FRAMEWORK FOR DECISION SUPPORT
4.4.1 Rules for proposing candidates and constructing arguments
Recall the example given in Section 4.2, where the decision procedure was used to generate two candidate solutions - benign change and breast cancer - given a problem of diagnosis where a women presents with a breast lump. Benign change and breast cancer were proposed because they were identified in the domain knowledge as possible causes of breast lumps and the task-specific knowledge about diagnosis
specified causes as appropriate candidates. The generic rule describing candidate proposal can be stated informally as follows:
Candidate Proposal: the candidates to be suggested include every item in the knowledge base which is linked to the focus (the initial dilemma the decision maker faces) by one of the appropriate proposal criteria for the kind o f decision.
The declarative reading of this rule provides an abstract definition of what constitutes a possible candidate. It can be stated formally, in predicate logic, as follows:
If proposaLcriteria( D ecision, Criteria ) and m edical_fact( Criteria, F ocus, Candidate )
Then p rop osed _can d id ates( D ecision, F ocus, Candidate )
An upper-case letter at the start of a word indicates a variable, and all variables are universally quantified. The rule states that the proposition in the conclusion, identifying the proposed candidates, is true for all values of D ecision , F ocu s and C andidate for which the two propositions referring to the proposal criteria and the facts are true, so if the task-specific knowledge includes the fact that causes are appropriate proposal criteria for diagnosis and the medical knowledge includes the fact that breast cancer can cause a breast lump, the rule can be used to generate the inference that breast cancer can be proposed as a candidate in the diagnosis of a breast lump. This results in the variable C andidate being instantiated with the value breast cancer.
The next rule governs the construction of arguments. In diagnostic decisions signs and symptoms are used in constructing arguments. The task-specific knowledge will, therefore, record that signs and symptoms are the appropriate criteria for use in constructing arguments. So, if the knowledge base records that ‘calcifications are a sign of breast cancer’, then this is a possible argument in favour of ‘breast cancer’. Abstracting from this, the generic rule is obtained:
Argument Construction: the possible arguments relating to each candidate include every item in the knowledge base which is linked to that candidate by one of the appropriate criteria for this kind of decision.
The argument in this example depends on the detection pf calcifications. The calcifications are what is termed a ‘finding’, a fact about the patient for which evidence may be obtained. In the Symbolic Decision Procedure described at the start o f this chapter, the evidence for arguments was either provided by the user, in response to a prompt of some kind, or was already available in the patient record. In the extension to the decision procedure described here, evidence is obtained from images. The rules which determine how this is done are called the Interpretation Rules and are described in the next sub-section.
Together, the rules for the proposal of candidates, the construction o f arguments and the collection of evidence describe how decision support is provided. A general rule linking these three together provides - in its declarative interpretation - a statement o f what information is provided as the decision support.
D ecision Support: the decision support consists of all of the evidence for each of the findings used in arguments proposed for and against every candidate sug gested from the specified focus.
D ecision Support: the decision support (the complete set of information provided by the system) consists of all of the evidence for each o f the findings used in arguments proposed for and against every candidate suggested from the specified focus.
If proposed_candidates( D ecision, F ocus, Candidate )
an d argum ent_for_candidate( D ecision, Candidate, Force, Finding )
an d evidence_for_argum ent( D ecision, Patient, Finding, E vidence )
Then decision_support( D ecision, Patient, F ocus, C andidate, Force, Finding, E vidence )
Candidate Proposal: the candidates to be suggested include every item in the knowledge base which is linked to the focus (the initial dilemma the decision-maker faces) by one of the appropriate proposal criteria for this kind of decision.
If proposal_criteria( D ecision, Criteria )
an d m edical_fact( Criteria, F ocus, C andidate )
Then proposed_candidates( D ecision, F ocus, Candidate )
Argument Construction: the possible arguments (with positive or nega tive force) relating to each candidate include every item in the know l edge base which is linked to a candidate by one of the appropriate criteria for this kind of decision.
If argum ents( D ecision, Force, Argum entType )
and m edical_fact( ArgumentType, Candidate, Finding )
Then argum ent_for_candidate( D ecision, Candidate, Force, Finding )
Table 2: the top level o f the Symbolic D ecision Procedure - rules defining what the decision support consists o f how candidates are proposed and how arguments are constructed
In a procedural reading, each rule defines an executable procedure. The conclusion of a rule serves as the ‘procedure call’. In order to generate decision support for a specified patient, decision and focus, the procedure defined in the Decision Support rule is called with the first three variables instantiated:
decision_support( d iagnosis, m s smith, breast lump. Candidate, Force, Finding, E vidence )
The rule functions as an instruction to call the other three procedures:
all possible values. The procedure p rop osed _can d id ates is called with D ecision set to
d ia g n o sis and F ocu s to breast lump, and the Candidate Proposal rule functions as an instruction to call the procedure proposal_criteria, then fact. These procedures consist not of rules but of sets of facts, such as those given in Table 3. So p rop osed _can d id ates
returns breast c a n c er and benign ch a n g e as solutions and argum ent_for_candidate is called for each. If the knowledge base contained only the facts given in Table 3, only one argument would be found: calcifications.
Task-Specific Knowledge: sets of facts defining decisions to which the decision procedure could be applied.
proposal_criteria( d iagnosis, c a u s e s )
argum ents( d iagnosis, supporting, positive sig n s )
Domain Knowledge: the set of facts describing the medical domain.
m edical_fact( c a u s e s , breast lump, breast can cer ) m edical_fact( c a u s e s , breast lump, benign ch a n g e )
m edical_fact( positive signs, breast cancer, calcifications )
Table 3: the task-specific knowledge and the domain knowledge required f o r the example o f a women presenting with a breast lump
The three rules in this section generate all the relevant findings for a given decision: that is, all the pieces of evidence required by the possible arguments which could be constructed for and against each of the relevant candidates. The rules described in the next section are used to establish whether the available procedures for processing the available images provide evidence that the findings are actually present.