~‘ompul. Bid Med Pergamon Prw 1978. Vol. 8, pp. 2540. Printed !n Great Britain.

GLAUCOMA CONSULTATION BY COMPUTER*

SHOLOMWEISS
Institute of Computer Science, Mt. Sinai School of Medicine, CUNY, NY
CASIMIR A. KULIKOWSKI
Department of Computer Science, Rutgers University, New Brunswick. NJ
and
ARAN SAFIR
Institute of Computer Science, Mt. Sinai School of Medicine, CUNY. NY

(Receioed 12 Junuary 1977)

Abstract-This paper describes a computer-based system for consultation in the diagnosis and
therapy of glaucoma. The reasoning procedures interpret the findings of a particular patient
in terms of a causal-associational network (CASNET) model that characterizes the pathophysio-
logical mechanisms and clinical course of treated and untreated diseases.
The major new features of this program are: (a) generation of complex interpretations from
a qualitative model of a disease process; (b) reasoning about detailed follow-up management
of a patient; (c) incorporation of alternative expert opinions about subjects under debate; and
(d) its testing and updating by a collaborative computer-based network of glaucoma researchers.

Computer-based medical consultation Artificial intelligence Glaucoma Causal-

associational network model Collaborative computer networking

INTRODUCTION

Consultation with a specialist is often sought by a practitioner who encounters a particu-

larly complex case, or one that fails to respond to conventional courses of therapy.
The consultant brings to bear detailed and up-to-date information from research and
clinical practice in several principal tasks: elicitation of patient history, performance
of the physical examination, selection of tests to be performed, and the subsequent
interpretation of the case and recommendations for therapy.
A computer consultation system that incorporates the knowledge and experience of
“expert” clinicians can be a flexible and useful tool for providing advice in at least
three of the above tasks. It can ensure that all the important medical history questions
are asked of the patient, and can suggest the performance of appropriate tests. It can
be provided with reasoning strategies that can be adjusted to reflect the most advanced
medical thinking and recent experience in a field. The tasks that it cannot simulate
well are those in which the clinician’s individual examining skills play a major part
in extracting useful diagnostic information. Although experts can incorporate into a
program some advice on examining procedures, they cannot in this way impart their
skills to other practitioners using the program. The examinations performed by these
practitioners may be carried out by techniques of observation different from those used
by the expert. However, with the growing reliance on standardized tests in many special-
ties, a consultation program may more closely encompass the tasks of the consultant.
In addition, the development of computer-based consultation systems can have a benefi-
cial effect in stimulating discussion and ultimate agreement among experts on standards.
These must be set for observational criteria, interpretation of intermediate test results,
the explication of concepts and the refinement of taxonomies in the medical specialty.
Almost from their inception, computer consultation systems have been envisioned
as repositories for the knowledge and expertise of medical specialists [l]. Most decision-
making programs have been concerned with providing diagnostic advice for a single
consultation session. Considerable experience has been developed in the application
* This research was supported in part by Grants: RR-643 of the NIH Biotechnology Resources Program
and l-ROl-MB-00161 of the Health Resources Administration. HEW.
26 SHOLOMWEISS, CASIMIRA. KULIKOWSKI and ARAN SAFER

of methods that summarize experiential knowledge as a series of statistical associations

between diseases and their manifestations [2, 31 or incorporate expert knowledge into
decision rules [4]. Practical data limitations usually restrict the use of statistical methods
to the characterization of disease as static “states” of a patient rather than dynamic
processes. Although decision tree structures may incorporate factors related to the time-
course of a disease, they provide no systematic means of reasoning with an explicit
representation of the process.
In the past few years, artificial intelligence (AI) ideas and methods have been proposed
and applied to problems of medical decision-making [S-9]. We have developed an
approach that uses a causal-associational network (CASNET) model of disease to rep-
resent the medical knowledge to be used in reasoning about a given patient [S, 9, lo].
In this paper we report on a consultation program for the diagnosis, prognosis, and
therapy of glaucoma based on a CASNET model. The program provides consultation
for complex clinical cases, including those with involved histories and multiple follow-up
visits. An important characteristic of our system is that it can present alternative
opinions derived from different consultants. To obtain a variety of opinions we have
established a national network of collaborating investigator-consultants who share in
the development and testing of the program [ll]. One of our goals is to develop a
mechanism for the routine inclusion of significant new results from current research.
Because these are included in the program as they are produced by clinical researchers,
they enhance the program’s capabilities as a teaching tool as well as a consultant.
Our approach has been to separate the model of disease from the decision-making
strategies. Although the data base of medical knowledge constantly changes to incorpor-
ate current research, the strategies of decision-making should be able to “roam over”
that data base without needing change. We also wanted a model of disease that could
generate strategies for explanation, question-answering, teaching, and testing as well
as those for diagnoses, prognoses, and therapy. Our representation is best suited to
qualitative descriptions of diseases as dynamically evolving processes. We use a causal
network of events to express the mechanisms of the disease. This formal framework
for the prediction and assessment of the time course of the disease before and after
treatment has proven to be both an efficient and realistic representation for several
types of diseases.
Members of our collaborative network present the computer consultation program
with a variety of complex cases, and weigh its performance against their own judgements.
Their subsequent suggestions are used to refine the diagnostic and therapeutic recom-
mendations, to improve the system’s assessment of signs and symptoms, and to perfect
specific techniques for acquiring and displaying clinical data. A model-building and
editing program for designing CASNET-type models has also been developed [12].
This facilitates the incorporation of new information into the glaucoma model or other
models of disease. To establish a data base composed of the glaucoma cases presented
to the system, we have designed data analysis programs [13] compatible with our
models. These permit a review of individual cases over time, as well as the selection
of groups of cases according to specified conditions. The facilities for summarizing case
information are designed to provide data in a form useful for updating and improving
the model.
Our programs currently use time-shared PDP-10 computers at Rutgers University
and at the SUMEX-AIM shared resource facility at Stanford University. The ophthalmo-
logical investigators, located at the Mt. Sinai School of Medicine, Johns Hopkins Univer-
sity, Washington University, the University of Illinois at Chicago, and the University
of Miami, access the SUMEX-AIM computer over a nationwide computer communica-
tions system (currently the TYMNET system).

BACKGROUND
Our fundamental approach has been to develop methods for describing diseases in
terms of models that incorporate current medical knowledge from a variety of sources.
 Glaucoma consultation by computer 27

Because we needed a model that could accommodate detailed descriptions of the disease
process, we restricted our initial design to a medical problem that is well-defined and
circumscribed, yet important enough to elicit interest from the clinical research com-
munity. Glaucoma was chosen with these constraints in mind. The eye can be adequately
described in terms of relatively few anatomical structures, and glaucoma is a disease
whose mechanisms are largely limited to the eye. Yet, glaucoma is a leading cause
of blindness in the United States, and a disease whose subtleties are often overlooked
until irreparable loss of vision has occurred.
The mechanisms of glaucoma are sufficiently well known that they can be used to
explain most of the observed patterns of patient findings in terms of causal models,
at least to a first approximation. Although some aspects of the disease are as yet not
well understood, those mechanisms that are known provide a rational framework for
diagnostic interpretation and therapeutic planning. In glaucoma, understanding of the
different alternative mechanisms and causative factors directly affects choices of thera-
pies. A clinically useful in-depth model can thus be developed. In contrast, a less well-
understood disease; such as uveitis, would be a poor choice for developing a prototype
model, because it follows a more varied and unpredictable course for which detailed
mechanisms are as yet poorly understood. Such a disease can be described best in
purely associational terms, for which other decision models may prove more appropriate.
Glaucoma’s causal structure is relatively easy to elicit from experts and from glaucoma
texts [14. 151. A summarized disease description can be designed from these. However,
significant methodological and practical problems arise in associating this descriptive
model with the inferential components needed for decision-making. These inferential
components (usually stated as implicational rules) relate qualitatively different elements
in the model: observations, intermediate hypotheses of pathophysiological states, and
higher level disease hypotheses. Maintaining consistency between a large number of
inferential rules and a related causal structure becomes difficult as the model grows
and becomes more complex by the addition of knowledge from various experts. Our
CASNET model [lo] overcomes these difficulties, in part, by including only those
elements appropriate for consultative reasoning. The scheme is flexible enough to handle
a variety of non-causal relations. It is here that the model designer has the greatest
latitude in building the information base for a specific domain such as glaucoma.
An important aspect of our approach is that we consider diagnostic interpretation
to bc much more than the simple assignment of a patient to some pre-specified category.
Evaluation of the patient’s status is an on-going dynamic process. The patient’s clinical
status is re-evaluated on successive visits as changes in the presenting signs occur. The
causal model summarizes the findings, and guides in the construction of diagnostic
and prognostic hypotheses. These hypotheses may be simple hypotheses. such as “very
elevated intraocular pressure,” or more complex hypotheses (composed of a set of related
simple hypotheses), such as “chronic angle closure glaucoma”. All hypotheses may in-
clude modifiers that further specify a condition by its intensity, duration, progression,
topographical distribution, or other characterizing features. Likewise, treatment recom-
mendations are not viewed as unique choices for either medical or surgical procedures.
Many problems may be encountered in the same patient at different stages of the disease.
The strategies of therapy selection must consider the alternatives and weigh them in
selecting the course of action that corresponds best to the particular circumstances
of an individual patient. Among the major factors affecting therapy selection, we consider
current diagnostic status, past history, and the desired expected outcome for the patient.
Once the patient is undergoing treatment, the effectiveness of the current medication
must be assessed and new factors considered, such as side effects, complications of
the disease for which the current therapy is not effective, conditions not detected at
the initial visit, etc. An important aspect of our system is that it does not “freeze”
past diagnoses as a permanent interpretation, though past diagnoses can be recalled.
At each follow-up visit, the entire past history and set of updated findings are re-
evaluated. and the possibility of a modified diagnosis considered. In addition to therapy
28 SHOLOM WEISS.CASIMIRA. KULIKOWSKIand ARAN SAFIR

recommendations, the program may suggest other possible actions, such as the perform-
ance of special diagnostic tests or advice to the patient to note certain key manifestations
as well as possible side effects of therapy.

CAUSAL ASSOCIATIONAL NETWORK

(CASNET) MODELS
A causal-associational network is a particular type of semantic network [16] designed
to:
(a) describe disease in causal terms;
(b) relate this description to an associational structure of external observations; and
(c) describe various classifications imposed on the model.
CASNET models can be used to describe many different complex processes, although
we have developed them to describe pathophysiological medical events [lo]. Knowledge,
in our scheme, is represented by three types of data elements, corresponding to the
three kinds of description outlined above. These are: observations of the patient; stat&
of the causal net; and the diagnostic, prognostic, and therapeutic categories. Observa-
tions are the direct evidence obtained about a patient. Pathophysiological states are
intermediate constructs that describe conditions or mechanisms that summarize results
from many different observations. Categories of disease are conceptually at the highest
level of abstraction, summarizing both states and observations.

SURE GLAUCOMA

FbthophysioloipM
States

fmpkco tioffo/
Links --j

Fig. 1.

A small illustrative example of some of the elements and interrelationships described

above is shown in Fig. 1. A simplified causal network for glaucoma is shown in Fig.
2. It contains cu. 40 pathophysiological states and their causal relationships, super-
imposed on a schematic diagram of the eye. Many of the states included in this figure
are summarizations of more detailed states that are defined in the current implemen-
tation of the glaucoma model, which includes approx. 120 such states.
Observation-state mappings
Many interdependent observations or tests, ti, are associated with the description
of a disease. Redundancies in the interpretation of these observations may be eliminated
by mapping Boolean combinations of tests, B(t,), into a single causal state. A mapping
is stated as an associational link that carries a confidence factor, Q:
II Ai+ sj. (1)
Fig. 2. A simplified causal network for glaucoma.
30 SHOLOM WEISS, CASIMIR A. KULIKOWSKI and ARAN SAFIR

Observations are first quantized so that tests in the form ti may be only: true, false
or undetermined. When the result of the test, ti,‘is obtained, the measures, Qij, determine
the confidence with which we can assert the presence, (Qij > 0), or absence, (Qij < O),
of state Sj, regardless of other factors.
Because tests are directly observable evidence, and states are hypothesized, we can
most readily interpret the confidence measures, Q, as weights of inverse inference. We
express these measures as positive integers to represent confidence in the presence of
Sj, and negative integers to represent confidence in its absence. The MYCIN system
has also used confidence factors in its reasoning procedures [17].

State network rules

States are linked to each other by causal rules of the form:
S iz!_)
Sj
(2)

In a causal rule, aij is interpreted as the likelihood with which we expect the effect
Sj to follow from the occurrence of the cause Si. It is assigned as a number from
0 to 1, but in clinical descriptions of disease we have found it useful to use only a
few levels of quantization. These correspond to simple frequency interpretations of causal
strength: sometimes, often, usually, almost always and always.
In a CASNET model we do not require that a set of mutually exclusive and exhaustive
effects be defined for every cause, nor a similar set of causes be defined for every effect.
The assigned transition weights are independent of the manner in which a cause is
itself produced. There are no constraints on the entire set of weights emanating from
a single cause. No state can be self causing; therefore, the causal net contains no loops.
Inverse causal weights (from effects backwards to their causes) depend strongly on
the entire context leading to the effects. They cannot be assigned statically when the
model is being designed unless we wish to exhaust all likely antecedent pathways for
every state in the network. Thus, the inverse causal weights are best calculated by
an algorithm that constructs contexts dynamically at the time a case is being considered
[lOI.
Some states have no antecedent causes described in the model. These “starting states”
represent basic causes of the disease. Other states in the network are terminal. No
effects are described for them. They limit the scope ‘bf the network by specifying the
furthest patterns of progression of the disease that are to be included as part of the
model.
The designer of a CASNET model has no fixed limit on the level of detail that
must be incorporated into the model. However, in an effective interpretational system,
the states must serve as summaries for the observations. The specific properties of states,
such as measures of magnitude or descriptive qualifiers, are left to the modeler to assign
in each particular medical domain.

Classijication rules
The “classification tables” are logically ordered rules for combining confirmed and
denied states in order to produce diagnostic, prognostic and explanatory conclusions.
They can be used to select different pathways through the causal net or to compose
any desired configuration of states or pathways of states.
Classification table rules are of the form:
4Si) --* D.x, (3)
where B(Si) is a Boolean combination of (confirmed or denied) states and D, is a diagnos-
tic statement.
The construction of a complete hypothesis for the patient is carried out dynamically
by various strategies [lo], and the classification tables serve only to elicit the appropriate
constituent parts, some of which may disappear in the final conclusion, having been
subsumed or excluded by others.
 Glaucoma consultation by computer 3I

GLAUCOMA CONSULTATION PROGRAM

Medical consultation programs can be addressed to various distinct classes of users.
As a research tool for the specialist, the consultation program serves as an interpretative
adjunct to a data base of specialty cases. It permits the expert to match his latest
information from research studies against an existing model. The computer scientist
with whom he works will be interested in studying the matching process to learn how
to update and improve the model. As a teaching device for the resident in training
or the physician in continuing education, a consultation program can be particularly
useful if it is able to provide alternative explanations from different experts about difficult
clinical cases. These can be supported by appropriate quotations and by facilities to
explore the logic of the program that led to the different conclusions. Without such
detailed output, the program could serve as a guide for the paramedic in acquiring
and storing patient data and as a summarizing system for producing patient records.
Several of the above applications require different types of input format, controlling
strategies, and vocabulary. We have several alternative options within the glaucoma
consultation program to make it useful to a variety of users. Because of speed and
efficiency considerations, our program is written in FORTRAN. It is an interactive
program running in 35 K words of memory, on the PDP-10 computer under either
the TOPS-10 or TENEX operating systems. The glaucoma program is an adaptation
of a general system for consultation based on CASNET-type models [I 81. Various
modifications of the glaucoma model are carried out by interaction with a separate
editing program [12], written in SNOBOL. This program checks the model for consist-
ency and compiles it so that it will run efficiently under the consultation program.
Figure 3 is a flow chart of the major modules of the consultation program. The
glaucoma system will be illustrated by examples that show results from each of these
sections.

Patierzt identijcation and case type selection

The program begins by requesting that the user select the type of case (real or hypothe-
tical) and the mode of data entry. The ability to distinguish hypothetical from real
cases is important. Hypothetical cases are often used for testing the system with delibera-
tely complex problems and for teaching with prototypical cases for the different diagnos-
tic categories. The coded mode of entry can accelerate input by a technician, and is
particularly useful for entering data from follow-up visits, when only a few observations
may have been taken.
There are three distinct types of cases: a new entry to the system; a case already
stored in the data base, to be retrieved for review; or an old case that is to be updated
by new information.

Questioning for disease-specific information

As stated above, the program will accept input information in two different modes:
as codes, and as responses to English language questions. The codes are numbers
assigned to each question taken from a printed questionnaire and entered by a user
when appropriate. Only positive responses and available measurements need be entered,
and they are preceded by the prefixes R, L, B, to indicate applicability to right, left,
or both eyes respectively.
A sophisticated and thorough consultation by computer requires a sequence of ques-
tioning that takes into account at every stage the pattern of previous responses. This
not only produces an intelligent dialogue, but can also significantly reduce the number
of questions asked during a session. For glaucoma, as for any ocular disease, the strategy
is somewhat complicated by the need to inquire about both eyes for many of the
measurements.
For CASNET models, several general strategies have been designed to select sequences
of questions on the basis of diagnostic and cost criteria [9]. Nevertheless, for the glau-
 SHOLOMWEISS, CASIMIRA. KULIKOWSKI and ARAN SAFW

INTERPRETATION
OF
FINDINGS

Patent Status

I Pattern of
I
SELECTION Evaluation of 1
OF I Dlagnostlc _ Observations
Associated wth __c Current
TREATMEFuT 1 State Treatment /
Treotments
I I

Recommendations

User Comments

Fig. 3. Flow chart of the glaucoma consultation program

coma model, a relatively well-structured domain-dependent sequencing was preferred

for routine use by the clinicians.
In this approach, the questions for the consultation have been organized to take
advantage of the inherent structure of many of the topics investigated during a glaucoma
examination. The program proceeds along the traditional line of history, physical find-
ings, and laboratory tests. Within each area, questions are grouped so that, for example,
all iris abnormalities, or all visual field results, can be listed together. A hierarchical
ordering can then be applied locally to various question topics. For example, one may
inquire as to whether there is any structural abnormality observed during a slit-lamp
examination. If this is the case, iris abnormalities may be considered, and if one is
found, the various specific abnormalities can be covered. This is not a pattern of ques-
tioning based on a global decision tree, but rather, a sequential questioning structure
that uses local tree orderings within small, related topics of inquiry. A second element
in the strategy of questioning is a list of conditions that must be met before a certain
question may be asked. If these conditions are not met, the question is eliminated
from consideration. These two elements of inquiry, i.e. local question structure and
conditions for question selection, form an effective basis for questioning by the glaucoma
program, or any diagnostic program that must guide the user within a well-circum-
scribed area of knowledge.
 Glaucoma consultation by computer 33

In all examples, responses are expressed as Y, N, or U for yes, no, and unavailable
respectively. Numerical responses are entered in the order of right eye. left eye. In
the examples that follow, the conventional notations of OD, OS, OU are used for right
eye, left eye, and both eyes respectively. An example of the data input is given in
Fig. 4.

Summurizing the findings

The program must be able to summarize, in an organized manner, the various com-
binations of results that have been entered by the ophthalmologist. This is essential
if the ophthalmologist is to confirm that the information has been correctly entered.
It is also needed for examining cases after retrieval from the data base. At the present
time, the glaucoma model has approx. 300 items of information that can be listed.
Because most of these may apply to either one or both eyes, the number of combinations
is much larger. However, only a subset of the items will be applicable to any given
patient. Since almost any kind of pattern of results may arise within the subset, a
general procedure is needed to produce an acceptable summary. Summarization is keyed
to the type of question for which a result has been received. Examples of question-types

1 . WHICH HISTORY TOPICS DO YOU WISH TO ENTER:

(1) CURRENT GLAUCOMA MEDICATIONS (OR STEROIDS)
(2) SYMPTOMS
(3) DRUG INTOLERANCES
(4) PAST HISTORY OF OCULAR SURGERY
(5) FAMILY HISTORY OF GLAUCOMA
(6) PATIENT HISTORY OF SYSTEMIC DISEASES
(7) PATIENT HISTORY OF OCULAR DISEASE OR TRAUMA
*VALUES: N
2. WHAT Is THE BEST CORRECTED VISUAL ACUITY-- 201 ? (ENTER l,2,3,4 FOR
Fc, HM, LP, NLP RESPECTIVELY)
'30,110
3. ENTER THE INTRAOCULAR PRESSURE (APPLANATION TENSION):
*34,3a
4. DO YOU HAVE ANY RESULTS FROM A SLIT LAMP/GONIOSCOPY EXAMINATION?
'Y
5. ANTERIOR CHAMBER DEPTH:
(1) FLAT (COLLAPSED)
(2) SHALLOW
1;; ;;;;"A'"

*2,2
6. ANGLES:
(0) GRADE 0 - COMPLETE OR PARTIAL CLOSURE
(1) GRADE 1 - NARROW EXTREME
(2) GRADE 2 - NARROW MODERATE
(3) GRADE 3 - OPEN
(4) GRADE 4 - OPEN
(5) SLIT
lo,o

18. DO YOU HAVE VISUAL FIELD DATA?

'Y
19. VISUAL FIELDS:
(1) NORMAL

I:;
(4)
NASAL STEP
TEMPORAL
ISOLATED
STEP
PARACENTRAL SCOTOMA
(5) INCOMPLETE ARCUATE SCOTOMA
(6) COMPLETE ARCUATE SCOTOMA
(7) ARCUATE SCOTOMA WITH BREAKTHROUGH T3 PERIPHERY
(8) DOUBLE ARCUATE SCOTOMA
(9) DOUBLE ARCUATE SCOTOMA WITH BREAKTHROUGH TO PERIPHERY
(10) CENTRAL ISLAND
(11) TEMPORAL ISLAND
(12) SPLITTING OF FIXATION
*VALUES:
*OD: 4
'OS: 6
Fig. 4. Example of data entry.
34 SHOLOM WEISS. CASIMIR A. KULIKOWSKI and ARAN SAFIR

are: those which accept numerical responses, binary questions, multiple choice, alterna-
tive choice, etc. Results are listed with binocular and monocular identification where
appropriate. Results that the program infers from changes in observations made during
follow-up visits are also included in the updated summary. An example of a summary
produced by the program is given in Fig. 5.

Computer-based diagnostic interpretation

The initial test results undergo several interpretational transformations before diagnos-
tic conclusions and therapeutic recommendations are formulated.
Test-to-state mappings. Each test has associated with it measures Q indicating the
level of confidence for the denial or confirmation of those states affected by the test.
Most rules will be found not applicable for a specific patient because the logical precon-
ditions have not been satisfied, or the tests not performed. When several consistent
rules apply to a single state, the rule with the greatest magnitude of confidence (for
confirmation or denial) is employed.
The Boolean logic of the test rules is examined to see which rules are satisfied, and
all applicable test-state mappings are performed. At the completion of this phase, each
state has been assigned a derived confidence measure, STATUS, which reflects whether
it is to fall in the confirmed, denied, or undetermined ranges. Because differences and
similarities between the subject’s eyes are often significant, comparisons are made and
the STATUS measures for both eyes compared and modified if necessary.
Diagnostic hypothesis generation. For any given instance of test results for a specific
patient, a configuration of confirmed, denied, and undetermined states is formed. The
starting states indicate the underlying basic mechanisms of disease. The most likely
mechanisms for a given patient are identified by finding those starting states for which
(causal) pathways may be generated which reach the largest number of states that are
confirmed (without traversing a denied state). This procedure is repeated until all con-

********************
* GLAUCOMA SUMMARY *
l *****+*****o*******

PERSONAL DATA:

NAME: H.E.
AGE: 40 RACE: W SEX: M CASE NO: 28

CLINICAL DATA SUMMARY FOR VISIT OF 515171

BEST CORRECTED VISUAL ACUITY:

OD: 20/ 30 OS: 20/ 40
IOP:
OD: 34 OS: 38
ANTERIOR CHAMBER DEPTH:
SHALLOW (OU)
ANGLES:
GRADE 0 (OU)
ANGLE CLOSURE:
OD:150 DEGREES OS:340 DEGREES
PERIPHERAL ANTERIOR SYNECHIAS:
0D:lOO DEGREES OS:270 DEGREES
HORIZONTAL CUP/DISC RATIO: 0.60 (OU)
OPTIC NERVE RIM WIDTH:
VARIABLY NARROW (OU)
VISUAL FIELDS:
ISOLATED PARACENTRAL SCOTOMA COD)
COMPLETE ARCUATE SCOTOMA (OS)
TONOGRAPHY C:
OD: 0.06 OS: 0.05

*******************
ANY REVISIONS (EXCLUDING THE COMMENTS) 7 N

Fig. 5. Example of data summary.

 Glaucoma consultation by computer 15

firmed states are covered. Each starting state has a pointer to those classification tables
of rules that are potentially applicable to the disease mechanisms related to that starting
state. Those rules that are logically satisfied will result in diagnostic conclusions for
the patient. The indicated classification tables are searched to see which rules are satis-
fied. Because the rules are ordered by seriousness of disease. if more than one rule
in a table is logically satisfied, the rule indicating the most serious degree of disease
is selected. The selected rule will point directly to a diagnostic statement. Once again
the conclusions for each eye are compared and the tentative conclusions may be modified
by binocular considerations.

Therapy selection is highly dependent on the diagnostic conclusions. A diagnostic

conclusion points to an ordered preference list of treatment states. These states, and
their associated STATUS measures, are interpreted somewhat differently from disease
states. The STATUS measure (derived, as usual, from test-state mappings) is here used
to indicate the measure of confidence in the success of a treatment. The treatment
with the largest STATUS value is selected from the list of treatments associated with
the diagnostic conclusion. If several different treatment recommendations are derived,
a master list is consulted to see if any tentatively recommended treatments are covered
by other treatments and are therefore unnecessary. Before the diagnostic and therapeutic
conclusions are displayed, the tentative conclusions for both eyes are again compared
and modifications made, if they are needed. An example of the conclusions is illustrated
in Fig. 6.

Fwihc~ ir~ftirmtion and research studies

There are currently over 200 possible diagnosis and therapy statements, many of
which are not mutually exclusive. For most cases, more than a single statement will
be appropriate. These conclusions and recommendations are often explanatory in nature,
and summarize particularly significant features of a case. These statements reflect the
judgement of our panel of experts and may contain alternative recommendations for
diagnoses and therapies. Even so, they must be concise and brief. Another section of
the program has been designed to amplify the conclusions in some cases. It emphasizes
citations from research studies, and quotations from experts to support the program’s
conclusions. Partial or alternative results from studies that may not yet be ready for
incorporation into the logic of the model are also included here. Figure 7 illustrates
the additional commentary for the case shown in Fig. 6.

A patient may return for frequent visits and may be followed for many years. The
program saves the test results from each visit. Computer recommendations for each
visit are not stored. because they can be regenerated, reflecting the latest stage of devel-
opment of the model and program logic. The chronological review feature allows the
ophthalmologist to examine the observations and computer conclusions for any specific
visit or sequence of visits of a patient. For this patient, one may also trace through
time the evolving patterns of disease, the computer’s analyses of the status of disease,
and the recommendations for treatment. In addition to saving the findings from each
visit (indicated by a “*” in the summary), the program generates and saves a master
summary which synthesizes the current findings with those found for this patient during
previous visits. The storage requirements for saving a case are not large because an
efficient bit encoding scheme is used for storing data [13]. An abstraction of visit sum-
maries is given in Fig. 8.

RESULTS AND DISCUSSION

The consultation program was developed in the course of an investigation into the
.use of models of disease processes for computer-based consultation. From this evolved
 36 SHOLOM WEISS, CASWIR A. KULIKOWSKI and ARAN SAFIR

DO YOU WANT TO SEE THE COMPUTER DX? Y

***********************
*DIAGNOSIS AND THERAPY*
*ii********************

l RIGHT EYE: *

[ll PRESENT DIAGNOSTIC STATUS:

CHRONIC (PRIMARY) ANGLE CLOSURE GLAUCOMA.

[21 TREATMENT RECOMMENDATIONS:

PERIPHERAL IRIDECTOMY IS INDICATED. INSTITUTE VEAK MIOTIC THERAPY

AS NEEDED TO NORMALIZE TENSIONS. (ADD DIAMOX IF NECESSARY)

l LEFT EYE: l

[ll PRESENT DIAGNOSTIC STATUS:

CHRONIC (PRIMARY) ANGLE CLOSURE GLAUCOMA.

[21 TREATMENT RECOMMENDATIONS:

PERIPHERAL IRIDECTOMY IS INDICATED. INSTITUTE WEAK MIOTIC 'THERAPY

AS NEEDED TO NORMALIZE TENSIONS. (ADD DIAMOX IF NECESSARY)
SECTOR IRIDECTOMY SHOULD BE CONSIDERED.
BECAUSE OF A SIGNIFICANT DEGREE OF PAS, FILTERING SURGERY
MAY BE CONSIDERED. HOWEVER, THE INCREASED RISK OF FILTERING SURGERY
MAKES IRIDECTOMY PREFERABLE. STUDIES SHOW AN EQUAL CHANCE OF SUCCESS
OF FILTERING SURGERY FOLLOWING FAILURE OF IRIDECTOMY.

Fig. 6. Example of conclusions from the consultation program

Do you WANT MORE INFORMATION (RESEARCH STUDIES OR OPINIONS)? Y

* RIGHT EYE: l

"THE MEDICAL THERAPY FOR ANGLE-CLOSURE GLAUCOMA IS USEFUL

ONLY AS A PRELUDE TO SURGERY... MIOTICS AND DIAMOX
ARE STILL MOST VALUABLE IN LOWERING TENSION.
THE ADDITION OF HYPEROSMOTIC AGENTS PREOPERATIVELY SHOULD
BE EMPLOYED IF NORMAL LEVELS OF PRESSURE HAVE NOT BEEN ATTAINED."--
(BECKER-SHAFFER, 1970)
"THE RESULTS OF 139 IRIDECTOMIES PERFORMED FOR NARROWNESS OF
THE ANGLE OF THE ANTERIOR CHAMBER INDICATE THAT THE OPERATION
CAN BE HELPFUL IN PATIENTS WITH ACUTE ANGLE CLOSURE OF RECENT ONSET,
WITH PROLONGED ATTACKS OF ANGLE CLOSURE, OR EVEN WITH UNDERLYING
CHRONIC OPEN-ANGLE GLAUCOMA. INTRAOCULAR PRESSURE BELOW 25 MM HG
WITH OR WITHOUT MEDICAL THERAPY FOLLOWING IRIDECTOMY WAS ACHIEVED IN
87% OF ALL CASES WITH ACUTE ANGLE-CLOSURE, INCLUDING 772 OF CASES
IN WHICH THE ATTACK HAD BEEN PRESENT FOR LONGER THAN THREE DAYS,
IN 80% OF CASES WITH CHRONIC ANGLE CLOSURE GLAUCOMA, AND IN
80% OF CASES WITH CHRONIC OPEN ANGLE GLAUCOMA THAT HAD A SUPERIMPOSED
ANGLE CLOSURE ATTACK. SUBSEQUENT SURGERY WAS NEEDED IN 8% OF THE
CASES... THE EXPERIENCE GAINED IN THE PRESENT SERIES OF CASES INDICATES
THAT EVEN AFTER SYNECHIAL CLOSURE HAS DEVELOPED, IRIDECTOMY MAY
STILL BE THE PROCEDURE OF CHOICE IN MOST INDIVIDUALS."
(MURPHY, SPAETH - ARCH. OPHTH./VOL 91, FEB. 1974)

l LEFT EYE: l

"IN A RECENT SERIES OF 35 IRIDENCLEISIS OPERATIONS ON EYES WITH

PREVIOUS IRIDECTOMIES THERE WERE 9 EYES (26%) THAT FAILED To BE
CONTROLLED. THIS IS QUITE COMPARABLE TO THE AVERAGE FAILURE RATE(26fb)
OF IRIDENCLEISIS AS A PRIMARY PROCEDURE." (BECKER-SHAFFER, 1970)
"IF LENS CHANGES ARE PRESENT, RETINAL PATHOLOGY, OR A PROBABLE
NEED FOR MIOTICS POST-OP, THEN SECTOR IRIDECTOMY SHOULD BE DONE.
IF LENS CHANGES ARE SEVERE ENOUGH LENS EXTRACTION." (M. KASS-1976)

Fig. 7. Supplementary information for a consultation.

 Glaucoma consultation by computer 37

the causal associational network (CASNET) representation, with glaucoma chosen as

the disease which would serve for the development of a prototype program. As work
progressed, the goals became more clearly those of developing an expert consultation
system, in collaboration with a network of glaucoma researchers.
Because the program’s logic is contained in general strategies that analyze the CAS-
NET model, it is relatively easy to incorporate new medical knowledge or improve
THERE ARE 3 VISITS. DO YOU WANT TO SEE THE CHRONOLOGY? Y

DO YOU WANT (1)ALL VISITS (2)A SPECIFIC VISIT: 1

DO YOU WANT (1)S'JMMARY ONLY (2)SUMMARY WITH DIAGNOSIS

(3)SUMMARY, DX & STUDIES: 2

(VISIT: 1) AGE: 35

CLINICAL DATA SUMMARY FOR VISIT OF 51 5171

BEST CORRECTED VISUAL ACUITY:
OD: 20/ 30 OS: 20/ 40
IOP:
OD: 34 OS: 38

(VISIT: 3) AGE: 39
CLINICAL DATA SUMMARY FOR VISIT OF 71 5175
CURRENT MEDICATIONS:
'EPINEPHRINE 2% BID (OS)
'HUMORSOL 0.12% (OS)

***1**it**ff*************
*DIAGNOSIS AND THERAPY+
l ****************,*****

l RIGHT EYE: *

El] PRESENT DIAGNOSTIC STATUS:

PREVIOUS SURGERY FOR ANGLE CLOSURE GLAUCOMA.

RESIDUAL GLAUCOMATOUS DAMAGE FROM ANGLE CLOSURE. THE GLAUCOMA
MECHANISM MAY HAVE BEEN ELIMINATED BY SURGERY.
OPEN ANGLE WITH PAS (SECONDARY MECHANISM).
CHARACTERISTIC VISUAL FIELD LOSS WITH CORRESPONDING DISC CHANGES.
EARLY FIELD LOSS.

[21 TREATMENT RECOMMENDATIONS:

Fig. 8. Example of a chronological summary of a case.

existing parts of the model by increasing its depth and breadth of description. The
prototype model, which covered the description of the disease in a single eye for a
single visit of the patient, was succeeded by a binocular model, which in turn’ evolved
into a model that incorporated follow-up visit information.
One measure of the performance of a decision-making system is the level of accuracy
of classification for the original sample from which the classifier has been derived. It
is not unusual to find classifiers for which no adjustment can produce a 100’~ accurate
zx SHOLOM WEISS. CASIMIR A. KULIKOWSKI and ARAN SACIR

classification for the sample. Early in our work, we collected a sample of 40 difficult
cases. Initially, the program did not classify (diagnose) all cases correctly. However,
as our model improved, it was soon able to diagnose the 40 cases correctly. This result
demonstrated, at a relatively early stage, that our approach did provide an incremental
means of improving the program’s performance. We became confident that poor or
inaccurate conclusions could be corrected, that cases diagnosed correctly would remain
correct, and that diagnostic and therapeutic recommendations could be improved.
This first cycle of development was followed by a second stage, which began when
an improved prototype of the glaucoma consultation program was presented at the
annual meeting of the Association for Research in Vision and Ophthalmology in 1973.
Interest in the program, arising from this demonstration of its capabilities, led to the
collaboration with several expert glaucoma researchers. It was hoped that their advice
would result in a program that could provide consultation at a sophisticated level.
Another motivation for this approach was the authors’ desire to test the applicability
of the consultation system as a tool for clinical research. The consultation system would
serve as a point of entry for difficult clinical cases. A data base of such cases would
then be accumulated to test the consultation program. Statistical analysis of the data
might provide new insights into the disease process itself.
Researchers from five different medical centers have, since 1975, been accessing and
testing the glaucoma consultation program. Each has chosen a sub-specialty of glaucoma
for testing and developing the program in depth. Many subtleties of description and
reasoning in several types of glaucoma (primary open angle, angle closure, and the
secondary glaucomas) have been added as the result of this form of collaborative clinical
testing.
Although we currently have several hundred cases on file, it is difficult to evaluate
program performance in a simple manner. Classifying conclusions as being merely
correct or incorrect is an oversimplification. The program’s conclusions are presented
not as single unique diagnoses but rather as combinations of judgements about a patient’s
status. These may include such factors as: the type and severity of disease, evaluation
of current therapy, and recommendations for future testing or therapy. In an objective
evaluation of program performance, each of these elements must be considered. Most
of the cases selected by our clinical collaborators are complex and difficult ones requir-
ing expert judgement. Our sample of cases has thus been deliberately biased to enable
us to develop an expert consultant program rather than one that merely does well
in a large percentage of typical glaucoma cases. The network of investigators estimates
that the program arrives at reasonable and often sophisticated judgements in CLI.75”:;
of the difficult cases of glaucoma [ 191.
The glaucoma consultation system was subject to an intensive evaluation by a large
and varied group of ophthalmologists during the 1976 meeting of the American Academy
of Ophthalmology and Otolaryngology. The consultation program was used to summar-
ize results of cases and present its recommendations, contrasting them to the opinions
of a panel of experts, at the glaucoma symposium just preceding the formal opening
of the Academy convention. The cases had been entered into the computer in advance
of the symposium, but the program’s conclusions were left unaltered. The panel gave
a variety of opinions about the cases, and in almost all of them the program included
in its alternatives the main interpretation given by the panel.
We also tested the program in a more detailed manner. It was one of the scientific
exhibits at the Academy meeting. It was displayed and made available for testing by
all conference attendees. Evaluation questionnaires were filled out by those ophthalmolo-
gists who tested the program. Forty nine responses were obtained. The results are sum-
marized in Table 1.
A 95’2, acceptance rate for clinical proficiency in the sample questioned is high given
the amount of unknown material presented to the consultation system by the ophthal-
mologists, who were encouraged to test it with their difficult cases. The two cases (5”;,)
in which the program was judged to perform inadequately corresponded to situations
 Glaucoma consultation by computer 39

Table 1. Summary of program evaluation responses

QUt?StlOIlS
(n = Number 01 I. Level of clinical proficiency 2 Applicablbty to glaucoma research
responses, (.v = 44) (N = JII

,a) Expert (a) Very applicable 71”,, (ii) Very important w,,
(b) Very compstent (bl Moderately applicable w,, ib) Moderately important 42” ,,
,c, Acceptably competent (c) Somewhat applicable 5” CC) Somewhat important I I‘I,,
IdI or illtIc mlporrnncc 70
Id) Inadequate (d) 01 httle value o’.:: - I,

in which diseases other than glaucoma formed a significant part of the diagnosis. and
the appropriate information had not yet been included in the model. The 77pd rate
of high competence (the “expert” and “very competent” responses) ascribed to the system
by this independent sample of ophthalmologists accords well with the previously cited
judgement of our glaucoma collaborators. Efforts currently being devoted to including
alternative expert opinions in complex cases are expected to improve this performance
index in the coming year. The answers to the second question listed in Table 1 indicate
the strong potential that the ophthalmologists saw in using the consultation program
as a support tool for organizing clinical trials, and for summarizing and analyzing
their results.
It is interesting to observe the differences between the responses to the second and
third questions. Clearly, the ophthalmologists see an important ultimate contribution
to health care (87qb for very or moderately important), but this is secondary to the
applicability to glaucoma research (95’4 for the two top responses).

SUMMARY
This paper describes a computer-based system for consultation in the diagnosis and
therapy of glaucoma. The reasoning procedures interpret the findings of a particular
patient in terms of a causal-associational network (CASNET) model that characterizes
the pathophysiological mechanisms and clinical course of treated and untreated diseases.
Glaucoma investigators, linked by computer networking, collaborate in the development
of the program as a clinically useful tool.
After reviewing the CASNET model representation, we describe the various com-
ponents of the consultation program and the types of recommendations it produces.
Examples from typical cases are used to illustrate the capabilities of the program
for: (a) interactive questioning; (b) summarization of acquired data; (c) diagnostic and
prognostic interpretation; (d) therapeutic recommendations: (e) citation of research
studies; and (f) chronological review of follow-up visits.
The major new features of this program are: (a) generation of complex interpretations
from a qualitative model of a disease process; (b) reasoning about detailed follow-up
management of a patient; (c) incorporation of alternative expert opinions about subjects
under debate; and (d) its testing and updating by a collaborative computer-based
network of glaucoma researchers.

REFERENCES
I. R. S. Ledley and L. B. Lusted, Reasoning foundation of medical diagnosis. symbolic logic. probabilities
and value theory aid our understanding of how physicians reason, Science 130, 9 (1959).
2. H. R. Warner, A. F. Toronto and L. G. Veasy. Experience with Bayes theorem for computer diagnosis
of congenital heart disease, Ann. N.Y Acad. Sci. 115. 558 (1964).
3. G. A. Gorry and G. 0. Barnett, Experience with a model of sequential diagnosis. Comput. Biomrd. Re.
1, 490 (1968).
4. H. L. Bleich, Computer evaluation of acid-base disorders, .I. Clin. Irmesr. 48. 1689-1696 (1969).
5. C. Kulikowski and S. Weiss, Computer-based models of glaucoma. Comput. Biomrd. Tech. Rpt No. 3.
Department of Computer Science, Rutgers University (1971).
6. E. H. Shortliffe, S. G. Axline, B. G. Buchanan, T. C. Herigan and S. N. Cohen, An artificial intelligence
program to advise physicians regarding antimicrobial therapy, Compuf. Eiomrd. Res. 6, 544 (1973).
7. H. E. Pople. Jr., J. D. Myers and R. A. Miller, DIALOG: A model of diagnostic logic for internal
medicine. Proc. 4th IJCAI 2, 8499855 (1975).
8. A. Rubin. The role of hypotheses in medical diagnosis. Proc. -If/l IJC.41. 2. 856--X62 (1975).
40 SHOLOM WEISS.CASIMIRA. KULIKOWSKI and ARAN SAFIR

9. C. Kulikowski and S. Weiss, Strategies of data base utilization in sequential pattern recognition. In
Proc. IY7_’ IEEE Con/I itI Decision and Control and It th Symp. on Aduptioe Processes. pp. 1033105.
(1972).
10. S. Weiss, A system for model-based computer-aided diagnosis and therapy. Ph.D Dissertation, Rutgers
University (1974).
11. C. Kulikowski, S. Weiss, M. Trigoboff and A. Safir. Clinical consultation and the representation of disease
processes: some Al approaches, Proc. Conf ON Artificial Intelligerlce arid the Simulatiorl of Behauiour,
166174 (1976).
12. C. Kulikowski and S. Weiss, An interactive facility for the inferential modeling of disease. In Proc. IY73
Princeton Conf: on Information Sciences and Systems, p. 524 (1973).
13. S. Weiss, K. Kern, C. Kulikowski and A. Safir, A system for interactive analysis of a time-sequenced
ophthalmological data base, Proc. 3rd Illinois Conf on Medical Information Systems (1976).
14. A. E. Kolker and J. Hetherington, Becker-Shqffer’s Diagnosis arld Therapy of the Glaucomas, Mosby,
St. Louis. 3rd Edn. (1970).
15. S. Duke-Elder, Systems of Ophthalmology. Vol. 11. Mosby. St. Louis (1971).
16. W. Woods, What’s in a Link. Foundations for Semantic Networks, in Representation und Understanding,
(Eds. Bobrow and Collins). Academic Press, New York (1975).
17. E. H. Shortliffe and B. Buchanan, A model of inexact reasoning in medicine. Math. Biosci. 23, 351-79
(1975).
18. S. Weiss, C. Kulikowski and A. Safir, A model based method for computer-aided medical decision-making,
Computers in Biomedicine Technical Report No. 63, Dept. of Computer Science, Rutgers University,
January (1976).
19. M. Kass and J. Wilensky. personal communications.

Ahout the Author-SHoLoM M. WEISS received the B.S. degree in mathematics in 1968 from
the City College of New York, and the M.S. and Ph.D. degrees in Computer Science from
Rutgers University in 1970 and 1974 respectively.
In 1974. he joined the staff of the Institute of Computer Science at the Mt. Sinai School
of Medicine of New York. There, he directs the research activities and manages software develop-
ment for the Rutgers-Mt. Sinai Computer Health Care Laboratory. Current activities include
the design and development of advanced medical decision-making and data base systems. Since
1974, he has been a visiting assistant professor of Computer Science at Rutgers University.

About the Author-CASIMIR A. KULIKOWSKI received the B.E. and M.S. degrees from Yale Univer-
sity and the Ph.D. degree in Electrical Engineering from the University of Hawaii. His research
has been in the field of pattern recognition, medical modeling and decision-making. Since 1970
he has been with the Department of Computer Science at Rutgers University, where he is
currently an associate professor. He is also Associate Director of the Rutgers Resource on
Computers in Biomedicine, within which he heads the group on medical systems, which is
developing various artificial intelligence approaches to problems of clinical reasoning and
research inquiry in biomedicine. In collaboration with the Mt. Sinai School of Medicine he
is also engaged in a number of computer modeling projects in vision and ophthalmology.

About the Author-ARm SAFERreceived his B.A. degree from New York University in 1950,
and his M.D. degree from New York University-Bellevue in 1954. He completed his ophthalmo-
logy residency at the New York Eye and Ear Infirmary in 19561959. He studied physiological
optics at Cambridge University, England in 1961-1962 as a Special USPHS Fellow. He has
been with the Department of Ophthalmology at the Mt. Sinai School of Medicine since 1964,
and is currently Professor of Ophthalmology. Since 1975 he has also been Director of the
Mt. Sinai Institute of Computer Science. His research interests include physiological optics,
clinical refraction, and computer modeling and medical decision making. He is the author of
the text Refraction-International Ophthalmology Clinics, Little, Brown, 1971 and of the text
Refraction und Clinicul Optics. Harper & Row, 1976.