5P41-RROO785-15 Description of Program Activities

ll. Description of Program Activities

This section corresponds to the predefined forms required by the Division of
Research Resources to provide information about our resource activities for their
computerized retrieval system. These forms have been submitted separately and are
not reproduced here to avoid redundancy with the more extensive narrative
information about our resource and progress provided in this report.

Il.A. Scientific Subprojects
Our core research and development activities are described starting on page 14, our

training activities are summarized starting on page 97, and the progress of our
collaborating projects is detailed starting on page 127.

ll.B. Books, Papers, and Abstracts

The list of recent publications for our core research and development work starts on

page 83 and those for the collaborating projects are in the individual reports starting
on page 127.

ll.C. Resource Summary Table

The details of resource usage, including a breakdown by the various subprojects, is
given in the tables starting on page 100.

3 E. H. Shortliffe
5P41-RROO785-15 Narrative Description

lil. Narrative Description
Ill.A. Summary of Research Progress

lil.A.1. Resource Overview

This is an annual report for year 15 of the SUMEX-AIM resource (grant RR-00785),
the second year of a 5-year renewal period to support further research on
applications of artificial intelligence in biomedicine. For the technical and
administrative reasons discussed in earlier reports, the SUMEX project now includes
the continuation of work on the development and dissemination of medical
consultation systems (ONCOCIN) that had been supported before 1986 as resource-
related research under grant RR-01631. Progress on core ONCOCIN research is
therefore now reported here as well.

These combined efforts represent an ambitious research program to:

e Continue our long-range core research efforts on knowledge-based
systems, aimed at developing new concepts and methodologies needed
for biomedical applications.

e Substantially extend ONCOCIN research on developing and disseminating
clinical decision support systems.

« Develop the core systems technology to move the national SUMEX-AIM
community from a dependence on the central SUMEX DEC 2060 to a
fully distributed, workstation-based computing environment.

« Introduce these systems technologies into the SUMEX-AIM community
with appropriate communications and managerial assistance to
responsibly phase out the central resource and DEC 2060 mainframe in
a manner that will support community efforts to become self-sustaining
and to continue scientific interactions through fully distributed means.

« Maintain our aggressive efforts at training and dissemination to help
exploit the research potential of this field.

lll.A.1.1. SUMEX-AIM as a Resource

SUMEX and the AIM Community

In the fifteen years since the SUMEX-AIM resource was established in late 1973,
computing technology and biomedical artificial intelligence research have undergone
a remarkable evolution and SUMEX has both influenced and responded to these
changing technologies. It is widely recognized that our resource has fostered highly
influential work in biomedical Al -- work from which much of the expert systems
field emerged -- and that it has simultaneously helped define the technological base
of applied Al research.

The focus of the SUMEX-AIM resource continues to emphasize research on artificial
intelligence techniques that guide the design of computer programs that can help
with the acquisition, representation, management, and utilization of the many forms of

5 E. H. Shortliffe
Resource Overview 5P41-RROO785-15

medical knowledge in diverse biomedical research and clinical. care settings
~~ ranging from biomolecular structure determination and analysis, to molecular
biology, to clinical decision support, to medical education. Nevertheless, we have
long recognized that the ultimate impact of this work in biomedicine will be realized
through its assimilation with the full range of methodologies of medical informatics,
such as data bases, biostatistics, human-computer interfaces, complex instrument
control, and modeling. From the start, SUMEX-AIM work has been grounded in real-
world applications, like systems for the interpretation of mass spectral information
about biomolecular structures, chemical synthesis, interpretation of x-ray diffraction
data on crystals, cognitive modeling, infectious disease diagnosis and therapy, DNA
sequence analysis, experiment planning and interpretation in molecular biology, and
medical instruction. Our current work extends this emphasis in application domains
such as oncology protocol management, clinical decision support, protein structure
analysis, and data base information retrieval and analysis. All of these research
efforts have demanded close collaborations with diverse parts of the biomedical
research community and the integration of many computational methods from those
domains with knowledge-based approaches. Even though in the beginning the "Al-
in-medicine" community was quite small, it is perforce no longer limited and easily-
defined, but rather is spreading and is inextricably linked with the many biomedical
applications communities we have collaborated with over the years. Driven both by
the on-going diffusion of Al and by the development of personal computer
workstations that signal the practical decentralization of computing resources, we
must develop new resource communication and distributed computing technologies
that will continue to facilitate wider intra- and inter-community communication,
collaboration, and sharing of biomedical information.

The SUMEX Project has demonstrated that it is possible to operate a computing
research resource with a national charter and that the services providable over
networks were those that facilitate the growth of Al-in-Medicine. SUMEX now has a
reputation as a model national resource, pulling together the best available interactive
computing technology, software, and computer communications in the service of a
national scientific community. Planning groups for national facilities in cognitive
science, computer science, and biomathematical medeling have discussed and
studied the SUMEX model and new resources, like the recently instituted BIONET
resource for molecular biologists, are closely patterned after the SUMEX example.

The projects SUMEX supports have generally required substantial computing
resources with excellent interaction. Today, with the dramatic explosion of high-
performance workstations that are more and more generally available, the need for a
central source of raw computing cycles has significantly diminished. In place of
being a distributor of CPU cycles, SUMEX has become a communications cross-
roads and a source of Al and computer systems software and expertise.

SUMEX has demonstrated that a computer resource is a useful “linking mechanism"
for bringing together electronically teams of experts from different disciplines who
share a common problem focus. Al concepts and software are among the most
complex products of computer science. Historically it has not been easy for
scientists in other fields to gain access to and mastery of them. Yet the collaborative
outreach and dissemination efforts of SUMEX have been able to bridge the gap in
numerous cases. About 40 biomedical Al application projects have developed in our
national community and have been supported directly by SUMEX computing
resources over the years -- many more have benefitted indirectly through access to
the software, information, and advice offered by the SUMEX resource.

The integration of Al ideas with other parts of medical informatics and_ their
dissemination into biomedicine is happening largely because of the development in
the 1970's and early 1980's of methods and tools for the application of Al concepts

E. H. Shortliffe 6
5P41-RROO785-15 Resource Overview

to difficult professional-level problem solving. Their impact was heightened because
of the demonstration in various areas of medicine and other life sciences that these
methods and tools really work. Here SUMEX has played a key role, so much so that
it is regarded as "the home of applied Al."

SUMEX has been the home of such well-known Al systems as DENDRAL (chemical
structure elucidation), MYCIN (infectious disease diagnosis and therapy), INTERNIST
(differential diagnosis), ACT (human memory organization), MOLGEN/BIONET (tools
for DNA sequence analysis and molecular biology experiment planning), ONCOCIN
(cancer chemotherapy protocol advice), SECS (chemical synthesis), EMYCIN (rule-
based expert system tool), and AGE (blackboard-based expert system tool). In the
past four years, our community has published a dozen books that give a scholarly
perspective on the scientific experiments we have been performing. These volumes,
and other work done at SUMEX, have played a seminal role in structuring modern Al
paradigms and methodology.

The Future of SUMEX~AIM

Given this background, what is the future need and course for SUMEX as a resource
-~ especially in view of the on-going revolution in computer technology and costs
and the emergence of powerful single-user workstations and local area networking?
The answers remain clear.

Basic Research on Al! in Biomedicine

At the deepest research level, despite our considerable success in working on
medical and biological applications, the problems we can attack are still sharply
limited. Our current ideas fall short in many ways against today's important health
care and biomedical research problems brought on by the explosion in medical
knowledge and for which Al should be of assistance. Just as the research work of
the 70's and 80's in the SUMEX-AIM community fuels the current practical and
commercial applications, our work of the late 80's will be the basis for the next
decade's systems.

The report of the panel on medical informatics [3], convened late in 1985 by the
National Library of Medicine to review and recommend twenty-year goals for the
NLM, listed among its highest priority recommendations the need to greatly expand
and aggressively pursue an_ interdisciplinary research program to develop
computational methods for acquiring, representing, managing, and using biomedical
knowledge of all sorts for health care and biomedical research. These are precisely
the problems which the SUMEX-AIM community has been working on so successfully
and which will require work well beyond the five year funding period we have
requested. It is essential that this line of research in the SUMEX-AIM community,
represented by our core Al research, the ONCOCIN research, and our collaborative
research groups, be continued.

The Changing Role of the Central Resource

At the resource level, there are changing, but still growing, needs for computing
resources for the active AIM research community to continue its work over the next
five years. The workstations to which we directed our attention in 1980 have now
demonstrated their practicality as research tools and, increasingly, as potential
mechanisms for disseminating Al systems as cost-effective decision aids in clinical
settings such as private offices. The era of highly centralized general machines for
Al research is rapidly coming to an end and is being replaced by networks of
distributed but heterogeneous single-user machines sharing common information

7 E. H. Shortliffe
Resource Overview 5P41-RROO785-15

resources and communication paths among members of the biomedical research
community.

Most of our community groups have been able to take advantage of local computing
facilities, with SUMEX-AIM providing a central cross-roads for communications and
the sharing of programs and knowledge. In its core research and development role,
SUMEX-AIM has its sights set on the hardware and software systems of the next
decade. We expect major changes in the distributed computing environments that
are just now emerging in order to make effective use of their power and to adapt
them to the development and dissemination of biomedical Al systems for professional
user communities. In its training role, SUMEX is a crucial resource for the education
of badly needed new researchers and professionals to continue the development of
the biomedical Al field. The "critical mass" of the existing physical SUMEX resource,
its development staff, and its intellectual ties with the Stanford Knowledge Systems
Laboratory (KSL -- see Appendix A for a summary of current KSL research
activities), make this an ideal setting to integrate, experiment with, and export these
methodologies for the rest of the AIM community.

We will continue our experimental approach to distributed systems, learning to build
and exploit distributed networks of these machines and to build and manage graceful
software for these systems. Since decentralization is central to our future, we must
learn its technical characteristics.

Resource Sharing

An equally important function of the SUMEX-AIM resource is an exploration of the
use of computer communications as a means for interactions and sharing between
geographically remote research groups engaged in biomedical computer science
research and for the dissemination of Al technology. This facet of scientific
interaction is becoming increasingly important with the explosion of complex
information sources and the regional specialization of groups and facilities that might
be shared by remote researchers [2, 1]. Another of the key recommendations of
the NLM medical informatics planning panel [3] was that high-speed network
communication links be established throughout the biomedical research community so
that knowledge and information can be shared across diverse research groups and
that the required interdisciplinary collaborations can take place. Recent efforts to
establish a national NSFNet, largely to support the supercomputer projects funded by
NSF but also to replace and upgrade part of the national research community linkage
that the now aging ARPANET has supported, have made important progress. Still,
these efforts do not encompass the broad range of biomedical research groups that
need national network access and to date, the NIH has not played an aggressive role
in the interagency Research Internet coordination efforts. We must work to build a
stronger institutional support for a National Research Network.

SUMEX continues to be an important pathfinder to develop the technology and
community interaction tools needed to expand community system and communication
resources. Our community building effort is based upon the developing state of
distributed computing and communications technology and we have therefore turned
our core systems research to actively supporting the development of distributed
computing and communications resources to facilitate collaborative project research
and continued inter-group communications.

E. H. Shortliffe 8
5P41-RROO785-15 Resource Overview

Summary of Long-term Goals

e Maintain the synergistic relationship between SUMEX core system
development, core Al research, our experimental efforts at disseminating
clinical decision-making aids, and new applications efforts.

« Continue to serve the national AIM research community, less and less as
a source of raw computing cycles and more and more as a transfer
point for new technologies important for community research and
communication. We will also continue our coordinating role within the
community through electronic media and periodic AIM workshops.

e Maintain our connections to ARPANET, TELENET, and our local Ethernet
and assist other community members to establish similar links by
example, by integrating and providing enabling software, and by offering
advice and support within our resources.

« Focus new computing resource developments on more effective
exploitation of distributed workstations through better communication and
cooperative computing tools, using transparent digital networking
schemes.

e Enhance the computing environments of workstations so that only
minimal dependency on central, general-purpose computing hosts
remains and these mainframe time-sharing systems can be phased out
eventually. Remaining central resources wiil include servers’ for
communications, community information resources, and special computing
architectures (€.g., shared- or distributed-memory symbolic
multiprocessors) justified by cost-effectiveness and unique functionality.

e Incrementally phase in, disseminate, and evaluate those aspects of the
local distributed computing resource that are necessary for continuing
national AIM community support within this distributed paradigm. This
will ultimately point the way towards the distributed computing resource
model that we believe will interlink this community well into the next
decade.

« Gradually and responsibly phase out the existing DEC 2060 machine as
effective distributed computing alternatives become widely available. We
expect this to be possible sometime during the next year of the
continuation resource.

¢ Continue the central staff and management structure, essentially
unchanged in size and function during the five-year transition period,
except for the merging of the core part of the ONCOCIN research with
the SUMEX resource.

lll.A.1.2. Significance and Impact in Biomedicine

Artificial intelligence is the computer science of representations of symbolic
knowledge and its use in symbolic inference and problem-solving processes.
Projects in the SUMEX-AIM community are concerned in some way with the
application of Al to biomedical research and the resource has given strong impetus
and support to knowledge-based system research in biomedicine. For computer
applications in medicine and biology, this research path is crucial. Medicine and
biology are not presently mathematically-based sciences; unlike physics and

9 E. H. Shortliffe
Resource Overview 5P41-RROO785-15

engineering, they are seldom capable of exploiting the mathematical characteristics
of computation. They are essentially inferential, not calculational, sciences. If the
computer revolution is to affect biomedical scientists, computers will be used as
inferential aids.

The growth in medical knowledge has far surpassed the ability of a single
practitioner to master it all, and the computer's superior information processing
capacity thereby offers a natural appeal. Furthermore, the reasoning processes of
medical experts are poorly understood; attempts to mode! expert decision-making
necessarily require a degree of introspection and a structured experimentation that
may, in turn, improve the quality of the physician's own clinical decisions, making
them more reproducible and defensible. New insights that result may also allow us
more adequately to teach medical students and house staff the techniques for
reaching good decisions, rather than merely to offer a collection of facts which they
must independently learn to utilize coherently.

Perhaps the larger impact on medicine and biology will be the exposure and
refinement of the hitherto largely private heuristic knowledge of the experts of the
various fields studied. The ethic of science that calls for the public exposure and
criticism of knowledge has traditionally been flawed for want of a methodology to
evoke and give form to the heuristic knowledge of scientists. Al methodology is
beginning to fill that need. Heuristic knowledge can be elicited, studied, critiqued by
peers, and taught to students.

The importance of Al research and its applications is increasing in general, without
regard for the specific areas of biomedical interest. Al is one of the principal fronts
along which university computer science groups are expanding. The pressure from
student career-line choices is great. Federal and industrial support for Al research
and applications is vigorous, although support specifically for biomedical applications
continues to be limited. All of the major computer manufacturers (e.g, IBM, DEC, TI,
UNISYS, HP, and others) are using and marketing Al technology aggressively and
many software companies are putting more and more products on the market. Many
other parts of industry are also actively pursuing Al applications in their own
contexts, including defense and aerospace companies, manufacturing companies,
financial companies, and others.

Despite the limited research funding available, there is also an explosion of interest
in medical Al. The American Association for Artificial Intelligence (AAAI), the principal
scientific membership organization for the Al field, has 7000 members, several
thousand of whom are members of the medical special interest group known as the
AAAI-M. Speakers on medical Al are prominently featured at professional medical
meetings, such as the American College of Pathology and American College of
Physicians meetings; a decade ago, the words artificial intelligence were never heard
at such conferences. And at medical computing meetings, such as the annual
Symposium on Computer Applications in Medical Care (SCAMC) and the international
MEDINFO conferences, the growing interest in Al and the rapid increase in papers on
Al and expert systems are further testimony to the impact that the field is having.

Al is beginning to have a similar effect on medical education. Such diverse
organizations as the National Library of Medicine, the American College of
Physicians, the Association of American Medical Colleges, and the Medical Library
Association have all called for sweeping changes in medical education, increased
educational use of computing technology, enhanced research in medical computer
science, and career development for people working at the interface between
medicine and computing. They all cite evolving computing technology and (SUMEX-
AIM) Al research as key motivators. At Stanford, we have vigorous special programs
for student training and research in Al -- a graduate program in Medical Information

E. H. Shortliffe 10
5P41-RROO0785-15 Resource Overview

Sciences and the two-year Masters Degree in Al program. All of these have many
more applicants than available slots. Demand for their graduates, in both academic
and industrial settings, is so high that students typically begin to receive solicitations
one or two years before completing their degrees.

!.A.1.3. Summary of Current Resource Goals

The following outlines the specific objectives of the SUMEX-AIM resource during the
current three-year award period begun in August 1986. It provides an overall
research plan for the resource and provides the backdrop against which specific
progress is reported. Note that these objectives cover only the resource nucleus;
objectives for individual collaborating projects are discussed in their respective
reports in Section IV. Specific aims are broken into five categories: 1) Technological
Research and Development, 2) Collaborative Research, 3) Service and Resource
Operations, 4) Training and Education, and 5) Dissemination.

1) Technological Research and Development

SUMEX funding and computational support for core research is complementary to
similar funding from other agencies (including DARPA, NASA, NSF, NLM, private
foundations, and industry) and contributes to the long-standing interdisciplinary effort
at Stanford in basic Al research and expert system design. We expect this work to
provide the underpinnings for increasingly effective consultative programs in
medicine and for more practical adaptations of this work within emerging
microelectronic technologies. Specific aims include:

« Basic research on Al techniques applicable to biomedical problems.
Over the next term we will emphasize work on blackboard problem-
solving frameworks and architectures, knowledge acquisition or learning,
constraint satisfaction, and qualitative simulation.

« Investigate methodologies for disseminating application systems such as
Clinical decision-making advisors into user groups. This will include
generalized systems for acquiring, representing and reasoning about
complex treatment protocols such as are used in cancer chemotherapy
and which might be used for clinical trials.

e Support community efforts to organize and generalize Al tools and
architectures that have been developed in the context of individual
application projects. This will include retrospective evaluations of
systems like the AGE blackboard experiment and work on new systems
such as BB1, CARE, EONCOCIN, EOPAL, Meta~ONYX, and architectures
for concurrent symbolic computing. The objective is to evolve a body
of software tools that can be used to more efficaciously build future
knowledge-based systems and explore other biomedical Al applications.

« Develop more effective workstation systems to serve as the basis for
research, biomedical application development, and dissemination. We
seek to coordinate basic research, application work, and system
development so that the Al software we develop for the next 5-10 years
will be appropriate to the hardware and system software environments
we expect to be practical by then. Our purchases of new hardware will
be limited to experimentation with state-of-the-art workstations as they
become available for our system developments.

11 E. H. Shortliffe
Resource Overview 5P41-RROO785-15

2) Collaborative Research

e Encourage the exploration of new applications of Al to biomedical
research and improve mechanisms for inter- and_ intra-group
collaborations and communications. While Al is our defining theme, we
may consider exceptional applications justified by some other unique
feature of SUMEX-AIM essential for important biomedical research. We
will continue to exploit community expertise and sharing in software
development.

¢ Minimize administrative barriers to the community-oriented goals of
SUMEX-AIM and direct our resources toward purely scientific goals. We
will retain the current user funding arrangements for projects working on
SUMEX facilities. User projects will fund their own manpower and local
needs; actively contribute their special expertise to the SUMEX-AIM
community; and receive an allocation of system resources under the
control of the AIM management committees. We will begin charging
"fees for service" to Stanford users as DRR support for the DEC 2060
is phased out. Fees to national users will be delayed as long as
financially possible.

e Provide effective and geographically accessible communication facilities
to the SUMEX-AIM community for remote collaborations, communications
among distributed computing nodes, and experimental testing of Al
programs. We will retain the current ARPANET and TELENET
connections for at least the near term and will actively explore other
advantageous connections to new communications networks and to
dedicated links.

3) Service and Resource Operations

SUMEX-AIM does not have the computing or manpower capacity to provide routine
service to the large community of: mature projects that has developed over the
years. Rather, their computing needs are better met by the appropriate development
of their own computing resources when justified. Thus, SUMEX-AIM has the primary
focus of assisting new start-up or pilot projects in biomedical Al applications in
addition to its core research in the setting of a sizable number of collaborative
projects. We do offer continuing support for projects through the lengthy process of
obtaining funding to establish their own computing base.

4) Training and Education

« Provide documentation and assistance to interface users to resource
facilities and systems.

« Exploit particular areas of expertise within the community for assisting in
the development of pilot efforts in new application areas.

« Accept visitors in Stanford research groups within limits of manpower,
space, and computing resources.

¢ Support the Medical Information Science and MS/Al student programs at
Stanford to increase the number of research personnel available to work
on biomedical Al applications.

¢ Support workshop activities including collaboration with other community

E. H. Shortliffe 12
5P41-RROO785-15 Resource Overview

groups on the AIM community workshop and with individual projects for
more specialized workshops covering specific research, application, or
system dissemination topics.

5) Dissemination

While collaborating projects are responsible for the development and dissemination
of their own Al systems and results, the SUMEX resource will work to provide
community-wide support for dissemination efforts in areas such as:

e Encourage, contribute to, and support the on-going export of software
systems and tools within the AIM community and for commercial
development.

e Assist in the production of video tapes and films depicting aspects of
AIM community research.

e Promote the publication of books, review papers, and basic research
articles on all aspects of SUMEX-AIM research.

13 E. H. Shortliffe
Details of Technical Progress 5P41-RROO785-15

lll.A.2. Details of Technical Progress

This section gives an overview of progress for the nucleus of the SUMEX-AIM
resource. A more detailed discussion of our progress in specific areas and related
plans for further work are presented beginning on Page 19. Objectives and progress
for individual collaborating projects are discussed in their respective reports on Page
127. These collaborative projects collectively provide much of the scientific basis
for SUMEX as a resource and our role in assisting them has been a continuation of
that evolved in the past. Collaborating projects are autonomous in their management
and provide their own manpower and expertise for the development and
dissemination of their Al programs.

lll.A.2.1. Progress Highlights

In this section we sttmmarize highlights of SUMEX-AIM resource activities over the
past year (May 1987 - April 1988), focusing on the resource nucleus.

» We made excellent progress in core Al research. We have begun to
explore the design and use of very large knowledge bases with the
hypothesis that both the problems of brittleness and over-specialization
in current knowledge-based systems can be addressed by constructing
large, raulti-use knowledge bases (LMKB). A LMKB would 1) encode
domain knowledge in greater depth and breadth than required for any
specific task, 2) encode knowledge that cuts across many domains of
expertise, and 3) serve as a core repository of knowledge to be
accessed by large numbers of specific applications.

Research has also progressed on several fundamental issues of Al,
including knowledge representation, blackboard frameworks, parallel
symbolic computing architectures, and machine learning. Work continues
in PROTEAN and BB1 on the explicit representation of control
knowledge, on the representation of geometric problem-solving
knowledge in PROTEAN and PEAKS, on the representation of diagnosis
expertise and the integration of a numerical simulator of a model system
with an expert system in ABLE, and on the flexible, rich representation
of control knowledge to facilitate modeling of problem-solving at the
strategic level as well as at the tactical level. We have continued to
develop the BB1 blackboard architecture for systems that reason about
(control, explain, and learn about) their own actions. The BB1 system is
being applied in a new domain, the real time monitoring of patients in an
intensive care unit (BBICU).

The parallel architectures work has developed a CAD (Computer Aided
Design) system for hierarchical, multiple-level specification of computer
architectures (SIMPLE); a parameterized, multiprocessor array emulation
defined in SIMPLE's specification languages and running on SIMPLE's
simulator (CARE); a set of extensions to Lisp for studying expressed
concurrency in functional programming, object-oriented, and shared-
variable models of concurrent computation (LAMINA); and two alternative
parallel blackboard frameworks for expressing application problems
(POLIGON and CAGE). These have been applied to several signal
understanding problems with promising problem- solving speed-up.

The machine learning work has concentrated on explanation-based
generalization and chunking work in the SOAR framework, inductive rule

E. H. Shortliffe 14
5P41-RROO785-15 Details of Technical Progress

learning, "apprenticeship" learning, and tools for acquiring knowledge and
debugging knowledge structures. Our research in machine learning has
focused on several distinct problem domains including medicine
(NEOMYCIN/HERACLES), physics (ABLE), and biochemistry (PROTEAN)
in addition to domain-independent investigations.

Work has also continued on reasoning with uncertainty to find ways of
combining formal and informal approximate reasoning methods.
Specifically, this research seeks to (1) develop techniques for using
knowledge about problem-solving tradeoffs to dynamically optimize
system performance for the user, (2) construct efficient algorithms for
probabilistic reasoning, and (3) investigate pragmatic techniques for the
elicitation of knowledge from experts.

« We have made continued significant progress in the core ONCOCIN
research to generalize the tools for clinical trial management from the
initial cancer chemotherapy management application. A major
accomplishment this year from our work on the examination of the
structures of protocols across medical subspecialties other than cancer
chemotherapy (e.g., hypertension and insulin diabetes treatment) was the
creation of a generalized knowledge acquisition tool designed to encode
descriptions of clinical trials. The system is called PROTEGE and
produces as its output a computer program which is an OPAL-like
clinical trial protocol definition system specifically tailored for an
arbitrary clinical area such as hypertension. This OPAL-like system can
then be used to create the knowledge base (e.g., for hypertension) that
drives an ONCOCIN-like patient/protocol management system.

We continued development of the OPAL system for graphical knowledge
acquisition to facilitate protocol definition and knowledge base entry for
the ONCOCIN oncology application area. We began to explore
alternative platforms for developing OPAL-like systems such as
HyperCard on the MAC Il. We also tested the usefulness of a "generic
protocol viewing tool” based on a relational data base storage
mechanism. We performed several experiments aimed at developing a
single underlying storage mechanism (termed a "cell") from which
various interface systems including the Interviewer flowsheet, a
generalized spreadsheet utility, and the OPAL schema entry flowchart
interface could be derived.

We have begun a project to explore the integration of speech-
recognition technology into the interface to ONCOCIN. The project uses
a commercially available continuous speech recognition product and a
prototype ONCOCIN adaptation now permits users to navigate the
graphical interface and enter clinical data using spoken commands. The
development of this system exploits the significant experience we have
gained in distributed computing since the phonetic device, initial parsing
software, and the ONCOCIN system all reside on different pieces of
hardware.

We released a new version of our ONCOCIN object language which has
proven to be the most stable and powerful version to date. We have
aiso continued to examine the issues of disseminating the ONCOCIN
system into actual clinical settings.

e We have made excellent progress on the core system development work
targeted at supporting the distributed AIM community. We elaborated

15 E. H. Shortliffe
Details of Technical Progress 5P41-RROO785-15

our systems development plans further at a site visit held in August
1987 in response to our request to the National Advisory Research
Resources Council to restore the final 2 years of our grant award.
Following a special study section review and reconsideration by the
Council, our plan and the full 5-year grant award were approved. In line
with the review group recommendations, we have moved to sharply
focus our development resources on a small number of standardized
hardware and software configurations.

After consideration of many requirements and alternative systems for AIM
community computing needs for the coming years to replace and
upgrade the services of the 2060, we have chosen Apple Macintosh I
workstations as the general computing environment for researchers and
Staff, Tl Explorer Lisp machines (including the microExplorer Macintosh
coprocessor) as the near-term high-performance Lisp research
environment, and a SUN-4 as the central system network server
(wide- and local-area network interfaces, file services, printing services,
etc.). The bulk of this hardware was purchased with DARPA research
funds and we are now beginning the installation and integration process.
We have concentrated our current systems efforts on getting basic
capabilities operational, such as for text processing, filing/archiving,
printing, graphics, office management, system building tools, information
resource access, and distributed system operation and management) etc.

Development work for the Mac/Explorer/SUN environments has been
limited because of manpower cuts necessitated by NIH cuts in our
funding awards. Available resources have focussed on providing remote
access between workstations, integrating a solid support of the TCP-IP
network protocol, ard building a distributed electronic mail system. We
have significantly refined the prototype distributed EMail system
developed for the Xerox D-machines last year (a number of people are
using the system routinely to manage their mail) and are porting this to
the Explorer and adapting it for the Mac Il.

We have started a project based on the MacWorkstation software
licensed from APPLE which is designed to allow the Macintosh to start
up programs on mainframes or other workstations and receive graphic
and text output to the MAC in a seamless fashion. We plan to write a
Common Lisp window package (based on Common Windows) that uses
MacWorkstation so we can connect (via Ethernet, AppleNet or
RS-232C/modem) to any of our Common Lisp engines and run the same
piece of code on any of them.

One of the key issues in selecting the systems for our distributed
computing environment was the performance of Common Lisp and to
help make this evaluation, we undertook an informal survey of the
performance of two KSL Al software packages, SOAR and BB1, on a
wide variety of machines. Within a factor of two of the best
performance, a considerable range of workstations based on stock
microprocessor chips as well as specially microprogrammed Lisp chips
have comparable performance. Even though performance gaps between
microprogrammed Lisp systems and stock workstation implementations
are narrowing, there still remains a significant difference in the quality of
the development environments. We have attempted to distill the key
features of the Lisp machine environments that would be needed in
stock machine implementations in order to make them attractive in a
development setting.

E. H. Shortliffe 16
5P41-RROO785-15 Details of Technicai Progress

We have installed an interface between our Ethernet environment and
the TELENET network by which many AIM users gain access to SUMEX
that allows full access to distributed SUMEX resources. We have also
continued to maintain the DEC 2060 environment to stay current with
operating system and other environmental upgrades.

e We have continued the dissemination of SUMEX-AIM technology through
various media. We have reorganized the distribution system for our Al
software tools (EMYCIN, AGE, and BBt) to academic, industrial, and
federal research laboratories, in order to make it more efficient and
require less research staff time. We have also continued to distribute
the video tapes of some of our research projects including ONCOCIN,
and an overview tape of Knowledge Systems Laboratory work to outside
groups. Our group has continued to publish actively on the results of
our research, including more than 45 research papers per year in the Al
literature and a dozen books in the past 5 years on various aspects of
SUMEX-AIM Al research. We assisted and participated actively in the
AIM Workshop sponsored by AAA! and held at Stanford (Ramesh Patil
from MIT was the Program Chairman) and hosted a number of AIM
community visitors at our Stanford research laboratory this past year.

« The Medical Information Sciences program, begun at Stanford in 1983
under Professor Shortliffe as Director, has continued its strong
development over the past year.. The specialized curriculum offered by
the MIS program focuses on the development of a new generation of
researchers able to support the development of improved computer-
based solutions to biomedical needs. The feasibility of this program
resulted in large part from the prior work and research computing
environment provided by the SUMEX-AIM resource. It has recently
received enthusiastic endorsement from the Stanford Faculty Senate for
an additional five years, has been awarded renewed post-doctoral
training support from the National Library of Medicine with high praise for
the training and contributions of the SUMEX-AIM environment from the
reviewing study section, and has received additional industrial and
foundation grants for student support. This past year, MIS students have
published many papers, including several that have won conference
awards.

e We have continued to recruit new user projects and collaborators to
explore further biomedical areas for applying Al. A number of these
projects are built around the communications network facilities we have
assembled, bringing together medical and computer science collaborators
from remote institutions and making their research programs available to
still other remote users. At the same time we have encouraged older
mature projects to build their own computing environments thereby
facilitating the transition to a distributed AIM community. A substantial
number of projects have moved to their own computing resources,
including SOAR, under Dr. Paul Rosenbloom at USC/ISI in Los Angeles;
the Logic Group projects (DART, Intelligent Agents, and MRS) under
Professor Michae! Genesereth at Stanford: Hierarchical Models of Human
Cognition (CLIPR), under Professors Walter Kintsch and Peter Polson at
the University of Colorado; Problem Solving Expertise (SOLVER), under
Professors Paul Johnson and William Thompson at the University of
Minnesota; RXDX, under Professor Robert Lindsay at the University of
Michigan; and Computer-Based Exercises in Pathophysiologic Diagnosis,
under Professor J. Robert Beck at Dartmouth College.

17 E. H. Shortliffe
Details of Technical Progress 5P41-RROO785-15

SUMEX user projects have made good progress in developing and
disseminating effective consultative computer programs for biomedical
research. These systems provide expertise in areas like cancer
chemotherapy protocol management, clinical diagnosis and decision-
making, and molecular biology. We have worked hard to meet their
needs and are grateful for their expressed appreciation (see Section IV).

E. H. Shortliffe 18
5P41-RROO785-15 Details of Technical Progress

ll.A.2.2. Core ONCOCIN Research

ONCOCIN is a data management and therapy advising program for complex cancer
chemotherapy experiments. The development of the system began in 1979, following
the successful generalization of MYCIN into the EMYCIN expert system shell. The
ONCOCIN project has evolved over the last eight years: the original version of
ONCOCIN ran.on the time-shared DEC computers using a standard terminal for the
time-oriented display of patient data. The current version uses compact
workstations running on the Ethernet network with a large bit-mapped displays for
presentation of patient data) The project has also expanded in scope. There are
three major research components: 1) ONCOCIN, the therapy planning program and its
graphical interface; 2) OPAL, a graphical knowledge entry system for ONCOCIN; and
3) ONYX, a strategic planning program designed to give advice in complex therapy
situations. Each of these research components has been split into two parts:
continued development of the cancer therapy versions of the system, and
generalization of each of the components for use in other areas of medicine. This
annual report will describe our work on each of the components: implementation of
ONCOCIN workstations in the Stanford clinic, knowledge acquisition research, and
research to generalize ONCOCIN for application in clinical trial domains other than
medical oncology (E-ONCOCIN).

A major highlight of this year was the creation of a generalized knowledge
acquisition tool designed to encode descriptions of clinical trials. The system,
named PROTEGE, was the Ph.D. thesis work of Mark Musen. The output of PROTEGE
is an OPAL-like input system designed for one clinical area such as hypertension.
This input system (HTN-OPAL) can then be used to create the hypertension
knowledge base for an E-ONCOCIN like system. This experiment was carried out
this year for both the hypertension and oncology domains. Details of this project are
described later in this report.

1 - Overview of the ONCOCIN Therapy Planning System

ONCOCIN is an advanced expert system for clinical oncology. It is designed for use
after a diagnosis has been reached, focusing instead on assisting with the
management of cancer patients who are receiving chemotherapy. Because
anticancer agents tend to be highly toxic, and because their tumor-killing effects are
routinely accompanied by damage to normal cells, the rules for monitoring and
adjusting treatment in. response to a given patient's course over time tend to be
complex and difficult to memorize. ONCOCIN integrates a temporal record of a
patient's ongoing treatment with an underlying knowledge base of treatment protocols
and rules for adjusting dosage, delaying treatment, aborting cycles, ordering special
tests, and similar management details. The program uses such knowledge to help
physicians with decisions regarding the management of specific patients.

A major lesson of past work in clinical computing has been the need to develop
methods for integrating a system smoothly into the patient-care environment for
which it is intended. In the case of ONCOCIN, the goal has been to provide expert
consultative advice as a by-product of the patient data management process, thereby
avoiding the need for physicians to go out of their way to obtain advice. it is
intended that oncologists use ONCOCIN routinely for recording and reviewing patient
data on the computer's screen, regardless of whether they feel they need decision-
making assistance. This process replaces the conventional recording of data on a
paper flowsheet and thus seeks to avoid being perceived as an additive task. In
accordance with its knowledge of the patient's chemotherapy protocol, ONCOCIN
then provides assistance by suggesting appropriate therapy at the time that the day's
treatment is to be recorded on the flowsheet. Physicians maintain control of the

19 E. H. Shortliffe
Details of Technical Progress 5P41-RROO785-15

decision, however, and can override the computer's recommendation if they wish.
ONCOCIN also indicates the appropriate interval until the patient's next treatment and
reminds the physician of radiologic and laboratory studies required by the treatment
protocol.

2 - Implementation of the ONCOCIN Workstation in the Stanford Clinic

In mid-1986, we placed the workstation version of ONCOCIN into the Oncology Day
Care clinic. This version is a completely different program from the version of
ONCOCIN that was available in the clinic from 1981-1985-- using protocols entered
through the OPAL program, with a new graphical data entry interface, and revised
knowledge representation and reasoning component. One person in the clinic (Andy
Zelenetz) became primarily responsible for making sure that our design goals for this
version of ONCOCIN were met. His suggestions included the addition of key
protocols and the ability to have the program be useful for clinicians as a data
management tool if the complete treatment protoco! had not yet been entered into
the system. Additional fellows are being trained on a very stable release of
ONCOCIN that became available in early 1988. A version of the system was sent to
the University of Pittsburgh for evaluation. In addition, ONCOCIN will become
available for presentation at the National Library of Medicine Artificial Intelligence
Demonstration Center. For these various efforts, Janice Rohn has created an
extensive user manual, sample patient interactions, and reminder cards to shorten the
training period for ONCOCIN.

The process of entering a large number of treatment protocols in a short period of
time led to other research topics including: design of an automated system for
producing meaningful test cases for each knowledge, modification of the design of
the time-oriented database and the methods for accessing the database, and the
development of metheds for graphically viewing multiple protocols that are combined
into one large knowledge base. These research efforts will continue into the next
year. In addition, some of the treatment regimens developed for the originai
mainframe version are still in use and can be transferred to the new version of
ONCOCIN.

We also received new insights about the design of the internal structures of the
knowledge base (e.g., the relationship between the way we refer to chemotherapies,
drugs, and treatment visits). We will continue to optimize the question-asking
procedure, the method for traversing the plan structure in the knowledge base, and
consider alternative arrangements used to represent the structure of chemotherapy
plans. Although we have concentrated our review of the ONCOCIN design primarily
on the data provided by additional protocols, we know that non-cancer therapy
problems raise similar issues. The E-ONCOCIN effort is designed to produce a
domain-independent therapy planning system that includes the lessons learned from
our oncology research.

3 - E-ONCOCIN: Domain Independent Therapy Planning

During the past two years, our E-ONCOCIN research has concentrated on
understanding how protocols in medicine vary across subspecialties. We are
examining several application areas: the intensive care unit, insulin treatment for
diabetes, hypertension protocols, and both standard and complex cancer treatment
problems. The diagnosis and therapy selection for patients in the intensive care unit
is a natural application area because it is based on changing data and the need to
determine the response to therapy interventions. In addition, it is an area where
reasonable mathematical models of the respiratory system can be integrated into the
expert system. We also felt that the area of insulin treatment for diabetes would be a

E. H. Shortliffe 20
5P41-RROO785-15 Details of Technical Progress

good area to explore. Like cancer chemotherapy, the treatments for diabetes
continues over a long period of time and has been the area of intensive protocol
development. Unlike cancer chemotherapy, the treatment plan must handle multiple
treatments in one day and deemphasizes the use of multiple drugs (although there
are a variety of types of insulin). During 1987, using the medical literature and
several internists in the medical computer science research group (Mark Frisse, Mark
Musen, and Michael Kahn), we performed knowledge acquisition experiments for
insulin treatment of diabetes. The proposed structure for the knowledge base was
implemented using the object-oriented programming language upon which ONCOCIN
has been based. These experiments, like those of adding more protocols to
ONCOCIN, demonstrated the need for changes in the way that the knowledge base
can access the time-oriented data base that records patient data and previous
conclusions. The relationships between the different doses and types of insulin
treatments will also require alternative ways of building treatment hierarchies. Thus,
our initial experiments have shown that many of the elements of the ONCOCIN
design are sufficiently general for other application areas, but that some specific
elements (particularly the representation of temporal events) will have to be
generalized. A description of our revised temporal representations will appear in a
forthcoming thesis by Michae! Kahn of University of California at San Francisco.
During the coming year, we will continue our knowledge acquisition experiments and
design a version of the E-ONCOCIN system that is separate from the ongoing "clinic
version."

4 - OPAL: Graphical Knowledge Acquisition Interface

OPAL is a graphical environment for use by an oncologist who wishes to enter a
new chemotherapy protocol for use by ONCOCIN or to edit an existing protocol.
Although the system is designed for use by oncologists who have been trained in its
use, it does not require an understanding of the internal representations or reasoning
strategies used by ONCOCIN. The system may be used in two interactive modes,
depending on the type of knowledge to be entered. The first permits the entry of a
graphical description of the overall flow of the therapy process. The oncologist
manipulates boxes on the screen that stand for various steps in the protocol. The
resulting diagram is then translated by OPAL into computer code for use by
ONCOCIN. Thus, by drawing a flow chart that describes the protocol schematically,
the physician is effectively programming the computer to carry out the procedure
appropriately when ONCOCIN is later used to guide the management of a patient
enrolled in that protocol.

OPAL's second interactive mode permits the oncologist to describe the details of the
individual events specified in the graphical description. For example, the rules for
administering a given chemotherapy will vary greatly depending upon the patient's
response to earlier doses, intercurrent illnesses and toxicities, hematologic status,
etc. For example, one form permits the entry of an attenuation schedule for an
agent based upon the patient's white count and platelet count at the time of
treatment. Tables such as this are generally found in the written version of
chemotherapy protocols. Thus OPAL permits oncologists to enter information using
familiar forms displayed on the computer's screen. The contents of such forms are
subsequently translated into rules and other knowledge structures for use by
ONCOCIN.

21 E. H. Shortliffe
Details of Technical Progress 5P41-RRO0785-15

4.1 - Status of the OPAL System

The OPAL is one of the few graphical knowledge acquisition systems ever designed
for expert systems. Even fewer are designed to be used as the main method for
entering knowledge as opposed to a proof of concept implementation. We have
pursued three directions in the development of the OPAL system, also in response
to the large number of protocols entered through this system during the last year.
The first direction is the modification of graphical forms needed to allow the entry of
facts that did not show up in the protocols used to test the initial version of OPAL.
OPAL continues to assume that most of the knowledge to be entered will have very
stereotyped forms, e.g., dose attenuations for most treatment toxicities are based on
a comparison of only one laboratory measurement at a time, such as using the BUN
to adjust for renal toxicity. We sometimes need much more complex ways of stating
the scenarios in which dose adjustments may be necessary. This need has led us
in a second direction, towards a "lower-level" rule entry approaching the syntax of
the reasoning component of ONCOCIN, but using graphical input devices where
applicable. A major accomplishment of this last year was to experimentally combine
the OPAL and ONCOCIN programs into one working program, and to completely
enter knowledge from OPAL using both the high level tools and lower level rule
editors, but without needing to make changes at the ONCOCIN side of the system.
Work on graphical replacements for low level rule entry is the master's project of
Eric Sherman.

The OPAL program maps the information provided on the graphical forms into a
complex data structure (called the IDS) that represents the required knowledge to
snecify the contents of a protocol. This data structure is used for copying
information from one protocol to another, and as the basis for the creation of the
ONCOCIN knowledge base. Our experiments with OPAL, and our intention to
generalize OPAL use outside of oncology protocols, suggests that we reorganize the
OPAL program to use a relational database to store its knowledge. We have
patterned the database after an existing database query syntax. Because no
databases exist for the InterLisp language upon which OPAL is based, we
reimplemented the database from its written description. The database structure is
completed, and was the basis for the PROTEGE knowledge acquisition experiments.

With changes in the future of dedicated lisp processors, we began to explore
alternative platforms for developing OPAL-like systems. We have begun experiments
using HyperCard on the MAC Il. We also tested the usefulness of a “generic
protocol viewing tool" based on a relational-database storage mechanism, in
preparation for a OPAL design based on relational DBMS technology. We also
performed several experiments aimed at developing a single underlying storage
mechanism (termed a "“cell") from which various interface systems including the
Interviewer flowsheet, a generalized spreadsheet utility, and the OPAL schema entry
flowchart interface could be derived.

5 - Generalized Knowledge Acquisition through PROTEGE

Mark Musen designed and implemented the first version of the PROTEGE
knowledge-acquisition-system development tool. PROTEGE is used to collect
information which describes the concepts (both entities and their relationships) in an
application area for which a skeletal-planning type of expert system would be useful,
concentrating on clinical trials. The system acquires the "ontology" of a domain
through a series of fill-in-the-blank forms and a "flowchart-entry" tool. These
concepts are then mapped onto a set of generic forms, which, in turn , create a
knowledge acquisition tool for the application area.

PROTEGE makes use of the forms management system built for the original OPAL,

E. H. Shortliffe 22
5P41-RROO785-15 Details of Technical Progress

and a newly developed relational database management system written for the Xerox
InterLisp-D workstations. The output of PROTEGE is an OPAL-like set of forms
tailored to the special structures of the application area. To test these ideas, we
first reimplemented portions of the OPAL interface from a high level description of
oncology. After the translation process to the ONCOCIN reasoning program, a
consultation was run that matched the manually built system. This experiment was
then repeated for the area of hypertension protocols for which ONCOGCIN had never
been specifically designed. With some minor generalizations to the oncocin
reasoner and interviewer, we were able to run a hypertension consultation.

6 - Speech Input to Expert Systems

6.1 - Prototype Speech Hardware/Software System

In. 1987 we began a project to explore the integration of speech-recognition
technology into the interface to ONCOCIN running on the XEROX lisp workstations.
The project uses a commercially available continuous-speech-recognition product
loaned by the vendor, Speech Systems, Inc. (SSI) of Tarzana, California. The
speech recognizer consists of a custom processor, called the Phonetic Engine(R)
and a suite of software modules called the Phonetic Decoder(TM). The Phonetic
Decoder(TM) is running on a SUN 3/75.

The development of this project requires significant experience in distributed
computing since the phonetic device, initial parsing software and the ONCOCIN
system all reside on different pieces of hardware. One of the early steps is to allow
the Lisp machine to remotely control the parsing software on the SUN. We built an
interpreter for communicating between the speech software library running on the
SUN and the Xerox Lisp machine. This interpreter reads lisp-style function calls
corresponding to the speech library routines and returns lisp-style results as remote
procedure call (RPC) mechanism. We then wrote the library of corresponding Lisp
stub routines and a function to connect to the SUN workstation (through a
programmatic TELNET) and start the server. In normal operation, we call the 'C'-
based speech library routines from Lisp as if they were Lisp functions. We did a
number of updates to both components of the server program (on the SUN and Lisp
machine) as various SSI updates came out, including a major revision when SSI
doubled the number of library routines, greatly increasing the server's capabilities.

Eventually we plan to replace the programmatic TELNET and custom RPC language
with a new version of the system that uses the SUN RPC language. This will include
the Xerox Lisp interface to the SUN RPC package developed by SUMEX; we were
not able to use this interface initially as it did not exist early enough and required a
release of Xerox Lisp that supports Common Lisp which the ONCOCIN system was
not yet using. This eventual change will allow more efficient communications
between the machines, allowing us to move larger data structures much faster. This
will also tie in to the proposed method of communications between Lisp and non-
Lisp routines that Xerox plans to use when their system migrates to the SUN
workstations. The Lisp routines that correspond to the various SSI library routines
will not appear to change to the application programs so they should not require any
modification when the underlying RPC mechanism is changed.

We created a prototype system that permits users to navigate the graphical interface
and enter clinical data using speech. The system uses the location of the cursor on
the screen to provide a context for choosing candidate grammars with which to
attempt recognition of a user's utterance. The system dynamically re-orders the list
of candidate recognition grammars based on the dialog history. Albeit with limitations
on the legal grammars, it is now possible to carry on most of the ONCOCIN data

23 E. H. Shortliffe
Details of Technical Progress 5P41-RROO785-15

acquisition steps using speech alone or speech plus pointing with the mouse. In
addition, some elements such as the neural toxicities can be entered as textual
descriptions and automatically encoded as one the 1-4 point scale used on
flowsheet forms.

In order to translate an utterance back into an action that can occur in the ONCOCIN
interface, we need the ability to reparse the text string returned by the SSI
equipment. The SSI equipment uses (potentially complex) syntax, built up of various
classifications, to understand sentences but returns just the ASCII component of the
actual sentence; you can not get it in terms of the original classifications in the
syntax (which are generally semantically significant). To overcome this problem, we
devised a whole new syntax format based on Lisp and our OZONE object language
in which one devises a grammar from which the SS! syntax files are generated.
Then when the ASCII string is returned, the original syntax object is used to parse
the string into a parse tree that relates directly to the grammar definition. We can
now process the returned information at a much higher level than was possible with
the simple ASCII text. In addition, the semantic elements of the parse can be added
to syntactic structures used to encode the sentence.

The latest additions to this part of the system include exhaustive and random
sentence generation as Lisp data structures (previously only partially available just
on the SUN workstation) as well as word list generation from the syntax objects.
These features generate both in-core representations as well as file representations
which will potentially free us from any dependency on SSI's equipment (versus
another manufacture's speech input device).

We wrote graphics software that plots the sentence returned by the speech
equipment along with the changes in amplitude, pitch, accuracy scores and other
information. This helps us to understand how SSI's ‘black box’ operates, particularly
in situations where it fails. In addition, we wrote programs to make it possible to
return all the best syntactic matches from the speech equipment. The current SS!
software can either return the one best answer or all the possible word mappings
(unfortunately regardless of syntax). We believe the latter would be useful when the
most likely parse is associated with a low reliability score. We could then look at all
the highest scoring candidates and evaluate what they do and do not have in
common for creating a clarifying query to the speaker. The candidate sentences are
composed from the word mapping data which is immense and needs to be
compressed, filtered and manipulated to be of any use (the parser/generator is used
as a filter to remove the non-syntactic word mappings). Unfortunately, though it runs
faster with each improvement, this doesn't look reasonable for real time use;
currently a three word utterance takes about two to four minutes to process into a
candidate list of half a dozen possible sentences. It gives us an indication of what
would be useful data structures to process on the SUN (to improve speed and lower
data transmission).

6.2 - Speech Experiments

We are performing experiments to (1) enhance the system's grammars with a wider
range of phrases clinicians actually use when talking to a computer and (2) gain
insights into clinicians' models of spoken interaction with advice systems so that we
may ground our interface design in observed practice.

In order to assess how physicians would speak to a computer in an ideal situation
without constraints or prior assumptions, we are conducting a series of experiments
which simulate continuous-speech understanding by computers. The setting of these
experiments includes a hidden computer operator simulating the output of ONCOCIN
if it had the ability to understand the spoken input, as well as a video camera to

E. H. Shortliffe 24
5P41-RROO785-15 Details of Technical Progress

record both audio and visual clues. Typed responses from the operator are
translated back as actions on the computer display as well as audio responses
through the use of a speech synthesizer. It appears to the subject as if the
computer is understanding and responding to their speech. The physicians use
ONCOCIN in the same manner as it is used in the clinic when they see patients, but
with the added capability of speech input. These experiments enable us to both build
up a basic vocabulary for the speech system as well as examine subtle linguistic
issues to guide future directions.

One component of these experiments was the use of speech synthesis as an output
medium. Using an inexpensive board based on the General Instrument ASClI-to-
phoneme and phoneme-to-voice chip set, we built a driver for the Xerox Lisp
machines. This driver included an application software transparent misspelling
dictionary facility to correct for inaccuracies in the speech board.

7 - Object Language Support for ONCOCIN Project

We released a new version of our object language at the start of this past year
which has proven to be the most stable and powerful version to date. There have
been a number of minor bug fixes and several feature additions over the course of
the year but for the most part the system as required much less attention than in
previous years. The number of new systems being built on it (like our speech work)
continues to increase. Future planning for the system consists of determining
whether or not it should be converted to Common Lisp, based on whether object
systems available under Common Lisp are sufficient for our needs, and if we do
convert it what it would it look like if properly integrated with that language.

8 - Personnel

Samson Tu has been primarily responsible for the design of E-ONCOCIN, Michael
Kahn has developed the temporal representations used by the system, Clifford
Wulfman has been involved with extensions the the data entry interface and the
extensions to the interface in order to add speech input. Samson and Cliff were
responsible for extensions to their programs to support the PROTEGE effort. David
Combs has been involved with the knowledge acquisition interface and provided
major programming support for the PROTEGE effort. Janice Rohn has been involved
with the entry of protocols, interaction with physicians using the system,
documentation of the system, and execution of the speech experiments. Ellen
Isaacs, a Ph.D. student in Psycholinguistics has helped to design the speech
experiments. Christopher Lane has developed the object-oriented systems software
upon which the entire ONCOCIN system is designed and the systems software and
parsing programs used in the speech project.

25 E. H. Shortliffe