Recognition-Based Diagnostic Reasoning

1. Introduction.

Expertise in many diagnostic problem domains depends on both an extensive .. detailed knowledge base about potential faults and their mamfestat1ons and on control structures capable of selecting only those por­ tions of the knowledge base likely to be relevant to the current problem. We have developed a model of dtag­ nostic reasoning for such problems which performs at expert levels through the usc of ~hrcc components. An initial recognition component tnggers a small set of working hypotheses about a fault .. Refinement rules, specific to each triggered hYJ>othests, work to accept, rE>jf:!ct., or modify thf:! hypothf:!sts, or to Sl.Jggf:!st. altf:!rnatf.' hypotheses for consideration. A causal model com­ ponent generates precise expet;~ations about each fault for use by the refinement rules. Successful diagnosis results from good first hypotheses about possible defects, efficient mechanisms for refining these hypotheses, and accurate techniques for determ1rung the plausibility of possible defects .

In fault diagnosis, expertise often consists of the recognition of specific situations as instances of ones that have been seen and successfully cor:rected in the past [1,2,3]. Such recognition-driven expertise is bas~d upon a large repertoire of domain specific kn~wledge m the !orm ol rules Lhal Lrigger models or pol:!Stble faulls . These models in turn contain expectations for testmg whether a proposed model fits the conditions of an observed situatlon. The complexity and level of detail of this . knowledge leads to serious problems with the com­ putational complexity of search processes. Our model focuses comp':ltational resources so that this detailed, Not all fault based task environments lend them­ selves to this treatment. of course. The complexity of possible defects may be such that predefined fault models cannot be constructed. In such cases, more complex causal reasoning i~ often required .

2. Problem Domain.

The above approach to diagnostic problem solving has been demonstrated in a program we call GALEN. GALEN diagnoses cases of congenital heart disease in children. In this domain, GALEN is presented With some observable data about a circulatory system that con­ tains one or more faults or defects. All possible defects that can occur within the circulatory system are assumed to be known in advance, as are the immediate effects they have on the rest of the system. Some sets of defects uniquely correspond to named diseases . It is GALEN's task to identify what disease is present (and therefore what defects are present) given this patient data. The disease whose expectations best match the observed data is selected as the final diagnosis .

3. Knowledge base.

The knowledge base in GALEN consists primarily of an organized collection of hypotheses. The hypotheses specify potential defects, expectations assoctated With these defects, and rules for refining the hypotheses or suggesting other possible hypotheses .

3.1. Hypotheses.

Each hypothesis has two main parts. The first part is a model of the hypothesized phenomena themselves . In congenital heart disease, this is a partial f!10del of a diseased circulatory system. The models provtde expec­ tations about what should be observed if the hypothesis is true. The second part is a group of production rules that describe how to investigate the truth or falsity of the hypothesis. These rules 4.escribe conditi~ns under which a hypothesis should be accepted, reJected, or modified. They also describe when to temporarily aban­ don a hypothesis in favor of another.

Three dillerenl Corms of hypolhesis have proven useful in diagnosing faults in the domain of congenital heart disease . Category hypotheses describe groups of diseases that have some important features in common. For example, the category hypothesis 'diseases.JVitUI1crease~BF' includes all diseases that involve increased pulmonary blood flow. The category hypothesis 'cyanotic..hearL.rlisease' includes hypotheses that could explain why the patient is cyanotic. The

James B . .Moen Computer Science Department models of a category hypothesis describe irr portant expectations that are common to all its member hypotheses . The production rules in a category hypothesis deal primarily with choosing the most appropriate member hypotheses for the given situation.

Disease hypotheses describe individual diseases . Disease hypotheses typically contain several models that represent the state of the circulatory system under several variants of the disease . These variants describe dit!erences in severity (e.g . mild aortic stenosis, moderate aortic stenosis) or the presence of significant co-occurring defects (e.g . aortic stenosis with mitral insufficiency) . The production ru1es in a disease hypothesis act to select an appropriate variant model of the disease, in much the same way that the ru1es in a category hypothesis act to select ail appropriate disease.

Pathophysiological hypotheses describe individual faults l~ithin the circu1atory system. They contain a sin­ gle model of the portion of the system that contains the fault. Pathophysiological hypotheses contain ru1es that trigger diseases (or categories of diseases) which resu1t from the faults the!' describe.

3.2. Hierarchical structure.

An important reason for the success of our model is the use of a hierarchically organized knowledge base which allows the program to consider possible diagnoses at appropriate levels of abstraction. These levels range from very general (categories of diseases), to mid-range (specific diseases), to very specific (defects within diseases).

Describing hypotheses at different levels of abstraction is important in achieving efficiency during di~gnosis. Category hypotheses are included because some data cues · suggest broad classes of hypotheses without necessarily suggesting any specific disease within the class . Further specification of the exact defect can then be postponed until analysis of additional data have been completed, rather than hypothesizing a disease that might have to be rejected later. We can also prune away large areas of the problem space with a minimum of effort if we can reject an entire category of hypotheses, instead of having to reject hypotheses one at a time. At the other end of the spectrum, pathophy­ siological hypotheses are included primarily as a way to organize patient data into mcaningfu1 chunks. We can use a pathophysiological hypothesis as a short form for the complex of data that led to its proposal.

The set of all hypotheses known to GALEN forms a search graph, in which the graph's vertices are the hypotheses themselves and the graph's edges are pro­ duction rules. A given hypothesis is "connected" to another one if the first hypothesis has a rule whose con­ sequent. specifies an operation on the second. Trigger­ ing a hypothesis activates its production rules, which in e!Tect starts a search of the graph at the corresponding point.

4. Control.

GALEN contains two major procedures for recogniz­ ing defects: propose and review. Together, these two procedures implement a highly focused search strategy for moving through the hierarchy of hypotheses. GALEN operates primarily by alternating cycles of propose and review: propose to suggest an initial place to begin searching the hypothesis graph. and review to continue the search until more data is needed.

4.1. Propose.

The propose procedure allows initial attention to be focused into the graph of hypotheses at some appropri­ ate point . This is done by applying rules which look for specific features of the incoming data and t7igger one or more hypotheses if they are successful. When a hypothesis is triggered, a specialized set of rules within the hypothesis are applied which specify t.he most. appropriate model for the hypothesis based on an exam­ ination of the data and .of other hypotheses . For exam­ ple, the disease hypothesis 'tetralogy....oUallot' will choose a model reflecting a 'severe' form of the disease if the po.tient is less than two months old:

(if (data (description (age (lt 2)))) then (specify severe....tetralogy))

These rules may also trig~er additional related hypotheses; the hypothesis 'increasedJ'BF' also triggers the category hypothesis 'diseases.JVith.increasedJ'BF' and the pathophysiologi­ cal hypothesis 'lefUo....righL.tl.ow':

(if (hypotheses (increasedJ'BF}) then (trigger lefUo..ri.ghL.tl.ow diseases.JVith.increas edJ'BF})

The level of abstraction of triggered hypotheses may vary due to differences in the usefulness and specificity of data cues. For example, clubbing of the fingertips can indicate cyanosis, which would resu1t in 'cyanotic...hearL.disease' being triggered. This is a category hypothesis that represents some six possible diseases; the cyanosis data cue alone is not powerful enough to select among them. An early diastolic mur­ mur heard near the pu1monary valve, however, would result in the more specific hypothesis 'insutiicienLpulmonary-Yalve' being directly triggered:

(if (diastolic (time early) (heard .best pulmonary.-area)) then (trigger insutricienLpulmonary_yalve))

4.2. Review.

Once the propose procedure has identified plausi­ ble places to begin searching the graph of hypotheses, the review procedure can carry out a search from there. The review procedure compares a hypothesis' models against accumulated data, then applies ru1es within the hypothesis that examine the re>:ult~ of the comparison. These rules typically look for violated expectations in the patient data and specify new models of the original hypothesis that do not give rise to the same violations. (The ru1es can also accept or reject current hypotheses or trigger new ones). For example, if the disease hypothesis 'tetralogy...,oL.fallot' detected that its model 'severe..ietralogy' expected increased vascu1arity on X-ray but none was observed, a new model, 'mild..ietralogy' would be specified as an alterna­ tive:

(if (unobtained (X-ray ( va:;cularily )}) then (specify milcU.etralogy))

Under some conditions, a hypothesis may have its models compared against other hypothesis' models in addition to being compared against data. The decision to make this extra comparison is made by rules in the affected hypotheses themselves. Such additional com­ parisons can be useful if there is overwhelming evidence in favor of one or more hypotheses, or if there is some important pathophysiological hypothesis that should be taken into account. whenever a new hypothesis is trig­ gered.

5. Expectations.

A third procedure is also critical to Galen's perfor­ mance. The modeler uses two types of causal rules to precompile models of the circulatory system that are associated with hypotheses in the knowledge base . One type of rule determines the hemodynamic results of a defect or abnormality in the circulatory system. Another type of rule determines the observable results of a defect or abnormality. For example, a leaky valve can cause changes in flow and pressure in its vicinity (hemodynamic results) as well as an audible murmur (observable results) .

The modeler ~lar-Ls wilh a model of Lhe normal cir·­ culatory system and adds a set of defects corresponding to those present in a specific disease. Next, all relevant rules are applied to the defects, yielding a circulatory system contu.iniog new hemodynamic u.nd observu.ble abnormalities. Rules are then applied in turn to the resulting abnormalities, until no new rules are applica­ ble . The result is a model of how a real circulatory sys­ tem would appear in the specific disease.

6. Performance.

GAT.F.N's knowledge base contains information sufiicienl to diagnose 70 congenital cardiac diseases and disease variants. It is able to respond to cases covering approximately 95% of the diseases found in the files of the Pediatric Cardiology Clinic at the University of Min­ nesota heart hospital. In a typical run, GALEN inspects 20~30 pieces of patient c:J.ata, distributed in tpe categories of history, physical examination, X-ray, and EKG. Hypotheses are proposed as each item of data is examined. In such a run, 10-20 specific disease models are usually considered. Following the last item of data, a priority rating of these models is established, based upon the degree of fit between GALEN's expectations for each disease and the actual patient data values. Results of validation studies carried out using patient cases from the hospital files indicate that: 1) GALEN is able to reach the same diagnosis as medical staff on the most common forms of congenital heart disease, 2) GALEN is also able to reach a correct diagnosis ·on selected cases of congenital heart disease that were initially misdiag­ nosed by medical staff, and 3) the reasoning steps employed by the program and the expert diagnosticians are similar l4,5].

~owiedgements.

GALEN was implemented on facilities provided by SUMEX-AIM. The model of problem solving in pediatric cardiology is based on earlier work done by David B. Swanson, Paul Johnson and James H. Moller. Richard M. Keller and Robert L. Sedlmeyer provided valuable assis­ ~~nce in the development of GALEN.

BJBlJOGRAPHY

• [ 1] D.B. Swanson, "Computer simulation of expert problem solving in medical diagnosis," Ph.D. dissertation, University of Minnesota, 1978.
• J. Aikins, "ProlcLypes and pr-oduclion rules: A knowledge representation for computer consultations," Ph.D. dissertation, Rep. # STAN-..CSD-B0-814, Stanford University, 1980.
• F. Hayes-Roth, D. A. Waterman, D. B. Lenat, "Principles. of pattern-directed inference systems," in Pattern Directed Inference Systems, F. Hayes-Roth & D. A. Waterman, eds. , Academic Press, 1978, pp. 577·601.
• P. Johnson, A. Duran, F. Hassebrock, J. Moller, M. Priet.ula, "F.xpertise and error in diagnostic reasoning," Cognitive Science, 1981, #5. pp. 235-283.
• P. E. Johnson, F. Hassebrock, ''Validating computer simulation models of expert reasoning," Cybernetics and Systems Research, North-Holland Publishing Company, 1982.