Acquisition of Cognitive Procedures (ACT) P41 RROO785-09

Another class of learning mechanisms are concerned with tuning
existing procedures so that they apply more appropriately. There are
various mechanisms concerned with extending or generalizing the range of
application of a procedure. In the past year we have been working at
reducing these differant generalization processes to a common partial
matching process. In addition to generalization, tuning occurs in the ACTF
system by means of discrimination and composition. Discrimination is a
process for restricting the range of applicability of a production.
Composition attempts to build macro-operators out of a series of
productions.

The third direction of our learning work has been concerned with
developing a flexible strength-based set of conflict resolution rules.
Here we are concerned with modelling the gradual improvement seen in human
cognitive skills and also providing the system with the resilience so that
it can recover from noise and changes in environmental contingencies.

We have been applying this theory in detail to a simulation of how
students acquire proof skills in geometry. We have a more or less thorough
analysis of how students learn new postulates of geometry; we initially use
these postulates in an interpretative fashion, integrating them with prior
knowledge; how they compile special purpose procedures that directly apply
this knowledge to proof generation; and how these procedures become tuned
with practice. This application has provided strong evidence for most of
the learning developments in the ACT system. It has also forced us to
develop formalisms for how planning and problem-solving should be
structured within a production-system framework.

D. List of Project Publications

{1] Anderson, J.R. Language, Memory, and Thought. Hillsdale, N.J.: L.
Erlbaum, Assoc., 1976.

(2] Kline, P.J. & Anderson, J.R. The ACTE User’s Manual, 1976.

[3] Anderson, J.R., Kline, P. & Lewis, C. Language processing by
production systems. In P. Carpenter and M. Just (Eds.). Cognitive
Processes in Comprehension. L. Erlbaum Assoc., 1977.

{4] Anderson, J.R. Induction of augmented transition networks. Cognitive
Science, 1977, 125-157.

(S] Anderson, J.R. & Kline, P. Design of a production system. Paper
presented at the Workshop on Pattern-Directed Inference Systems,
Hawaii, May 23-27, 1977.

{6] Anderson, J.R. Computer simulation of a language acquisition system:
A second report. In D. LaBerge and S.J. Samuels (Eds.). Perception
and Comprehension. Hillsdale, N.J.: L. Erlbaum Assoc., 1978.

{7] Anderson, J.R., Kline, P.J., & Beasley, C.M. A theory of the

acquisition of cognitive skills. In G.H. Bower (Ed.). Learning and
Motivation, Vol. 13. New York: Academic Press, 1979.

E. A. Feigenbaum 176
P41 RROO78S-09 Acquisition of Cognitive Procedures (ACT)

[8]

[9]

[10]

[11]

[12]

[13]

{14]

(15)

[16]

[17]

Il.

Anderson, J.R., Kline, P.J., & Beasley, C.M. Complex Learning. In R.
Snow, P.A. Frederico, 2 W. Montague (Eds.). Aptitude, Learning, and
Instruction: Cognitive Processes Analyses. Hillsdale, N.J.: Lawrence
Erlbaum Assoc., 1980.

 

Anderson, J.R. & Kline, P.J. A learning system and its psychological
implications. In the Proceedings of the Sixth International Joint

Conference on Artificial Intelligence, 1979.

 

 

 

Reder, L.M. & Anderson, J.R. Use of thematic information to speed
search of semantic nets. Proceedings of the Sixth International Joint
Conference on Artificial Intelligence, 1979, 708-710.

 

 

 

Neves, D.M. & Anderson, J.R. Knowledge compilation: Mechanisms for
the automatization of cognitive skills. In J.R. Anderson (Ed.),
Cognitive Skills and their Acquisition. Hillsdale, N.J.: Lawrence

Erlbaum Associates, 1981.

Anderson, J.R., Greeno, J.G., Kline, P.J., & Neves, D.M. Acquisition
of Problem-solving skill. In J. R. Anderson (Ed.), Cognitive Skills

and their Acquisition. Hillsdale, N.J.: Lawrence Erlbaum Associates,
1981.

Anderson,J.R. Tuning of search of the problem space for geometry
proofs. Proceedings of IJCAI-81,1981,165-170.

 

Anderson, J.R. A theory of language acquisition based on general
learning principles.Proceedings of IJCAI-81,1981,97-170.

 

Anderson,J.R. A general learning theory and its application to the
acquisition of proof skills in geometry. In R.Michalski, J.Carbonell,
and T. Mitchell(Eds) Machine Learning, forthcoming.

Anderson, J.r. A.C.T. Cambridge, MA>: Harvard University press, 1983.

Kline, P. J. The superiority of relative criteria in partial matching
and generalization. Proceedings of IJCAI-81, 1981,296-303.

 

E. Funding Support

An information-processing analysis of learning in geometry
John R. Anderson, Principal Investigator

National Science Foundation (IST-80-15357)

$186,000 Feb 15, 1981 - Feb 15, 1984

INTERACTION WITH THE SUMEX-AIM RESOURCE

A. Collaborations, Interactions, and Sharing of Programs via SUMEX.

We have received and answered many inquiries about the ACT system

over the ARPANET. This involves sending documentations, papers, and copies

177 E. A. Feigenbaum
Acquisition of Cognitive Procedures (ACT) P41 RROO785-09

of programs. The most extensive collaboration has been with Greeno and
Lesgold who were also on SUMEX.

We find the SUMEX-AIM workshops ideal vehicles for updating ourselves
on the field and for getting to talk to colleagues about aspects of their
work of importance to us.

Due to memory space problems encountered by ACT we expect that soon
we will need to make use of the smaller version of INTERLISP developed at
SUMEX for use in the CONGEN program.

B. Critique of Resource Management

The SUMEX-AIM resource has been well suited for the needs of our
project. We have made the most extensive use of the INTERLISP facilities
and the facilities for communication on the ARPANET. We have found the
SUMEX personnel extremely helpful both in terms of responding to our
immediate emergencies and in providing advice helpful to the long-range
progress of the project. Despite the fact that we are not located at
Stanford, we have not encountered any serious difficulties in using the
SUMEX system; in fact, there are real advantages in being in the Eastern
time zone where we can take advantage of the low load on the system during
the morning hours. We have been able to get a great deal of work done
during these hours and try to save our computer-intensive work for this
time.

Two location changes by the ACT project (from Michigan to Yale in the
summer of 1976 and from Yale to Carnegie-Mellon in the summer of 1978) have
demonstrated another advantage of working on SUMEX: In both cases we were
back to work on SUMEX the day after our arrival.

III. RESEARCH PLANS (8/80-7/86)

 

A. Project Goals and Plans

Qur long-range goals are: (1) Continued development of the ACT
system; (2) Application of the system to modeling of various cognitive
processes; (3) Dissemination of the ACT system to the national AI
community.

This is a period of major evolution for the ACT theory. We have been
developing three special versions of the ACTF learning that allow us to
more efficiently simulate learning in three domains: proving theorems in
geometry, speaking a new language, and writing programs in LISP. We are
also performing special purpose simulations of the processes of spreading
activation in memory retrieval and of pattern-matching processes in
reading. We will be assimilating our experiences with these special
purpose simulations in putting forth a major revision of the ACT theory. A
research monograph is being written setting forth this theory and is
scheduled for publication in Spring,1983. Subsequent to the writing of
this monograph we intend to implement the ACT* successor to ACTF that will
embody the new conceptions.

E. A. Feigenbaum 178
P41 RROO785-09 Acquisition of Cognitive Procedures (ACT)

B. Justification for Continued Use of SUMEX:

Our goal for the ACT system is that it should serve as a ready-made
"programming language" available to members of the cognitive science
community for assembling psychologically-accurate simulations of a wide
range of cognitive processes. Our intention and ability to provide such a
resource justifies our use of the SUMEX facility. This facility is
designed expressly for the purpose of developing and supporting such
national AI resources and is, in this regard, clearly superior to the
facilities we have available locally from the Carnegie-Mellon computer
science department. Among the most important SUMEX advantages are the
availability of INTERLISP on a machine accessible by either the ARPANET or
TYMNET and the existence of a GUEST login. It appears that, at least for
the time being, ACT has no hope of being a national resource unless it
resides at SUMEX.

C. Needs and Plans for Other Computational Resources

Carnegie-Mellon’s is upgrading its PDP-1i0 hardware to emerging state-
of-the-art machines (PERQ,VAX, LISP machines, etc.) promises to provide a
excellent resource, and we hope to have access to that resource as it
develops. A major decision that needs to be made in the next 6 months is
whether to implement the ACT* system locally or on SUMEX.

D. Comments and Suggestions for Future Resource Goals

We are beginning to feel squeezed by various limitations of the SUMEX
facility. The problem of peak load is quite serious. We have also been
struggling with the address limitations of the current INTERLISP which is
made more grievous by the amount of space INTERLISP requires. The
computation time and address space limitations have meant that we have not
been able to pursue certain projects that we would have otherwise. We
applaud any efforts to increase computational power, to increase the
address space of INTERLISP (e.g. VAXes), or to create significantly more
space efficient versions of INTERLISP.

It seems clear that the eventual trend will be to do more and more
computation locally on the ever more powerful and cheaper machines. SUMEX
could help us in the choices we are facing in terms of local computation.
If we could have experimental access to the new machines like Dolphins,
related software, and software advise, it would be easier to make decisions
about new resources. It would also be very helpful to develop programs on
the SUMEX machines which could eventually be transferred to local machines
when acquired. Finally, and perhaps most important, it seems important to
make an effort to promote communication among local machines to maintain
the community that has been built up through SUMEX.

179 E. A. Feigenbaum
CADUCEUS Project (INTERNIST) P41 RROO785-09

II.A.2.2 CADUCEUS Project (INTERNIST)

 

CADUCEUS Project
J. D. Myers, M.D. and H. Pople, Ph.D.

University of Pittsburgh
Pittsburgh, Pennsylvania

I. SUMMARY OF RESEARCH PROGRAM

 

A. Medical Rationale

The principal objective of this project is the development of a high-
level computer diagnostic program in the broad field of internal medicine
as an aid in the solution of complex and complicated diagnostic problems.
To be effective, the program must be capable of multiple diagnoses (related
or independent) in a given patient.

A major achievement of this research undertaking has been the design
of a program called INTERNIST-I, along with an extensive medical knowledge
base. This program has been used over the past 10 years to analyze many
hundreds of difficult diagnostic problems in the field of internal
medicine. These problem cases have included cases published in medical
journals (particularly Case Records of the Massachusetts General Hospital,
in the New England Journal of Medicine), CPCs, and unusual problems of
patients in our Medical Center. In most instances, but by no means all,
INTERNIST-I has performed at the level of the skilled internist, but the
experience has high-lighted several areas for improvement.

The INTERNIST-I system, in complex cases with multiple diagnoses,
operates in a sequential manner. The leading diagnostic hypothesis, i.e.
the one with the most supportive clinical evidence both positive and nega-
tive, is worked on initially until confirmed or downgraded and replaced by
a second hypothesis which is now the stronger. When the first diagnosis is
confirmed, the system awards credit to all other diseases which are related
("linked") to the confirmed diagnosis so as to favor, as is logical, a set
of related diagnoses rather than independent ones. This sequential
analysis can be time consuming and at times involves "false starts." The
skilled clinician, instead of focusing on disease A, confirming it, and
then moving on to related disease B, etc., will consider from the onset of
hypothesis of disease A plus disease B, in other words a diagnostic
complex. We believe that our second generation system, CADUCEUS, can be
built to accomplish this same form of logical analysis. To do so requires
refinement of the computer diagnostic algorithms as well as some
reorganization of the medical knowledge base.

The concept of "facets" of diseases has been introduced. Facets are
selected pathophysiologic processes or concepts which apply to a broad
range of individual diseases. They include such items as cholestasis,
coagulopathy of liver disease, acute hemorrhage, and respiratory acidosis.

E. A. Feigenbaum 180
P4i RROO785-09 CADUCEUS Project (INTERNIST)

Facets are comprised of individual clinical manifestations of disease, and
the number of the latter in a facet varies from a few to as many as
fifteen. A more general or higher level facet may include a number of more
specific or limited facets in its structure. Drs. Myers and Miller have
reorganized the 57 individual hepatobiliary diseases profiled in the
medical knowledge base into facets. Some manifestations of disease are
specific predisposing factors and others are specific findings, e.g.
bacterial cultures, but the vast majority of manifestations are subsumed
into facets. It is reassuring that, taking facets, predisposing factors
and specific findings together, all significant manifestations of each of
the 57 hepatobiliary diseases are included in the reorganization. The
number of facets per disease varies from a half-dozen to more than a dozen.

The use of facets provides certain advantages. The diagnostic analy-
sis starts from broader concepts and more basic pathophysiological and
patho-biochemical ones. The use of facets should overcome a present
disadvantage of INTERNIST-I, namely that the first diagnosis with which a
clinical finding is compatible explains that finding even though some other
disease present in the patient is a better and more potent explanation. In
the new system, individual manifestations are often assigned to multiple
facets and thus may be utilized more prudently in the diagnostic analysis.

Manifestations, as well as facets, are being modified so as to
portray important anatomical information, a point often underplayed in
INTERNIST-I.

It is planned that CADUCEUS will as well include a better time axis
in regard to the development of a disease process or the sequence of
diseases.

B. Medical Relevance and Collaboration

The program inherently has direct and substantial medical rele-vance.

The institution of collaborative studies with other institutions has
been deferred pending completion of the programs and knowledge base en-
hancements required for CADUCEUS. The installation of our own, dedicated
VAX computer will be expected to aid considerably any future collaboration.

C. Highlights of Research Progress

Accomplishments This Past Year:

 

During the past year, there has been continued expansion of the
CADUCEUS knowledge base within the context of the original INTERNIST-I
system, which is operational at SUMEX-AIM. In addition, there has been a
concerted effort to develop an in-house capability to enable programs
originally written in INTERLISP to be run in our new local VAX/FRANZLISP
environment. To this end, we have developed a set of macro capabilities
that handle the major dissimilarities of these two LISP host processors.
In particular, we now have within FRANZLISP reasonably complete CLISP and
RECORD package facilities.

181 E. A. Feigenbaum
CADUCEUS Project CINTERNIST) P41 RROO785-09

Our objective has been to make operational in the local setting the
new CADUCEUS diagnostic programs and knowledge base editors, which had been
developed at SUMEX. At this writing, this conversion process is virtually
complete. Because of extensive machine dependencies in the original
INTERNIST-I code, and because of known deficiencies in the logic and
structure of that program there are no current plans to convert INTERNIST-I
to run in the VAX/FRANZLISP environment.

The modified manifestations and the newly devised facets are in
process of entry into our new VAX computer, and extensive testing of the
reorganized medical knowledge base and the computer operating program (in
FRANZLISP) is planned for the next year.

The medical knowledge base has continued to grow both in the
incorporation of new diseases and the modification of diseases already
profiled so as to include recent advances in medical knowledge. The
knowledge base now in-cludes 536 individual disease profiles, 3,679
Manifestations of disease, and about 1,500 "links" or interrelationships
among diseases as well as a myriad of miscellaneous pieces of information
which are essential for the correct operation of the system. Twenty-two
new diseases have been profiled during the past year and the pediatrics
knowledge base has continued to grow as well. In addition, neurologists
Gordon Banks and Bruce Weimer, who have recently become associated with the
project, have begun developing CADUCEUS knowledge structures for selected
problem areas in neurology.

Research in Progress:

There are five major components to the continuation of this research
project:

1) The completion, continued updating, refinement and
testing of the extensive medical knowledge base required
for the operation of INTERNIST-I.

2) The completion and implementation of the improved
diagnostic consulting program, CADUCEUS, which has been
designed to overcome certain performance problems
identified during the past five years’ experience with
the original INTERNIST-I progran.

3) Institution of field trials of CADUCEUS on the clinical
services in internal medicine at the Health Center of the
University of Pittsburgh.

4) Expansion of the clinical field trials to other
university health centers which have expressed interest
in working with the system.

5) Adaptation of the diagnostic program and data base of

CADUCEUS to subserve educational purposes and the
evaluation of clinical performance and competence.

E. A. Feigenbaum 182
P41 RROO785-09 CADUCEUS Project (INTERNIST)

Current activity is devoted mainly to the first two of these, namely,
the continued development of the medical knowledge base, and the
implementation of the improved diagnostic consulting program.

D. List of Relevant Publications

1. Pople, H.E. “The Formation of Composite Hypotheses in Diag-nostic
Problem Solving: An Exercise in Synthetic Reasoning", Proceedings of
the Fifth International Joint Conference on Artificial Intelligence,
Boston, August 1977.

2. Pople, H.E. "On the Knowledge Acquisition Process in Applied A.I.
Systems", Report of Panel on Applications of A.I., Proceedings of Fifth
International Joint Conference on Arti-ficial Intelligence, 1977.

3. Pople, H.E., Myers, J. D. & Miller, R.A. "The DIALOG Model of
Diagnostic Logic and its Use in Internal Medicine, Proceedings of the
Fourth International Joint Conference on Artificial Intelligence,
Tbilisi, USSR, September 1975.

4. Pople, H.E. "Artificial Intelligence Approaches to Computer-based
Medical Consultation, Proceedings IEEE Intercon, New York, 1975.

5. Myers, J. D. "The Process of Clinical Diagnosis and Its Adaptation to
the Computer," in The Logic of Discovery and Diagnosis in Medicine,
University of Pittsburgh Series in the Philosophy and History of
Science, University of California Press (in press).

6. Myers, J. D., Pople, H. E. & Miller, R. A. “INTERNIST: Can Artificial
Intelligence Help?" in Clinical Decision and Laboratory Use, University
of Minnesota Press (in press).

7. Pople, H. E. "Coming to Grips with the Multiple Diagnosis Problem," in
Computer-Assisted Decision Making Using Clinical and Paraclinical
(Laboratory) Data. B. Statland &@S. Bauer (eds.) Mediad Inc.,
Tarrytown, N. Y., 1980, pp. 81-88. Reprinted in The Logic of Discovery
and Diagnosis in Medicine, University of Pittsburgh Series in the
Philosophy and History of Science, University of California Press (in
press).

8. Pople, H. E. “Heuristic Methods for Imposing Structure on I11l-
structured Problems: The Structuring of Medical Diagnostics," in
Artificial Intelligence in Medicine, AAAS Symposium Series, Westview
Press (forthcoming 1982).

183 E. A. Feigenbaum
CADUCEUS Project (INTERNIST) P41 RROO785-09

E. Funding Support

1.

E. A. Feigenbaum

Clinical Decision Systems Research Resource

Harry E. Pople, Jr., Ph.D.
Associate Professor Business

Jack D. Myers, M.D.
University Professor (Medicine)
University of Pittsburgh

Division of Research Resources
National Institutes of Health

5 R24 RRO1101-05

07/01/80 - 06/30/85
$1,607,717

07/01/81 - 06/30/82
$202 ,632

INTERNIST: A Computer-Based Diagnostic Consultant

Harry E. Pople, Jr., Ph.D.
Associate Professor of Business

Jack D. Myers, M.D.
University Professor (Medicine)
University of Pittsburgh

National Library of Medicine
National Institutes of Health

5 ROL LMO03710-02

07/01/80 - 06/30/85
$817,884

07/01/81 - 06/30/82
$126,746

184
P41 RROO785-09 CADUCEUS Project (INTERNIST)

3. New Computer-Based Patient Case Simulator
Randolph A. Miller, M.D.
Assistant Professor of Medicine
University of Pittsburgh

National Library of Medicine - New Investigator
National Institutes of Health

5 R23 LM03589-02

07/01/80 - 06/30/83
$87 ,005

07/01/81 - 06/30/82
$29,555

4. Neurologic Consultation Computer Program
Gordon E. Banks, M.D.
Assistant Professor of Medicine
University of Pittsburgh
National Library of Medicine - New Investigator
National Institutes of Health
5 R23 LMO3889-01

04/01/82 - 03/31/85
$107,675

04/01/82 - 03/31/83
$35 ,725

IT. INTERACTIONS WITH THE SUMEX-AIM RESOURCE

 

A, B. Collaborations and Medical Use of Program Via SUMEX

CADUCEUS remains in a stage of research and development. As noted
above, we are continuing to develop better computer programs to operate the
diagnostic system, and the knowledge base cannot be used very effectively
for collaborative pur-poses until it has reached a critical stage of
completion. These factors have stifled collaboration via SUMEX up to this
point and will continue to do so for the next year or two. In the
meanwhile, through the SUMEX com~munity there continues to be an exchange
of information and states of pro-gress. Such interactions particularly
take place at the annual AIM Work-shop.

185 E. A. Feigenbaum
CADUCEUS Project (INTERNIST) P41 RROO785-09

C. Critique of Resource Management

SUMEX has been an excellent resource for the development of CADUCEUS
Our large program is handled efficiently, effectively and accurately. The
staff at SUMEX have been uniformly supportive, cooperative, and innovative
in connection with our project’s needs.

III. RESEARCH PLANS (7/81-6/86)

 

A. Project Goals and Plans

The prototype CADUCEUS programs and the trial reorganization of the
liver and biliary tract diseases will be tested in the VAX/FRANZLISP
environment over the summer and fall of this year. As rapidly as possible
and pending further refinement and reorganization from experience with the
new system, the remainder of the medical knowledge base will be entered.
Local and later collaborative field trials must necessarily be postponed
until this development has been accomplished.

At least 200 important medical diseases remain to be pro-grammed.
Renewed effort in this direction is now being expanded now that other tasks
have been surmounted. Expanded efforts in the fields of neurology and
pediatrics are included as described above.

B. Justification and Requirements for Continued SUMEX Use

Our use of SUMEX will obviously decline with the installation of our
VAX. Nevertheless, the excellent facilities of SUMEX are expected to be
used for certain developmental work. It is intended, further, to keep
INTERNIST-I at SUMEX for comparative use as CADUCEUS is developed here.
Our team hopes to remain as a component of the SUMEX community and to share
experiences and developments.

C. Needs and Plans for Other Computing Resources Beyond SUMEX-AIM

Our predictable needs in this area will be met by the dedicated VAX
computer recently installed.

D. Recommendations for Future Community and Resource Development

Whether a program like CADUCEUS, when mature, will be better operated
from centralized, larger computers or from the developing self-contained
personal computers is difficult to predict. For the foreseeable future it
would seem that centralized, advanced facilities like SUMEX will be
important in further program development and refinement.

E. A. Feigenbaum 186
P41 RROO785-09 Hierarchical Models of Human Cognition
II.A.2.3 Hierarchical Models of Human Cognition
Hierarchical Models of Human Cognition (CLIPR Project)
Walter Kintsch and Peter G. Polson

University of Colorado
Boulder, Colorado

I. SUMMARY OF RESEARCH PROGRAM

 

A. Project Rationale

The two CLIPR projects have made progress in their research in this
past year. This progress is almost completely due to our access to the
SUMEX facility. The prose comprehension group has completed one major
project, and is currently interacting with other SUMEX projects with the
goal of building a prose comprehension model that reflects state-of-the-art
knowledge from psychology and artificial intelligence.

The main activity of the planning group during the last year has been
completion of the analysis of thinking-out-loud protocols collected from
both expert and novice software designers. SUMEX facilities have been used
to store, edit, and reformat the raw protocols to facilitate later
analysis. Results of successive analyses are then input to SUMEX, and SUMEX
facilities are used to collate the various results. In addition, we have
just started a new project studying the psychological factors underlying
device complexity and the difficulties that nontechnically trained
individuals have in learning to use devices like work processors.

Technical Goals:

The CLIPR project consists of two subprojects. The first, the text
comprehension project, is headed by Walter Kintsch and is a continuation of
work on understanding of connected discourse that has been underway in
Kintsch’s laboratory for over eight years. The second, the planning
project, is headed by Peter Polson of the University of Colorado and
Michael Atwood of Bell Telephone Laboratories,Denver, formerly of Science
Applications Incorporated, Denver, and is studying the processes of
planning using software design tasks. In addition, Polson in collaboration
with David Kieras of the University of Arizona, Tucson, is studying the
learning and problem solving processes involved in the utilization of
devices like word processors.

The goal of the prose comprehension project is to develop a computer
system capable of the meaningful processing of prose. This work has been
generally guided by the prose comprehension model discussed by Kintsch and
van Dijk (1978), although our programming efforts have identified necessary
clarifications and modifications in that model (Miller & Kintsch, 1980,
1981; Kintsch & Miller, 1981; Miller, 1982). In general, this research has
emphasized the importance of knowledge and knowledge-based processes in

187 E. A. Feigenbaum
Hierarchical Models of Human Cognition P41 RROO78S-09

comprehension, and we are accordingly working with the AGE and UNITS groups
at SUMEX toward the development of a knowledge-based, blackboard model of
prose comprehension. We hope to be able to merge the substantial artificial
intelligence research on these systems with psychological interpretations
of prose comprehension, resulting in a computational model that is also
psychologically respectable.

The primary goal of the planning project is the development of a
model of human performance on software design tasks. We have concentrated
on developing a deeper understanding of our protocol data, to increase our
knowledge of the details of the planning processes and the knowledge
structures that experts use in the process of planning. We have developed a
Method of protocol analysis that essentially involves the transforming of
the protocol into a low level theoretical description of the processes used
to solve the design problem. We have assumed a very simplified version of a
blackboard model that is described in Jeffries, Turner, Polson, and Atwood
(1981). We currently carry out our analysis by hand, developing a form of
this low level model for each protocol.

The goal of the device complexity project is to develop explicit
models of the user-device interaction. Our current plan is to model the
device as a nested automata and the user as a production system. The
purpose of developing such models would be to make explicit kinds of
knowledge that are required to operate different kinds of devices and to
specify the processing loads imposed by different implementations of a
device. We feel that this offer are tools that are being developed at
SUMEX--in particular AGE and the UNIT package--will dramatically facilitate
our abilities to generate such models of the user-device interface.

B. Medical Relevance and Collaboration

The text comprehension project impacts indirectly on medicine, as the
medical profession is no stranger to the problems of the information glut.
By adding to the research on how computer systems might understand and
summarize texts, and determining ways by which the readability of texts can
be improved, medicine can only be helped by research on how people
understand prose. Development of a more thorough understanding of the
various processes responsible for different types of learning problems in
children and the corresponding development of a successful remediation
strategy would also be facilitated by an explicit theory of the normal
comprehension process.

Note that our goal of a blackboard model is particularly relevant to
the understanding of learning difficulties. One important aspect of a
blackboard model is the separation of cognitive processes into a set of
interacting subprocesses. Once such subprocesses have been identified and
constructed, it would be instructive to observe the model’s performance
when certain of these processes are facilitated or inhibited. Many
researchers have shown that there are a variety of cognitive deficits
(insufficient short-term memory capacity, poor long-term memory retrieval,
and such) that can lead to reading problems. Having a blackboard model in
which the power of individual components could be manipulated would be a
significant step in determining the nature of such reading problems.

E. A. Feigenbaum 188
P41 RROO785-09 Hierarchical Models of Human Cognition

The planning project is attempting to gain understanding of the
cognitive mechanisms involved in design and planning tasks. The knowledge
gained in such research should be directly relevant to a better
understanding of the processes involved in medical policy making and in the
design of complex experiments. We are currently using the task of software
design to describe the processes underlying more general planning
mechanisms that are also used in a large number of task oriented
environments like policy making.

Both the text comprehension project and the planning project involve
the development of explicit models of complex cognitive processes;
cognitive modelling is a stated goal of both SUMEX and research supported
by NIMH.

The on-going development of the prose comprehension model would not
be possible without our collaboration with the AGE and UNITS research
groups. We look forward to a continued collaboration, with, we hope,
mutually beneficial results. Several other psychologists have either used
or shown an interest in using an early version of the prose comprehension
model, including Alan Lesgold of SUMEX’s SCP project, who is exporting the
system to the LRDC vax. We have also worked with James Greeno -- another
member of the SCP project -- on a project that will integrate this model
with models of problem solving developed by Greeno and others at the
University of Pittsburgh. Needless to say, all of this interaction has
been greatly facilitated by the local and network-wide communication
systems supported by SUMEX. There has been considerable communication
between members of the prose comprehension and AGE/UNITS groups as program
bugs have been discovered and corrected; the presence of a mail system has
made this process infinitely easier than if telephone or surface mail
Messages were required. The mail system, of course, has also enabled us to
maintain professional contacts established at conferences and other
meetings, and to share and discuss ideas with these contacts.

C. Progress Summary

The prose comprehension project has completed an initial version of a
model of prose comprehension (Miller & Kintsch, 1980). This model has been
applied to a large number of texts, and has yielded quite reasonable
predictions of recall and readability. Psychologists from other
universities have used this system to derive reading time and recall
predictions for their own experimental materials. We are currently using
the AGE and UNITS packages to extend this model toward one that can make
use of world knowledge in its analyses; this model is discussed in Miller
and Kintsch (1981) and Miller (1982).

The planning group has completed the detailed analysis of several
long thinking-out-loud protocols collected from both expert and novice
software designers. These analyses involved the development of a lower
level model for each of the protocols. See Atwood and Jeffries (1980) for
details and examples.

189 E. A. Feigenbaum
Hierarchical Models of Human Cognition P41 RROO7&5-09

The device complexity project is just getting underway. One of our
primary objectives for this year is to develop an explicit model for the
knowledge structures involved in the user-device interaction and we will
then consider designs for possible simulation programs.

D. List of Relevant Publications

Atwood, M. E., & Jeffries, R. Studies in plan construction I: Analysis of
an extended protocol. Technical Report SAI-80-028-DEN, Science
Applications, Incorporated, Denver, Co. March, 1980.

Atwood, M. E., & Jeffries, R. Studies in plan construction II: Novice
design behavior. Technical Report SAI-80-154-DEN, Science Applications,
Incorporated, Denver, Colorado, December, 1980.

Atwood, M. E., Polson, P. G., Jeffries, R., and Ramsey, H. R. Planning as
a process of synthesis. Technical Report SAI-78-144-DEN, Science
Applications, Incorporated, Denver, Co. December, 1978.

Jeffries, R., Turner, A., , Polson, P. G., & Atwood, M. E. The process of
designing software. J. R. Anderson (Ed.), Cognitive skills and their
acquisition. Hillsdale, N.J.: Erlbaum, 1981 ,254-285.

Kieras, D. E. and Polson, P. G. An outline of a theory of the user
complexity of devices and systems. Working Paper No. 1, Device
Complexity Project, Universities of Arizona and Colorado, May 1982.

Kintsch, W. On modelling comprehension. Invited address at the American
Educational Research Association convention. San Francisco, April 10,
1979.

Kintsch, W. and van Dijk, T. A. Toward a model of text comprehension and
production. Psychological Review, 1978, 85, 363-394.

Miller, J. R., & Kintsch, W. Readability and recall of short prose
passages: A theoretical analysis. Journal of Experimental Psychology:
Human Learning and Memory, 1980, 6, 335-354.

Miller, J. R., & Kintsch, W. Readability and recall of short prose
passages. Text, 1981, 1, 215-232.

Miller, J. R. A knowledge-based mode of prosa comprehension: Applications
to expository text. Paper presented at the American Educational
Research Association meeting, April, 1981.

Kintsch, W., & Miller, J. R. Knowledge-based processes in prose
comprehension. Paper presented at the 21st annual meeting of the
Psychonomics Society, St. Louis, November, 1980.

Miller, J. R. A knowledge-based model of prose comprehension: Applications

to expository text. In B. K. Britton and J. B. Black (Eds.),
Understanding expository text. Hillsdale, N. J.: Erlbaum, 1982.

E. A. Feigenbaum 190
P41 RROO785-09 Hierarchical Models of Human Cognition

E. Funding Support Status

1. Readability and Comprehension.
Walter Kintsch, Professor, University of Colorado
National Institute of Education
NIE-G-78-0172
9/1/78 - 8/31/81: $96,627
9/1/80 - 8/31/81: $46,537

2. Text Comprehension and Memory
Walter Kintsch, Professor, University of Colorado
National Institute of Mental Health
5 Rol MH15872-14-16
7/1/81 ~ 6/30/84: $148,693
7/1/81 - 6/30/82: $37,370

3. Comprehension and Analysis of Information in Text
Walter Kintsch, Professor, University of Colorado, and
Lyle E. Bourne, Jr., Professor, University of Colorado
Office of Naval Research, Personnel and Training Programs
ONR NO00014-78-C-0433
6/1/78 - 6/30/82: $225,266
7/18/81 - 6/30/82: $78,125

4. Procedural Net Theories of Human Planning and Problem Solving
Michael Atwood, Research Psychologist, Science Applications,
Incorporated, Denver, Colorado
Qffice of Naval Research, Personnel and Training Programs
ONR NOO14-78-C-0165
1/25/78 - 12/31/80: $230,000
1/1/80 - 6/30/81: $85,000

5. User Complexity of Devices and Systems
David Kieras, Associate Professor, University of Arizona
Peter G. Polson, Professor, University of Colorado
International Business Machines Corporation
1/1/82-12/31/82: $104,000

IT. INTERACTIONS WITH THE SUMEX-AIM RESOURCE
A. Sharing and Interactions with Other SUMEX-AIM Projects

Our primary interaction with the SUMEX community has been the work of
the prose comprehension group with the AGE and UNITS projects at SUMEX.
Feigenbaum and Nii have visited Colorado, and one of us (Miller) recently
attended the AGE workshop at SUMEX. Both of these meetings have been very
valuable in increasing our understanding of how our problems might best be
solved by the various systems available at SUMEX. We also hope that our
experiments with the AGE and UNITS packages have been helpful to the
development of those projects.

igt E. A. Feigenbaum
Hierarchical Models of Human Cognition P4i RROO785-09

We should also mention theoretical and experimental insights that we
have received from Alan Lesgold and other members of the SUMEX SCP project.
The initial comprehension model (Miller & Kintsch, 1980) has been used by
Dr. Lesgold and other researchers at the Universivy of Pittsburgh, as well
as researchers at Carnegie-Mellon University, the University of Manitoba,
Rockefeller University, and the University of Victoria.

B. Critique of Resource Management

The SUMEX~AIM resource is clearly suitable for the current and future
needs of our project. We have found the staff of SUMEX to be cooperative
and effective in dealing with special requirements and in responding to our
questions. The facilities for communication on the ARPANET have also
facilitated collaborative work with investigators throughout the country.

III. RESEARCH PLANS
A. Long Range Projects Goals and Plans

The primary long-term goal of the prose comprehension group is the
development of a blackboard-based model of prose comprehension.
Correspondingly, we anticipate continued use of the AGE and UNITS packages.
These packages allow us to model the knowledge structures possessed by
people and the inferential processes that operate upon those structures,
and are essential to our work.

The primary goal of the device complexity project is the development
of a theory of the processes and knowledge structures that are involved in
the performance of routine cognitive skills making use of devices like word
processors. We plan to model the user-device interaction by representing
the users processes and knowledge as a production system and the device as
@ hested automata. We are also studying the role of mental models in
learning how to use then.

B. Justification and Requirements for Continued SUMEX Use

The research of the prose comprehension project is clearly tied to
continued access to the AGE and UNITS packages, which are simply not
available elsewhere. We hope that our continued use of these systems will
be offset by the input we have been and will continue to provide to those
projects: our relationship has been symbiotic, and we look forward to its
continuation.

C. Needs and Plans for Other Computational Resources

We currently use two other computing systems located at the
University of Colorado. One is the Department of Psychology'’s VAX 11/780,
which is used primarily to run real-time experiments to be modeled on
SUMEX. The second is the University of Colorado’s CDC 6400, which is used
for various types of statistical analysis.

E. A. Feigenbaum 192
P41 RROO785-09 Hierarchical Models of Human Cognition

When the ARPA-sponsored Vax/Interlisp project is completed, we would
be most interested in experimenting with becoming a remote AGE/UNITS site.
It would seem that this sort of development is the ultimate goal of the
package projects, and this type of interaction, once it becomes feasible,
would be a logical extension of our association with the SUMEX facility.

D. Recommendations for Future Community and Resource Development

Our primary recommendation for future development within SUMEX
involves (a) the continued support of INTERLISP, which is needed for AGE
and for other work we have underway on SUMEX and (b) the continued
development of the AGE and UNITS projects. In particular, we would like to
see an extension of AGE to include a wider variety of control structures so
that our psychological models would not be confined to one particular view
of knowledge-based processing. The limited physical capacity of SUMEX,
both in terms of address space and overloading, is, as before, a major
problem. The prose comprehension group can no longer use the publicly
released AGE/UNITS system due to its severely limited address space, and
has had to build a personal AGE system from a stripped-down version of
Interlisp and a selected subset of AGE and UNITS. We heartily endorse the
plans underway to obtain more computing capacity for the SUMEX project.

Given our acquisition of a VAX, we particularly support the ongoing
and continued development of INTERLISP for the VAX, so that local use of
AGE and UNITS would be possible. Since we, as well as other psychologists,
need the real-time capability of VAX/VMS to run on-line experiments, we
hope that the INTERLISP system to be developed will be compatible with VMS.
Note that this need for real-time work coincides with real-world
applications of SUMEX programs, in which a VAX might be devoted to both
real-time patient monitoring and diagnostic systems such as PUFF or MYCIN.

193 E. A. Feigenbaum
PUFF-VM Project P41 RROO785-09

IT.A.2.4 PUFF-VM Project

PUFF-VM: Biomedical Knowledge Engineering in Clinical Medicine

John J. Osborn, M.D.
The Medical Research Institute of San Francisco
Pacific Medical Center

and

Edward H. Shortliffe, M.D., Ph.D.
Department of General Internal Medicine
Stanford University Medical Center
Stanford University

The immediate goal of this project is to develop knowledge-based
programs to interpret physiological measurements made in clinical medicine.
The interpretations are intended to be used to aid in diagnostic decision
making and in therapeutic actions. The programs will operate within
medical domains which have well developed measurement technologies and
reasonably well understood procedures for interpretation of measured
results. The programs are:

(1) PUFF: the interpretation of standard pulmonary function
laboratory data which include measured flows, lung volumes,
pulmonary diffusion capacity and pulmonary mechanics, and

(2) VM: management of respiratory insufficiency 1n the intensive care
unit.

The second, but equally important, goal of this project is the
dissemination of Artificial Intelligence techniques and methodologies to
medical communities that are involved in computer aided medical diagnosis
and interpretation of patient data.

I. SUMMARY OF RESEARCH PROGRAM
PUFF :

A. Project Rationale

The task of PUFF program is to interpret standard measures of
pulmonary function. It is intended that PUFF produce a report for the
patient record, explaining the clinical significance of measured test
results. PUFF also must provide a diagnosis of the presence and severity

of pulmonary disease in terms of measured data, referral diagnosis, and
patient characteristics.

E. A. Feigenbaum 194
P41 RROO785-09 PUFF-VM Project

B. Medical Relevance and Collaboration

Interpretation of standard pulmonary function tests involves
attempting to identify the presence of obstructive airways disease (OAD:
indicated by reduced flow rates during forced exhalation), restrictive lung
disease (RLD: indicated by reduced lung volumes), and alveolar-capillary
diffusion defect (DD: indicated by reduced diffusivity of inhaled CO into
the blood). Obstruction and restriction may exist concurrently, and the
presence of one mediates the severity of the other. Obstruction of several
types can exist. In the laboratory at the Pacific Medical Center (PMC),
about 50 parameters are calculated from measurement of lung volumes, flow
rates, and diffusion capacity. In addition to these measurements, the
physician may also consider patient history and referral diagnosis in
interpreting the test results and diagnosing the presence and severity of
pulmonary disease.

Currently PUFF contains a set of about 400 physiologically based
interpretation "rules". Each rule is of the form "IF <condition> THEN
<conclusion>*. Each rule relates physiological measurements or states to a
conclusion about the physiological significance of the measurement or
state.

C. Progress Summary

The results of the PUFF system are reviewed in more detail in the
1978 SUMEX annual report and {3]. A version of the PUFF system is now in
routine daily use at Pacific Medical Center. Reports are reviewed by a
physician pulmonary physiologist. Over 85 % of the reports are accepted by
the physician without change; they are signed and entered into the patient
record. Most of the remaining reports are edited on-line to modify a small
point in the test interpretation.

The PUFF system was developed on SUMEX in INTERLISP. It was
converted to BASIC to run on a PDP11. It has interpreted the data of about
4000 patient cases since it became clinically operational late in 1978.

The hospital has paid the Medical Research Institute for each clinical test
interpreted by the system since it became fully operational. The program
has been extended and maintained by research personnel during this period.
PMC is now contracting to have PUFF converted to run on its own stand-alone
micro computer (in BASIC). At that time, PUFF will have been fully moved
from the research environment to the full responsibility of the end user.

D. Relevant Publications

[1] "A Physiological Rule-Based System for Interpreting Pulmonary Function
Test Results", J.C. Kunz, R.J. Fallat, D.H. McClung, B.A. Votteri,
J.S. Aikins, H.P. Nii, L.M. Fagan, E.A. Feigenbaum, HPP 78-154,
Stanford Heuristic Programming Project, 1978.

[2] "Prototypes: An Approach to Knowledge Representation for Hypothesis
yP PP & yP

Formation", Aikins, J.S., HPP-79-10 (working paper), 1979. Also Int.
Joint Conf. on Artif. Intell., Tokyo, Japan, August, 1979.

195 E. A. Feigenbaum
PUFF-VM Project P41 RROO785-09

{3] “A Physiological Rule-Based System for Interpreting Pulmonary Function
Test Results", J.C. Kunz, R.J. Fallat, D.H. McClung, B.A. Votteri,
J.S. Aikins, H.P. Nii, L.M. Fagan, E.A. Feigenbaum, Proceedings of
Computers in Critical Care and Pulmonary Medicine, IEEE Press, i979.

[4] "The Art of Artificial Intelligence: Themes and Case Studies of
Knowledge Engineering", E.A. Feigenbaum, Proceedings of the IJCAI,
(1977). (Also Stanford Computer Science Department Memo STAN-CS-77-
612).

IB

A. Project Rationale

The Ventilator Manager program (VM) interprets the clinical
significance of time varying quantitative physiological data from patients
in the ICU. This data is used to manage patients receiving ventilatory
assistance. An extension of a physiological monitoring system, VM (1)
provides a summary of the patient’s physiological status appropriate for
the clinician; (2) recognizes untoward events in the patient/machine system
and provides suggestions for corrective action; (3) suggests adjustments to
ventilatory therapy based on a long-term assessment of the patient status
and therapeutic goals; (4) detects possible measurement errors; and, (5)
maintains a set of patient-specific expectations and goals for future
evaluation. The program produces interpretations of the physiological
measurements over time, using a model of the therapeutic procedures in the
IcU and clinical knowledge about the diagnostic implications of the data.
These therapeutic guidelines are represented by a knowledge base of rules
created by clinicians with extensive ICU experience.

B. Medical Relevance and Collaboration

To assist in the interpretation process, VM must be able to recognize
unusual or unexpected clinical events (including machine malfunction) in a
manner specifically tailored to the patient in question. The interpretation
task is viewed as an ongoing process in the ICU, so that the physiological
measurements must be continually reevaluated producing a current clinical
picture.

This picture can then be compared with previous summary of patient
status to recognize changes in patient condition upon which therapy
selection and modifications can be made. The program must also determine
when the measurements are most likely to be sensitive to error or when
external measurements would be of diagnostic significance.

VM offers a new approach toward more accurate recognition of alarm
conditions by utilizing the history and situation of the patient in the
analysis. This is in contrast to the use of static limits applied to
Measurements generated to fit the “typical patient" under normal

E. A. Feigenbaum 196
P4i RROO785-09 PUFF-VM Project

conditions. Our program uses a model of interpretation process, including
the types and levels of conclusions drawn manually from the measurements to
provide a summary of patient condition and trends. The program generated
conclusions are stated at levels more abstract than the raw data; for
example, the presence of hemodynamic stability/instability rather than in
terms of heart rate and mean arterial pressure. When the data is not
reliable enough to make these conclusions, additional tests may be
suggested. The recognition of important conclusion for which external
verification is sought, will also elicit the suggestion for confirming
tests from the program.

C. Progress Summary

VM is provided with the values of 30 physiclogical measurements on a
2- or 10-minute bases by an automatic monitoring system. The output is in
the form of suggestions to clinicians and periodic summaries . During the
past year, the VM system was run extensively to validate the contents of
its knowledge base. The knowledge representation and system output are
discussed more extensively in the 1981 SUMEX report and in (Fagan 80).

During the past year, VM was tested on several dozen patients. Rule
modifications were made to improve system accuracy. Limits of expected
values were refined, and rules relating ventilatory management to
hemodynamic function were extended.

During the past year, basic research was started on knowledge
representation for a physiological model. This research project integrates
Artificial Intelligence (AI) and simple mathematics to analyze a
physiological model. In the domain of renal physiology, a computer program
is being implemented based on these techniques. It analyzes physiological
behavior, diagnoses abnormality, and explains the rationale for its
analyses. The program fits data to the model, identifies whether the data
are abnormal, and identifies the possible causes and effects of any
abnormalities. The physiological model is based on knowledge about
anatomy, the behavior of the physiological system, and the mechanism of
action of the system.

In this project, mathematics is used in an AI system to compute
quantitative values based on available data. Concurrently, AI enhances the
simple mathematical models. AI is used to represent diverse knowledge
about the problem, to match a solution technique with problems needing
solution, and to interpret results. Results may be quantitative and
qualitative. Results are explained in terms of their contribution to
diagnostic and therapeutic decisions.

During this year, work was also initiated to convert alarm rules from
VM into computer programs which routinely analyze patient monitoring data.
VM rules have been elaborated and recoded to detect instrument error in
respiratory and hemodynamic monitoring data. -Additional VM rules have been
recoded to detect simple physiological abnormalities. One of the original
VM goals was to provide prompt feedback to clinicians about the presence of
error conditions. Alarm checking rules have been incorporated into the

197 E. A. Feigenbaum
PUFF-VM Project P41 RROO785-09

operating data interpretation programs. Integrating these rules with the
production monitoring system software allows this feedback to be prompt and
associated naturally with the other monitoring system output.

D. Relevant Publications

Fagan, L.M., Kunz, J.C., Feigenbaum, E.A. and Osborn, J.J.: A symbolic
processing approach to measurement interpretation in the intensive care
unit. Proc. Third Annual Symposium Computer Applications in Medical
Care, Silver Spring, Maryland, October, 1979, pp. 30-33.

Fagan, L.M., Shortliffe, E.H. and Buchanan, B.G.: Computer-based medical
decision making: From MYCIN to VM. Automedica 3(2), 1980.

Fagan, L.M.: VM: Representing Time-Dependent Relations in a Medical
Setting, Ph.D. dissertation, Stanford University, 1980.

Osborn, J.J., Fagan, L.M., Fallat, R.J., et al: Managing the data from
respiratory measurements. Med. Instrumentation, November-December,
1979. (Winner of the "Best Article of the Year’ Award for AAMI - 1979.)

Sierra, D.: Development of a smart respiratory alarm system using
artificial intelligence (abstract), presented at AAMI meeting, May
1982.

Mitchell, R.R.: The Need for closed loop therapy. Accepted for
publication, Critical Care Medicine, 1982.

Mitchell, R.R., Wilson, R.M., Holzapfel, L., Benis, A.M., et al: The Oxygen
washing method for monitoring Functional Residual Capacity. Accepted
for publication, Critical Care Medicine, 1982.

E. Funding Support

PUFF-VM was supported by NIH grant GM24669 from 1978 - 1981. A
renewal application was approved and not funded.

Research is now supported by a $50,000 grant from the Johnson and
Johnson foundation, John J. Osborn, PI. This grant includes some indirect
costs. The grant is for one year.

II. INTERACTIONS WITH THE SUMEX-AIM RESOURCE
A. Collaborations and Medical Use of Programs via SUMEX

The PUFF-VM project requires very close collaboration between
investigators at two institutions separated by fifty miles. This kind of
collaboration, in which program development and testing proceeds
concurrently on the same application system, requires a computer network
facility for sharing of code, data and ideas. SUMEX has been used at PMC
for running programs developed concurrently by Stanford and PMC staff, and

E. A. Feigenbaum 198
P4i RROO785-09 PUFF-VM Project

data has been taken from the PMC computer system and transferred to SUMEX
on magnetic tape for program development and testing. The SUMEX staff has
developed a cooperating set of computer programs to allow the PMC computer
and the SUMEX/2020 systems to actively exchange files and program data and
output.

We also use the SUMEX system for purposes other than program
development. Joint PMC-Stanford reports on PUFF and VM are being prepared
entirely through the the communications and word processing capabilities of
SUMEX. Investigators from the two institutions have collaborated in
writing reports together; the separate contributions are prepared on SUMEX,
edited and merged with an exchange of messages but without ever requiring
actual meetings. We have also used the system for trading bibliographic
information with other AIM users. We have also experimentally run the
Internist program using SUMEX.

B. Sharing and Interactions with Other SUMEX-AIM Projects

We have participated in the AIM workshop and had very fruitful
interaction with a number of other SUMEX users, directly influencing our
perception of important problems and potentially appropriate solutions.
Personal contacts at other conferences, at Stanford AI weekly meetings, and
at PMC with visiting members of the AIM community, have also been very
helpful in keeping abreast of the current thinking of other members of the
AI community and with members of the medical community interested in
computer based physiological analysis and diagnosis. We believe that the
use of a common machine and the existence of the AIM conference encourages
increased recognition and better communication with other AIM workers.
Within AIM we most closely collaborate with the MYCIN, MOLGEN and DENDRAL
projects, who share common space, common techniques, and common attitudes.

C. Critique of Resource Management

The SUMEX community continues to be an extremely supportive
environment in which to do research on uses of artificial intelligence in
clinical medicine. The community has two equally vital resources -- the
people with knowledge and interest in AI and the facility on which AI
system development can proceed. They are equally excellent as resources,
helping hands when faced with problems, and friendly support for continued
productive research. The availability of INTERLISP; of a facility on which
routine data processing functions (eg. manipulating magnetic tapes and
making long listings) can take place; and of message-sending among remote
users are all vital functions for our project. SUMEX provides them in an
environment which is friendly and reliable. Management of the SUMEX
facility is consistent and excellent.

199 E. A. Feigenbaum
PUFF-VM Project P41 RROO785-09

III. RESEARCH PLANS
A. Long Range Goals and Plans

The main emphasis of this project changed in 1979 from the
development of the PUFF system to the development, extension and evaluation
of the VM system. That change was consistent with the transition of PUFF
from a research project to an operating program in a clinical setting. In
addition, we want to pursue some promising research opportunities in the
richer VM setting. Some long term interests, such as consensus building
between experts, may be examined using both application areas.

The long range clinical research goal of the VM project is to develop
and evaluate an interpretation system that will improve patient care in the
Icu. The basic research goal continues to be to develop new Artificial
Intelligence techniques which contribute to progress toward that clinical
research goal. Within our relatively limited current funding
opportunities, we plan to continue research toward these goals. We plan to
exploit the knowledge in the VM rules in the current generation of
monitoring system software and to develop improved AI techniques for doing
diagnosis using models of physiology.

The VM alarm detection rules will be refined and incorporated into
the operating monitoring system analysis programs. Some of these rules
could not be fully exploited with VM itself because, running at SUMEX, VM
typically received data only every 10 minutes. Incorporated into the
analysis routines, the alarm detection rules can immediately obtain
additional confirming data as necessary. This proximity to the data will
allow both improved rules and more effective clinical use of the available
rules.

Extending VM rules into the operating data analysis programs is an
activity which follows in the tradition of PUFF. The original problem was
analyzed and formalized in the computationally rich SUMEX environment.

Once the knowledge of the problem is relatively well understood then it can
be represented in the comparatively restrictive environments of BASIC on a
PDP11 (in the case of PUFF), or the C language on a microcomputer (in the
case of the VM alarm rules). This mode of transfer of knowledge seems to
be one important way for knowledge developed in AI systems to be brought
into wide use.

The near-term goals for the physiological modeling activity are to
continue to develop a theory of physiological modeling within an AI system
and to continue development of a computer program which demonstrates the
power of that theory. The computer program is designed to demonstrate the
following processing capabilities: identify the causes and effects of
abnormal values of routine physiological measurements; define an
abnormality (e.g. oliguria) identifying its possible causes, effects, and
associated therapeutic goals; identify the normal physiological
consequences of routine physiological behavior, such a drinking a fluid;
identify the physiological consequences of abnormal physiological or
anatomical situations.

E. A. Feigenbaum 200