DIALOG: A Model Of Diagnostic Logic For Internal Medicine

1. Introduction

1.1. Internal Medicine as a Task Environment for Artificial Intelligence

At first glance, the task environment of internal medicine might appear to be eminently suitable for attack by means of the state-space methods (1) of artificial intelligence. A medical problem may be characterised in terms of an initial state (disease), a goal state (health), and a collection of available operators (drugs, surgical procedures, diets, etc.) by which one state may be transformed into another. The main difficulty with this view of the task is that ordinarily, evidence available at the out­ set provides only a partial description of the initial state of a medical problem. While so­ called 'presenting signs and symptoms' may be suggestive of one or more abnormal or patho­ logical conditions, in most cases, conclusive evidence concerning the underlying disorder is

not available.

Because he can be much more selective in the choice of therapeutic measures when attempt­ ing to treat the cause rather than its manifes­ tations, the clinician is generally obliged to formulate a model, comprising one or more hypothesized abnormal or pathological conditions, as a basis for clinical problem solving. This process of generating and testing hypotheses with respect to unobserved - perhaps unobservable­ pathological conditions is the essence of medical diagnosis. Because this task is one of 'problem finding' or 'problem formulation,' not 'problem solving' as this term is generally construed, it is not clear .that the solution­ oriented heuristic procedures of artificial In certain special cases, where the diagnos­ tic task can be taken to be that of identifying the presence of a single clinical problem, any of a host of deterministic or probabilistic 'recognizer&' can be employed to deal with the problem. (See for example: Ledley and Lusted (2), Gorry and Barnett (3), Bleich (4)J Such methods are of very limited scope, however, as they rest on the assumption that the clinical cases to be analyzed are 'pure'-i.e., there are no erroneous data (sometimes called red­ herrings) to be dealt with, and only one clinico­ pathological diagnosis to be discerned in each case.

In the real world of internal medicine, these assumptions are rarely satisfied. None­ theless, the skilled clinician manages to sort out the facta of a case, disregarding some, while on the basis of others formulating a clinical picture that may comprise a number of interrelated or distinct disease entities.

In what follows, we describe and illustrate a model of the diagnostic process (called DIALOG, for DIAgnostic LOGic) that was designed to mimic the data structures and reasoning processes of the skilled clinician. The process employs a novel heuristic procedure for ranking and partitioning of diagnostic inferences, which enables construction of clinically relevant diagnostic models even in the presence of errone­ ous and misleading data. Using an extensive medical data base which encompasses approximately fifty percent of the major diseases of internal medicine, the DIALOG procedure has accurately analyzed many complex, real-world clinical problems, involving as many as five distinct

disease entities.

In the following sections, we describe first the logic of the diagnostic process and the structure of medical knowledge that underlies the DIALOG data base. There follows a descrip­ tion of the DIALOG procedure, and an illustration

of the system's behavior.

1. 2. The Logic of Diagnosis _

Given a set of observations, which are taken to be consequences of some unknown cause, the task of diagnosis is to devise reasonably cogent hypotheses which would be sufficient to enable derivation, via deductive inference, of the ob­ served consequences. This method of inference, sometimes referred to as abduction, or the method of hypothesis (5, 6), requires the design of procedures and data structures to deal with the following issues:

1. Observations must be able to 'trigger' or evoke hypotheses of disease entities with which they are associated. Since certain findings may suggest one disease more strongly than another, it is necessary to provide for some measure of strength of association between obser­ vations and the hypotheses they evoke.
2. Hypotheses must be able to generate expectations concerning likely consequences, which may be posed as questions regarding ad­ ditional observations (or laboratory ~tudies perhaps) in order to 'test' the hypotheses. This suggests the need for a second set of associations - those relating disease entities to the possible consequences thereof; here, a useful measure of the degree of association is the frequency of occurrence of a given finding in a given disease.
3. Since the business of hypothesis for­ mation is an inexact process (some would call it an art), it is necessary to provide some means for deciding among contending hypotheses. This suggests that the candidates be ranked on some basis, and that criteria be established for deciding when the weight of evidence is suf­ ficient to permit reasonably confident judgments to be rendered.
4. The final and perhaps most important consideration is that some means must be devel­ oped to group hypotheses into mutually exclusive subsets corresponding to coherent problem areas. This is essential in order to be able to deal with cases where more than one disease may be present, and hence more than one hypothesis correct. The discriminating procedure sketched above must not be forced to choose between two complementary hypotheses; somehow, it must focus on those evoked hypotheses that constitute alternatives to one another, not complements.

In section 2, we show how each of these issues is addressed in the DIALOG model. Before pursuing these matters further, however, we turn our attention next to a consideration of the characteristics of the knowledge base underlying internal medicine.

1.3. Structure of Medical Knowledge

There are three significant relations that we have discerned requiring representation in the knowledge base for internal medicine. There are the two relations described above, which may be expressed as:

1. EVOKES (M, D) denoting the evocative association which manifestation M has for disease D, also
2. MANIFEST (n, M) which is the reverse association between disease D and its manifesta­ tion M.

In addition, a number of investigators (7,8) including ourselves, have independently arrived at the conclusion that a hierarchy of disease categories constitutes a significant organizing principle that should be represented in the information structure of internal medicine. Thus we define the relation:

1. FORM-OF (Dl, Dl.l) which asserts that disease (or category) Dl.l is a form of disease category Dl. For example, Dl.l mi~ht stand for hepatocellular disease and Dl for liver disease. Then further elaboration of this region of the hierarchy would eventually lead to hepatitis A, infectuous mononucleosis, and so forth.

The reason that FORM-OF appears to be a significant relation in internal medicine is the economy of representation (and processing) made possible by the following sort of quantification:

Yd(FOAA-OF(Dl,d):>MANIFEST (d, Ml))

expressing the common situation wherein all diseases of a certain class can give rise to the same manifestation,

Because of the enormity of the data base, and because of the need to qualify the EVOKES and MANIFEST relations with respect to strength of associations, it is impractical to express the axioms of internal medicine by means of verbal expressions, as above. Instead, we use an explicit network to encode the hierarchy of disease categories and the related sets of associations, as illustrated in Fig, 1. Here, the solid lines denote the FORM-OF relation; dashed lines (with upward pointers) express EVOKES; and dotted lines (with downward pointers) stand for MANIFEST, The quantified concept that 'all diseases of a certain type share a common manifestation' is denoted in this network by MANIFEST and EVOKES links attached to non­ terminal nodes. D2.1.1

Note that the MANIFEST and EVOKES relations can exist between nodes of the network as well as between nodes and manifestations; this reflects the fact that one disease may be caused by, and hence be a manifestation of another.

In the following section, we give a somewhat more detailed account of the construction and interpretation of the DIALOG information structure.

2. The DIALOG System

2.1. Development of the Data Base

In constructing the disease hierarchy, suc­ cessive general areas of internal medicine such as liver disease and heart disease are selected. Subcategories of each general area are chosen on the basis of similarity of pathogenetic mechanism; thus, members of a subcategory share in good part a common mode of clinical presentation. Further subdivision of subcategories is done until finally the level of individual diagnoses is reached.

Once the superstructure is completed, the appropriate manifestations are entered under each diagnostic level (terminal) node of the hierarchy. The following data are obtained: (1) a list of all manifestations of each terminal node; (2) an estimate of L(Di/Ma): the likelihood that if manifestation Ma is observed in a patient, diag­ nosis Di is its cause (relative to all other causes of Ma). Estimates of this "evoking strength" are given on a 0-5 scale, with 0 indicating that the manifestation is too nonspecific to draw any diagnostic conclusions, and 5 indicating that Ma is pathognomonic for Di; (3) an estimate of F(Mb/Dj), the frequency with which patients with proven diagnosis Dj will display Mb as a manifes­ tation of that disease. A 1 to 5 scale is used, with 1 indicating Mb occurs rarely in Dj and 5 indicating Mb is a sine qua non for Dj.

In addition to listing historical findings, symptoms, signs, and laboratory data as manifes­ tations of a given disease, diagnoses or syndromes themselves can be listed. For example, portal hypertension would be listed under Laennec's cirrhosis; secondary neoplasm of the liver under adenocarcinoma of the colon; and congestive heart failure under aortic insufficiency. These are examples of the unidirectional "causal link": D 1 is a manifestation of D 2 if D2 is a known cause of D 1 ·

Once all manifestations are entered for each of the diagnosis-level nodes, a program is invoked to carry out the 'generalization' process. For each node of the disease hierarchy, the inter­ section of the manifestation lists of its subnodes is determined. The result is a list of the manifestations common to all the subnodes of a given node; it is thus the manifestation list for that higher node. By this method, jaundice becomes a manifestation of hepatocellular injury (and of cholestasis and others), and the presence of a markedly increased alkaline phosphatase indicates that a certain category of cholestatic diseases is to be considered. Further discrimina­ ting information can be obtained only by examining the manifestation lists of subnodes in the class. This process allows the diagnostic program to construct the patient's differential diagnosis on as general a level as is possible; thus, the program can attempt to rule out general classes of disease before working on individual diagnoses.

The final process in constructing the data base is to enter data about each manifestation. Mbst important in this area are the TYPE of the manifestation (history, symptom, sign, laboratory data, diagnosis) and the IMPORT (an integer from one to five). The TYPE allows the diagnostic program to work on the least expensive, least dangerous differential points first (history, symptoms, signs) before going on to the more costly items (there are 3 levels of laboratory data based on complexity, cost and danger to the patient). The IMPORT is an index of how readily an observed manifestation can be ignored. i.e., to what degree it could be considered a "red herring." A history of shellfish ingestion has an IMPORT of 1 and is easily ignored; a liver biopsy showing caseating granulomas has an import of 5 and must be explained by one of the final diagnoses. Various other properties are encoded for each manifestation.

2. 2. The Logic of DIALOG

The DIALOG program has been designed to mimic the problem-solving procedures of the clinician. The program begins by accepting any given sequence of presenting manifestation& of illness. It then inquires about relevant histor­ ical items, symptoms. signs. and laboratory data (proceeding via the TYPEs, exhausting the useful questions in each TYPE before going on to a more coatly TYPE).

2.2.1. Case Analysis: Entry of Presenting Manifestations (Phase I)

As each observed manifestation is entered. nodes of the disease hierarchy 'evoked by' that manifestation are processed, as follows:

• (a) if the indicated node has not previously been considered- i.e ·· if no manifestation has previously caused this node to be evoked - a new hypothetical model is created that reflects the explanatory power of the newly considered disease. Each disease model consists of four parts:
• a list of all manifestations that have been observed but which cannot be explained by this particular disease. This list is referred to as the 'shelf' of that model.
• a list of all observed manifestations. along with associated evoking strengths, that are consistent with this disease.
• a list of all manifestations that would ordinarily be expected to occur (with reasonable frequency) in the presence of this disease, but which have been found absent.

1. a list of all other manifestations that are consistent with the newly evoked node, but about which nothing is yet known.
2. (b) if the node had previously been evoked, then all associated sublists of the model are simply updated to reflect the new observation.

2.2.2. Case Analysis: The Interrogative Phase of DIALOG (Phase II)

Having recorded all of the initial input data, the system proceeds to weigh; the evidence for and against each hypothesis on the evoked list. The weight assigned to each of these is determined by the following factors:

• (a) counting in favor of each model is a factor proportional to the combined evoking strengths of all observed manifestations that it explains.
• (b) counting against a model are two factors: data not explained are weighed in proportion to their IMPORT; data expected but found absent in the patient are weighed in proportion to their frequence of occurrence in the considered disease.
• (c) a 'bonus' is awarded to those models causally linked with disease nodes that have already been con£ irmed.

The set of evoked hypotheses, ordered on the basis of weights computed as above, is next processed to determine which of several modes of analysis is to be pursued. A partitioning process is utilized that employs the following concept of dominance: hypothetical model A dominates model B if the net shelf of A (i.e., the shelf minus those items explained by pre­ viously confirmed diagnoses) is a subset of the net shelf of B. Each member Mi of the evoked models list is compared with the top-ranked model T. If Mi either dominates or is dominated by T, it is placed on the 'considered' list; otherwise it is placed on the deferred list which is temporarily set aside.

This partitioning heuristic has the effect of grouping with the top ranked model those diagnoses that may reasonably be considered mutually exclusive alternatives to it. Since the combina­ tion of any of these models with the top-ranked model would add nothing to the explanatory power of the individual models taken separately, the diagnostic process reduces (for the moaent) to a discrimination among these alternatives.

Once a. 'considered' list is selected, only those diagnoses within a fixed range of the top-ranked model are used to determine which mode of questioning is to be used. When this reduced list contains five or more models, 'RULEOUT' mode is used. 'RULEOUT' asks about manifestations with very high frequency of occurrence in the diseases being processed. Such questions stand a good chance of eliminating one or more of the considered models. The level of questions asked is incremented·via the TYPE, so that inexpensive items are asked first. Because of the high cost associated with the acquisition of laboratory data, 'RULEOUT' mode is not used when the TYPE of questioning has reached the level of laboE&tory procedures. Instead, the key word 'NARROW' is printed out, and the field of considered diag­ noses is artificially narrowed so that 'DISCRIMINATE' mode can be used, which normally applies only when the reduced considered list contains from two to four models. In this mode, the top two diagnoses are selected for discrimi­ nation; items that count heavily for one model while counting heavily against the other are the desiderata for questioning. Finally, if the reduced considered list contains only one model, 'PURSUING' mode is used. Questions are selected that are thought to have a good chance of being 'clinchers.' Manifestations which have a strong evoking strength with respect to the considered model are asked, The system continues in 'PURSUING' mode until either the initial spread between the two top models has reached criterion, or until the spread has been reduced to the point that the top node no longer stands alone on the major list. In the former case, the system 'CONCLUDE's that the considered disease is present; in the latter, processing reverts to the 'DISCRIMINATE' mode.

In each mode, a small number of questions are selected and asked. The responses to the set of queries are processed in a manner e~sentially the same as described for Phase I: new nodes are evoked, old nodes updated, etc. A new ranking of all evoked models is determined; the processes of partitioning, mode selection, and interrogation are then repeated as often as necessary. When­ ever the presence of a particular disease is con­ cluded, the list of manifestations explained by that disease is deleted from further consideration; diseases causally related to the confirmed diag­ nosis are given appropriate bonus scores (dependent on evoking strength and frequency of causal relationship). Phase II processing is then repeated in an effort to discover and confirm additional problems.

2. 3. DIALOG Case Analysis: An Example

Due to space constraints, it is not possible to include a complete protocol of the dialog that took place in analyzing the case discussed below. Rather than include all questions and answers that were exchanged between the physician and machine, we have chosen to highlight the key decision making points of the process. Anyone desiring further information may contact the authors for complete protocols of this and other cases analyzed by DIALOG.

Before turning to the substantive aspects of the transcript, some explanation of the format of the protocol is in order.

As many data as desired can be entered initial­ ly, and the order of entry does not matter. The program prints an asterisk (*) when requesting input, and will continue to do so until the respondent types FINIS to signal that he has finished.

After processing the initial data, the program prints out its current list of 'considered nodes,' prefaced by one of the keywords (RULEOUT, NARROW, DISCRIMINATE, PURSUING, CONCLUDE) discussed in the preceding section. If any data are not explained by the top-ranked model, they are displayed with the notation 'DISREGARDING.'

The program then solicits additional informa­ tion by one of two question types, If it displays the name of a manifestation followed by a question mark, it expects a response of YES, NO, or NA (meaning 'not available'). When it asks 'PLEASE ENTER FINDINGS OF ··· ,' information is being requested concerning a group of manifestations. In response, either all information about the inquired group of manifestations can be entered, or else 'NO' can be typed in order to learn by the next printout what specific question in the group the program had been considering. If the user types in a manifestation name in response to the question, the program will continue to request input (by displaying a '*') until the user types FINIS.

This oase demonstrates hor.J ths program manages the problem of multiple diagnoses in the same patient. The diagnostic problem is easier when more than one diagnosis are causally related, but the machine can diagnose several completely unrelated conditions. In this case, four of the abno:rmal states--pulmonary emphysema, congestive heart failure, congested U.ver, and ths transuda­ tive pleural effusion are interrelated.

• *(DOCTOR)

SYMPT(Jo[S?

*AGE-GTR-THAN-50

• *SEX-MALE
• *CIGARETTE-SMOKING-HX
• *HEART-FAILURE-CONGESTIVE-HX
• *WHEEZING-HX
• *CHEST-PAIN-ANTERIOR-HX
• *DYSPNEA
• *DYSPNEA-EXERTIOMAL
• *ORTHOPNEA
• *COUGH
• *ANOREXIA
• *EDEMA-SUBCUTANEOUS
• *BREATHING-ACCESSORY-MUSCLES-USED
• *RALES-LOCALIZED
• *CHEST- PERCUSSION-HYPERRESONANT-BILATERAL
• *LIVER-ENLARGED
• *HEART-SOUNDS-54-LEFT-ATRIAL-GALLOP

*HEART-SOUNDS-53-LEFT-VENTRICULAR-GALLOP

*SPUTUM-MUCOID

*BREATHING-PROLONGED-EXPIRATORY-PHASE

*BREATHING-INTERCOSTAL-RETRACTIONS-INSPIRATORY

*CHEST-PERCUSSION-DULL-LOCALIZED

*BREATH-SOUNDS-DECREASED-LOCALIZED

*HEART-MURMUR-APICAL-SYSTOLIC-EJECTION

*PRESSURE-VENOUS-INCREASED

*HEART-XRAY-LEFT-VENTRICULAR-ENLARGEMENT .

*HEART-XRAY-RIGHT-VENTRICULAR-ENLARGEMENT

*TUBERCULIN-TEST-NEGATIVE

*FINIS

After the primary data arfJ entsrfJd, thfJ progl'QII'I a lrsady has adequate svidsnae to diagnosfJ congestive heart failure.

(DISREGARDING: CHEST-PERCUSSION-HYPERRESONANT­ BILATERAL HEART-SOUNDS-P2-INCREASED HEART-SOUNDS54-LEFT-ATRIAL-GALLOP BREATHING-PROLONGEDEXPIRATORY-PHASE BREATHING-INTERCOSTAL-RETRACTIONSINSPIRATORY CHEST-PERCUSSION-DULL-LOCALIZED BREATH-SOUNDS-DECREASED-LOCALIZED HEART-MURMURAPICAL-SYSTOLIC-EJECTION)

(CONCLUDE: HEART-FAILURE-cONGESTIVE)

The machine then taokZes the ne:r:t prob Z.., - the one about which it has the most conclusive information. That problem is the differentiation of pulmonary emphysema f1'cm bronchial aathna. This is aaaompUshed with a modest amount of simple clinical information.

(DISREGARDING: HEART-SOUNDS-P2-INCIEASED HEART­

SOUNDS-54-LEFT-ATRIAL-GALLOP CHEST-PERCUSSION-

DULL-LOCALIZED BREATH-SOUNDS-DECREASED-LOCALIZED

HEART-MURMUR-APICAL-SYSTOLIC-EJECTION)

(DllCRIMIRATE: PUUfONARY-EHPHYSEMA BRONCHIAL­

ASTHMA)

PLEASE ENTElt FINDINGS OF EYE-oPHTHALMOSCOPY

(DISREGARDING: HEART-SOUNDS-P2-INCREASED HEART-

SOUNDS-54-LEFT-ATRIAL-GALLOP CHEST-PERCUSSION-

DULL-LOCALIZED BREATH-SOUNDS-DECREASED-LOCALIZED HEART-MURMUR-APICAL-SYSTOLIC-EJECTION) (PURSUING: PULMONARY-EMPHYSEMA)

PLEASE ENTER FINDINGS OF CHEST-PERCUSSIONPULMONARY

(DISREGARDING: HEART-SOUNDS-P2-INCREASED HEART-SOUNDS-54-LEFT-A!RIAL-GALLOP CHESTPERCUSSION-DULL-LOCALIZED BREATH-SOUNDSDECREASED-LOCALIZED HEART-MURMUR-APICALSYSTOLIC-EJECTION)

(CONCLUDE: PULMONARY-EMPHYSEMA)

The ne=t problem considered is the left ventricuLar disease. Simple bedside observations are not productive so the prog~ turns its attention to the problem of pneumonia versus pleural effusion. This process of s!J'itching from one aspect of the overall diagnostic problem to another is one strength of the p:rogram and is determined by the balance of evidence favoring each problem. The deferred left ventricular problem is not permanently disregarded~ however~ and IJ'i n be toe turned to in due time.

(DISREGARDING: HEART-SOUNDS-P2-INCREASED

CHEST-PERCUSSION-DULL-LOCALIZED BREATH-SOUNDSDECREASED-LOCALIZED)

(RULEOUT: AORTIC-STENOSIS ACUTE-MYOCARDIAL­ INFARCTION CARDIOMYOPATHY-PRIMARY IDIOPATHICHYPERTROPHIC-SUBAORTIC-STENOSIS AORTICINSUFFICIENCY)

PLEASE ENTER FINDINGS OF HEART AUSCULTATION

(DISREGARDING: HEART-SOUNDS-P2-INCREASED

HEART-SOUNDS-54-LEFT-ATRIAL-GALLOP HEART-

MURMUR-APICAL-SYSTOLIC-EJECTION)

(DISCRIMINATE: PLEURAL-EFFUSION-TRANSUDATE

GRAM-NEGATIVE-ENTERIC-PNEUMONIA)

PLEASE ENTER FINDINGS OF CHEST-XRAY-LUNG-FIELDS

PLEURAL-FLUID-SPECIFIC-GRAVITY-GTR-THAN-1:013? *YES

(DISCRIMINATE: PLEURAL-EFFUSION-EXUDATE

PLEURAL-EFFUSION-TRANSUDATE)

PLEASE ENTER FINDINGS OF PLEURAL-FLUID-SPECIALEXAMINATION

*PLEURAL-FLUID-PROTEIN-LESS-THAN-3-GMS-PERCENT

(DISREGARDING: PLEURAL-FLUID-SPECIFIC-GRAVITY-GTR-

THAN-1:013 . CHEST-XRAY-LOWIR-LUNG-FIELD-DENSITY[I!S]

HEART-SOUNDS-P2-INCREASED HEART-SOUNDS-54-

LEFT-ATRIAL-GALLOP HEART-MURMUR-APICAL-SYSTOLIC-

EJECTION)

(PURSUING: PLEURAL-EFFUSION-TRANSUDATE)

(CONCLUDE: PLEURAL-EFFUSION-TRANSUDATE)

As would be expected~ the diagnosis of "pleural-effusion-transudate· is est.abUshed after thoracentesis. But note that the evidence is conflicting--the specific gravity is too high for a transudate (probably an actuaZ laboratory el'l'or) but on teaming that the protein concentration is appropriate for a transudate~ the col'l'ect conclusion is cboaJJM.

N~t~ the program turns its attention to the enlarged liver and in this setting has tittle trouble in diagnosing hepatic congestion. The alternative diagnosis~ cardiac cirrhosis~ is a most warthwhite consideration but tittle support ws found for it.

(DISREGARDING: PLEURAL-FLUID-SPECIFIC-GRAVITY-

GTR-THAN-1:013 CHEST-XRAY-LOWER-LUNG-FIELD-

DENSITY[IES] HEART-SOUNDS-P2-INCREASED

HEART-SOUNDS-54-LEFT-ATRIAL-GALLOP HEART-MURMUR-

APICAL-SYSTOLIC-EJECTION)

(DISCRIMINATE: HEPATIC-CONGESTION CARDIAC-CIRRHOSIS)

PLEASE ENTER FINDINGS OF LIVER-FUNCTION-TESTS

(DISREGARDING: PLEURAL-FLUID-SPECIFIC-GRAVITY-

GTR-THAN-1:013 CHEST-XRAY-LOWER-LUNG-FIELD-

DENSITY[IES] HEART-SOUNDS-P2-INCREASED

HEART-SOUNDS-54-LEFT-ATRIAL-GALLOP HEART-MURMUR-

APICAL-SYSTOLIC-EJECTION)

(CONCLUDE: HEPATIC-CONGESTION)

Lastly, attention is zteturned to ths Zeft ventl'icuZal' pztobZem and, aftezt Zeaming of the EKG abnormalities, myocardial infaztction is diagnosed.

(DISREGARDING: PLEURAL-FLUID-SPECIFIC-GRAVITY­

GTR-THAN-1:013 CHEST-XRAY-LOWER-LUNG-FIELD-

DENSITY [ IES] HEART-SOUNDS-P2-INCREASED HEART-

MURMUR-APICAL-SYSTOLIC-EJECTION)

(RULEOUT: ACUTE-MYOCARDIAL-INFARCTION AORTIC­ STENOSIS MITRAL-INSUFFICIENCY CARDIOMYOPATHYSECONDARY MYOCARDITIS IDIOPATHIC-HYPERTROPHIC SUBAORTIC-STENOSIS)

CHEST-PAIN-ANTERIOR-LASTING-GTR-THAN-2o-HINUTES? *YES

(DISREGARDING: PLEURAL-FLUID-SPECIFIC-GRAVITYGTR-THAN-1:013 CHEST-XRAY-LOWER-LUNG-FIELDDENSITY [ IES] HEART-SOUNDS-P2-IRCREASED HEART-MURMUR-APICAL-SYSTOLIC-EJECTION) (RULEOUT: ACUTE-MYOCARDIAL-INFARCTION CARDIOMYo­ PATHY-SECONDARY CARDIOMYOPATHY-PRIMARY · MYOCARDITIS-ACUTE-INFECTIOUS IDIOPATHICHYPERTROPHIC-SUBAORTIC-STENOSIS) PLEASE ENTER FINDINGS OF ELECTROCARDIOGRAM *EKG-LEFT-AXIS-DEVIATION *EKG-P-WAVES-TALL *EKG-QRS-PROLONGATION *EKG-ABNORHAL-Q-WAVES

• *EKG-ST-SEGHENT-ELEVATION

(DISREGARDING: !KG-LEFT-AXIS-DEVIATION EKG-P-

WAVES-TALL PLEURAL-FLUID-SPECIFIC-GRAVITY-GTR-

THAN-1:013 CHEST-XRAY-LOWER-LUNG-FIELD-

DENSITY[IES] HEART-SOUNDS-P2-INCREASED HEART-

MURMUR-APICAL-SYSTOLIC-EJECTION)

Ths patient from whom the data of this case azte derived had died some weeks eaztZiezt and aZZ five conclusions wezte vel'ified by autopsy.

Suaaary

We have presented in this paper an approach to the problem of hypothesis formation, which is central to the task environment of internal medicine and to many other areas of intellectual endeavor aa well. There are, of course, many other ways to attack the problem, some of which have been reviewed in Pople (9),

In our view, the moat significant aapact of the DIALOG approach ia the focusing heuristic that partitions hypotheses into coherent problem areas, thereby inducing system behavior that resembles the 'problem oriented' approach of the skilled clinician.

This approach would appear to have validity in any problem formulation task characterized by the potential concurrence of multiple entities, requiring ad hoc assembly of elementary hypotheses, We commend it to those investigating the 'Theory of Framea'(lO).

Bibliosraphy

1. Nilsson · J. J. · Problem Solving Method a in Artificial Intellisence. New York: MCGraw-Hill, 1971.
2. Ledley, R.s., and Lusted, L.B., Reasoning Foundation of Medical Diagnosis: Symbolic Logic, Probability and Value Theory and our Understanding of How Physicians Reason, Science, 130:9, 1959.
3. Gorry, G.A., and Barnett, G.o., Experience with a Hodel of Sequential Diagnosis. Computer Biomedical Research, 1:490, Hay, 1968.
4. Bleich, H.S. · Computer Evaluation of Acid­ Base Disorders. J. Clin. Invest ·· 48: 1689, 1969.
5. Peirce, c.s. · Collected Papers of Charles Sanders Peirce. C. Harthshorne, P. Weiss. and A. Burks (Eda,) 8 vola., Cambridge, Haas.: 1931-1958 1 eap. Vol. II, pp.272-607.
6. Pople, H.E., Jr., On the Mechanization of Abductive Logic, Proc., 3, IJCAI.
7. Wortman, P.M., "Medical Diagnosis: An Information Processing Approach,'' Comput. Biomed. Rea. 5, 315-328, 1972.
8. Rubin, A.D. · "Hypothesis Forma'tion and Evaluation in Medical Diagnosis," Masters Thesis, Dept. of Electrical Engineering, M.I.T., 1975.
9. Pople, H.E., Jr., Artificial- Intelligence Approaches to Computer-Based Medical Consultation, IEEE Intercon Conference, 1975.
10. Minsky, M. "A Framework for Representing Knowledge," in The Psycholosy of Computer Vision, PaH. Winston, ed., McGrawHill: New York, 1975.