HYOROID PROJECT Section 4.2.4

A number of large computational applications are being analyzed in order to
assess their potential for decomposition into modules for distributed processing.
The current applications are:

a) Programs which use heuristic methods in decision-making. Heuristic programs
frequently employ recursive decomposition of problems into subsidiary
problems which themselves may be suitable for distributed processing.

b} Programs which use multi-faceted databases to retrieve and abstract
information. The process of intelligent data retrieval and analysis often
depends on data or knowledge sources which are being maintained at
geographically distributed processing sites.

c) Programs which acquire data from multiple, possibly dissimilar, sensors and
attempt to reduce this data to simpler hypotheses.

Further candidates for analysis are:

d) Programs which solve large numerical problems, such as those found in image
processing applications.

e) Operating system functions.

How the parts of those programs work together depends on the coupling required
between their processes and on the facilities to support this coupling by the
hardware. Parameters which describe the computations to be simulated include:

a) The computational kernel size: the cycle and memory demand of a computational
unit between interprocessor reference requirements.

b)} The computation definition message size: the amount of data required to
transmit sufficient information to sufficient information to initiate a
computational kernel.

c) The database size: the amount of data or program text required to sustain a
computational kernel, and its availability and residence in the network.

The behaviour of the system can be varied through the adjustment of other
parameters. These parameters may be set to reflect the architecture of specific
hardware systems, or may be varied to obtain optimum performance. In addition to
obvious parameters (as the number and power of the processors), we expect the
following parameter types to be important in developing an understanding of the
spectrum of distributed processor architectures:

a) Interconnection density. As the density decreases, the message delay and
congestion increase. This parameter will provide a high level abstraction of
multi-processor connectivity schemes. Gecgraphical distribution will
increase message delay and transmission const.

b) Computational Jocality. A high degree of locality (of database or procedural
information in the network) will enhance the probability that relevant
knowledge exists in closely linked nodes, thus counteracting the effects of a
lon interconnection density.

151 J. Lederberg & E. Feigenbaum
Section 4.2.4 HYDROID PROJECT

c) Database viscosity. A database, including the programs required to carry out
the computations at a node, may be more or less fixed to one specific node.
This therefore encourages the use of certain nodes for specific functions.
Many current processor networks are completely rigid in this sense, and for
these networks optimal initial program and database allocations may be
determined. However, we hypothesize that a greater degree of dynamic
resource allocation is desirable to cope with changing loads and in order to
enhance reliability. For this reason this parameter needs to be included.

d) Redundancy. In order to assess the cost and benefits in terms of
responsiveness and reliability, the redundancy of database and computations
Will also be made a parameter. In order to utilize the redundancy well, the
computational resources (programs or data) which effect system performance
most must be identifiable.

e) Error rate. In order to test the effectiveness of reliability strategies,
node and communications channel failures will be simulated.

An important aspect of this model is that we intend to keep the
abstractions at a sufficiently high level to allow analytic and intuitive
verification of the model behaviour when applied to well understood computations.
In the past several large applications have been mapped into specific parallel
machines, but these results are not easily transferred to new architectures. The
distributed processor systems now being built may have characteristics with
unpredicted effects on system behaviour. We expect to be able to use the model
to find potential bottlenecks, which then will define areas where extra design
attention has a high payoff.

We do not intend to build hardware which is based literally on the abstract
model. We hope to verify results obtained from the model using existing
distributed processor systems and, assuming that our model (with appropriate
parameters describing the load and architecture) matches the given system, be
able to advise on system utilization or development aspects.

We are currently active in the support of one specific architecture, the S-
1 Advanced Technology Processor. This system promises to deliver very high
performance through the use of quite powerful processors nodes (on the level of
the DEC KL-10 or more), large shared address spaces (30 bit addresses), and a
moderate number of paraliel processors (16 for the Mark II version). The fact
that memory is shared allows the coupling to be quite strong, but partitioning of
problems with Jess coupling will yet incur less overhead and interference.

The hardware for the prototype processor (Mark I) node is now beginning to
be operational, basic system software has been written. The PASCAL language is
now available, a FORTRAN compiler is being developed, and a LISP compiler is
being considered as the next task. This project is funded through Lawrence
Livermore Laboratory and supported by various agencies. While this
implementation is only one of the possible architectural choices for multi-
processors it is encouraging that we are obtaining realistic resources to verify
and test notions of parallel processing.

J. Lederberg & £. Feigenbaum 152
HYDROID PROJECT Section 4.2.4

B. Medical Relevance and Collaboration

Many applications at SUMEX consume large quantities of computational
resources. They are also being hampered by architectural limitations of the
current hardware, of which the 18 bit address is probably the worst constraint.
The use of multiple distributed processors may provide a means to gain the
required processing capabilities in an economic manner. In this sense the
medical relevance of this study is indirect. We are attempting to develop tools
which will be of use in medical computation problems.

Our studies in distributed database applications have a more direct medical
relevance. To this end, we are maintaining contact with Dr. Jim Fries, whose
ARAMIS database network collects data for the analysis of disease progress and
treatment efficacy in rheumatoid arthritis from a variety of institutions.
Sharing of data to provide a broader base for analysis is also a feature of
programs in cardiology and oncology in which physicians at Stanford participate.
We are also discussing with Or. Collen and his collaborators at Kaiser Permanente
the benefits, costs, and trade-offs of distributed computing. The development of
adequate concepts and models is important to lay a bridge which supports
communications between Medical and Computer Scientists.

In each of these instances the distributed nature of the data resources
leads to differences in the meaning of data items, so that simple aggregation of
the data may not be valid. Distributed precessing may provide a powerful
alternative. We see here flow of partially processed data, i.e., information,
between the nodes, rather than row data of unspecified semantic value.

Cc. Progress Summary

The HYDROID project has been underway since the fall of 1976. We have been
involved since that time in developing a basic understanding of important problem
areas in distributed processing and problem solving.

A weekly research seminar, begun in December 1976 has brought together
members of the faculty and students from a variety of disciplines, and has
included several speakers from application areas where distributed processing may
be beneficial. Recently this seminar has been combined nith a research seminar
on communication issues, and this quarter a joint course on the topic of
distributed computing is being presented for the first time and has attracted
about 150 student from Stanford and industry. Development of industrial
participation is deemed important since the resources for these systems require
joint efforts.

A formalism to express the control of distributed problem solving in
loosely-coupled processor networks is the subject of the nearly completed Ph.D.
thesis of Reid Smith. This CONTRACT NET protocol makes the interprocessor
interactions explicit. It is the cost associated with these interactions which
appears to generate one of the performance boundaries for distributed processor
systems.

Problems of distributed search, distributed data acquisition, and
distributed databases have been simulated, as well as a model which studied
network topologies.

153 J. Lederberg & E. Feigenbaum
Section 4.2.4 HYORQID PROJECT

P)

2)

3)

4)

5)

6)

7)

8)

9)

10)

11)

12)

13)

Publications

H. Garcia-Molina, "Overview and Bibliography of Distributed Databases",
Computer Science Department, Heuristic Programming Project, Stanford
University.

H. Garcia-Molina, "Distributed Database Coupling", Heuristic Programming
Project, Paper HPP-78-4, Computer Science Department, March 1978.

H. Garcia-Molina, "Performance Comparison of Update Algorithms for
Distributed Databases",Heuristic Programming Project, Paper HPP-78-6, March
1978. .

K. Knudsen, "Some Issues in the Design of Large Mul ti-Microprocessor
Networks", Stanford Heuristic Programming Project, Paper HPP-77-31.

R. G Smith, "A Formalism for Distributed Problem Solving", Proceedings of the
Fifth International Joint Conference on Artificial Intelligence, MIT AI Lab,
Cambridge, August 1977.

R. G. Smith, “Issues in Distributed Sensor Net Design", Heuristic Programming
Project, HPP-78-2, January 1978.

R. G. Smith and R. Davis, "Distributed Problem Solving: The Contract Net
Approach", to be presented at the 2nd National CSCSI Conference, Toronto,
July 1978.

Thomas McWilliams, Lawrence C. Widdoes, Jr., and Lowell L. Wood, "Advanced
Digital Processor Technology Base Development for Navy Applications; The S~-1
Project",Lawrence Livermore Laboratory, Report NCI0-17705, September 1977.
G. Wiederhold and I. Kuhn, "Effective Services of Automated Ambulatory
Medical Record Systems", Policy Analysis and Information Systems, Vol. 1 2,
January 1978, pp. 21-32.

G. Wiederhold, "Binding in Information Processing", in preparation.

G. Wiederhold and R. E1 Masri, "A Structural Model for Database Systems",
submitted for publication.

G. Wiederhold, "Introducing Semantic Information into The Database Schema",
Proceedings of the Annual Conference of the Canadian Information Processing
Society, May 1978.

G. Wiederhold, "The Data Model of MUMPS Globals", Proceedings of the MUMPS
Users Group Meeting, June 1978.

Funding Support Status

The HYDROID project is currently supported under the Heuristic Programming

Project contract with the Advance Research Projects Agency of the DOD, contract
number MBA 903-77-C-0322, E. A. Feigenbaum, Principle Investigator.

J.

Lederberg € E. Feigenbaum 154
HYDROID PROJECT Section 4.2.4

IT. INTERACTIONS WITH SUMEX-AIM

SUMEX-AIM currently provides mainly the resources for communication and
documentation for the project. Recently, simulations, although initially
developed at SUMEX, have been carried out on IBM 370 computers at SCIP and SLAC
at Stanford due to their very high demands on computing cycles. System work for
the S-1 project is carried out at the SLAC and at the SAIL facilities. Further
funded work on databases will use resources at SRI. The availability of the
ARPA-net has made it possible to shift work to processors where resources and
funding are most appropriate. We thus enjoy a high degree of interaction with
other projects involved in the problems which result from construction of large
programs. Other points of contact are related to the use of the same programming
languages as well as the abundance of AI expertise residing around the resource.
This latter point is especially important considering that one of our aims is
discovery of suitable mappings of well understood AI methods onto highly parallel
asynchronous processor networks.

SUMEX-AIM is also an excellent medium for informal transmission of reports,
recent results and bulletins to users with related interests and problems. The
powerful screen-oriented editors available greatly enhance our capabilities for
writing both text and programs.

Finally, the development of simulation programs generally requires a highly

interactive computing environment - the sort of environment that SUMEX-AIM
provides.

155 J. Lederberg & E. Feigenbaum
Section 4.2.5 MOLGEN PROJECT

4.2.5 MOLGEN PROJECT

MOLGEN - An Experiment Planning System for Molecular Genetics

Prof. J. Lederberg (Genetics, Stanford)
Prof. N. Martin (Computer Science, U. of New Mexico)
Prof. E. Feigenbaum (Computer Science, Stanford)

I. SUMMARY OF RESEARCH PROGRAN
A. Technical Goals

The MOLGEN project is developing a computer system capable of generating
the experiment-planning sequences needed to solve given structural problems in
molecular genetics. In particular we have developed a system which is capable of
acquiring and representing information about genetic objects, transformations,
and strategies. The knowledge base presently includes information on DNA
structures, restriction enzymes, laboratory techniques, and a growing collection
of genetic strategies for discovering information about various aspects of DNA
molecules. Several specific subproblems such as simulating Ligase enzymes,
determining safe restriction enzymes for gene excision, and inferring DNA
structures from segmentation data have been explored. We have designed our
effort to facilitate generalization to other domains beyond genetics in future
research and applications.

The MOLGEN project has both an applications and a computer science
dimension. Along the tatter dimension, we seek to deepen our knowledge of the
art and science of creating programs that reason with symbolic knowledge to aid
human problem solvers. The task domain, molecular genetics, serves as a rich
intellectual and scientific environment in which to develop and test our ideas.

The major computer science issues we are addressing during the current
grant period are:

(1) Creation of a knowledge representation system with a knowledge
acquisition package. The system, known as the Units Package, may be used
te build a knowledge base in any suitable domatn. It provides an object-
centered approach for storage of both declarative and procedural
information concerning all entities in the domain.

(2) Structured representation of process information. Procedures which
simulate the action of the various processes in the domain form an
integral part of the knowledge base. Moreover, the representation
framework allons for inspection and acquisition of those procedures.

(3) Creation of program schemata and instances for general problem solving
steps. Domain-independent knowledge about general problem solving
methods also fits into the knowledge representation structure we have

devised.

J. Lederberg & E. Feigenbaum 156
MOLGEN PROJECT | Section 4.2.5

C4) Domain Specific Critics. Mechanisms for the activation of vartous domain
specific strategies when certain predefined situations occur during the
course of experiment design.

(5) Development of a specific planning strategy designed to provide high-
performance for the class of genetic experiments known as discrimination
experiments. The idea is based on indexing abstracted experimental
designs to the types of structural features for which they have proven
useful.

Along the applications dimension, we are attempting to develop tools that
can benefit molecular geneticists. We believe there is substantial benefit to be
derived from programs that act as "intelligent assistants" to scientists. First
of all, the sheer amount of detailed knowledge a scientist is expected to know
makes it likely that good experiments are being missed. Second, we believe that
an intelligent planning assistant can offer some help in reasoning about the
consequences of combining experimental facts in many possible ways.

A third motivation for applying artificial intelligence techniques to an
experimental science like molecular genetics is to help us better understand the
scientific method. The rigorous detail required for creating computer programs
that assist in the performance of scientific tasks forces us to explicate
concepts and procedures much more carefully than practicing scientists usually
do.

B. Medical relevance and collaboration

Molecular genetics has at least tno major connections to medical research.
Learning about the basic mechanisms which control the operation and transmission
of genetic information is necessary to understand and treat the wide range of
diseases and health conditions that are genetically controlled. Also, recent
developments in molecular genetics offer the promise of using genetic mechanisms
to produce essentially limitless amounts of drugs ard other biomedical
substances.

The MOLGEN project is a joint effort of the Computer Science Departments of
Stanford and the University of New Mexico and the Genetics Department of
Stanford. Major participants are Professor Nancy Martin and James Challenger of
the University of New Mexico; Professor Edward Feigenbaum, Professor Bruce
Buchanan, Dr. Randall Davis, Peter Friedland, and Mark Stefik of the Stanford
Computer Science Department, Professor Joshua Lederberg, and Jerry Feitelson of
the Stanford Genetics Department, and Professor Lawrence Kedes of the Stanford
Medical School. Jim Case, a graduate student in Professor Douglas Wallace's
laboratory, and Dr. John Sninsky, a molecular biologist working in Professor
Stanley Cohen's laboratory, are also collaborating in the MOLGEN project.

Cc. Progress summary
The major effort in MOLGEN has been the creation of a knowledge management

system. In addition, several specific problems which arise in genetics have been
examined in sufficient detail to result in reports and/or special purpose

157 J. Lederberg & E. Feigenbaum
Section 4.2.5 MOLGEN PROJECT

programs. We report briefly on two such programs, SAFE written by Peter
Friedland, and GA-1J written by Mark Stefik.

Knowledge management system

The success of MOLGEN as an experiment planner will depend on the quality
of its knowledge base. Therefore, much of the research effort to date has been
in the design and implementation of a knowledge representation and acquisition
system. All of the information relevant to the planning process will be an
explicit part of the knowledge base. The motivation for this aspect of the
design is the necessity to expand the program capabilities in a modular fashion
and to explain the rationale behind the program's planning behavior. We need to
represent concepts (e.g. enzyme}, instances (e.g. EcoRI), relationships among
concepts, and relationships among instances. In addition, we need to represent
processes. We have purposely limited the expressive power of our representations
to enable us to clearly define their semantics.

The result of this work is the Unit Package. Although this package has
been designed in the context of our genetics application, the package does not
contain any genetics knowledge.

One important aspect of the design of the system is that the knowledge base
contains knouledge about its own data representations. We have provided what we
term a "bootstrap knowledge base." It contains domain independent knowledge about
commonly used data types. When using our knowledge base in a new domain, an
artificial intelligence researcher would probably start with the bootstrap
knowledge base and then proceed to create units for the specific knowledge of his
task area. Both the AGE and genetics knowledge bases have been started in this
manner. The bootstrap knonledge base serves to illustrate our approach to
extensibility. Most of the bootstrap knowledge base is made up of primitive
datatypes. To add a new datatype to our system, one needs to provide the
knowledge base with procedures for some basic operations -- such as editing and
printing. Actually, the same approach is used in the unit package for defining a
new datatype as is used for defining a new enzyme. The process of defining new
datatypes requires, however, an understanding of Interlisp because the primitive
processes in the system are grounded in that language. New datatypes must be
defined together with their basic operations and entered into the knowledge base.

Knowledge base contents

The genetics knowledge base is growing rapidly. Approximately 602 of the
commonly used enzymes have been characterized. A beginning has been made on the
characterization of organisms such as bacteria and phages, plasmids and other
vectors, and genes. Our knowledge base also contains a growing collection of
genetic strategies for discovering information about various aspects of DNA
molecules, as well as a hierarchy of laboratory techniques which are used to
instantiate the strategies. The hierarchy of techniques includes modification,
separation, visualization, sequence analysis, and bacteriological techniques at
many levels of abstraction.

J. Lederberg & E. Feigenbaum 153
MOLGEN PROJECT Section 4.2.5

Safe program

The geneticist needs to predict what restriction enzymes can safely be used
to excise a gene, i.e. which ones can be guaranteed not to cut the functional
part of the gene. We would also like to know the approximate location of the
possible cutting sites of other restriction enzymes. This would all be very easy
if the complete DNA sequence of the gene was known. Sequence information is
becoming more and more prevalent, but it is still uncommon to know the complete
sequence of a gene. However, it is not unusual to know what protein the gene
codes for and to know the amino acid sequence of a protein.

Knowing the amino acid sequence does not provide full information because
of the degeneracy in the genetic code. One codon (a triplet of nucleotides)
specifies only one amino acid, but up to six different codons may specify the
same amino acid. The problem therefore, is combinatorically difficult. Typical]
proteins are up to 300 amino acids long (900 nucleotides), and all possible
nucleotide sequences which would produce the protein in the three possible phases
have to be considered.

The SAFE program lists the restriction enzymes that are currently stored in
the knowledge base and allows the user to add new ones. Besides determining
which enzymes are safe to use for gene excision in a particular DNA molecule, the
program also gives the position in the amino acid sequence where the possible
cutting site would be located.

GA-1 program

A common task in molecular genetics laboratories is the analysis of DNA
structure from restriction enzyme segmentation data. This task is one of the
simplest, although time-consuming, analysis tasks in molecular genetics. Two
standard approaches to solving this problem were examined: a data-driven strategy
and a model-driven strategy. These approaches are discussed and compared in
terms cf sensitivity to missing data, efficiency in the use of data, and other
measures of performance in [Stefik, 78]. A program was designed and implemented
which is superior to human performance on smaller problems, both in speed and
reliability. However, on large problems human problem solvers can use extra
structural constraints to out perform the program. The current program uses only
constraints derived from segmentation data itself. Geneticists usually known
additional information -- eg. that a given segment is on the end or that certain
segments must be adjacent.

A further benefit of this work was the suggestion of two new lab
techniques: combining multiple enzyme digests and incomplete digests. These
ideas arose from a systematic examination of evidence and inference rules that
went into building the program.

159 J. Lederberg & E. Feigenbaum
Section 4.2.5 MOLGEN PROJECT

D. Publications

Challenger J., A Program for Printing DNA Structures, CIS Report 78-3 CApril
1978)

Feitelson J., Stefik M.J., A Case Study of the Reasoning in a Genetics
Experiment, Heuristic Programming Project Report HPP-77-18 (Working
Paper) (May 1977)

Martin N.,» Friedland P., King J., Stefik M.J., Knowledge Base Management for
Experiment Planning in Molecular Genetics, Fifth International Joint
Conference on Artificial Intelligence. 882-887 CAugust 1977)

Stefik M., Friedland P., Machine Inference for Molecular Genetics: Methods and
Applications, Proceedings of the National Computer Conference, (June
1978)

Stefik M.J., Martin N., A Review of Knowledge Based Problem Solving As a Basis
for a Genetics Experiment Designing System, Stanford Computer Science
Department Report STAN-CS-77-596. (March 1977)

Stefik M., Inferring DNA Structures From Segmentation Oata: A Case Study,
Heuristic Programming Project Report HPP-78-3 (January 1978)

E. Funding Support Status

MOLGEN is funded by two NSF grants for the 2-year period 7/1/76 through
7/1/78. These grants, including indirect costs, are:

MNCS-7611935 PI: Nancy Martin, UNM Amount of $68,000
Univ. of New Mexico

MCS-7611649 PI: Edward Feigenbaum Amount of $110,706
Joshua Lederberg,
Stanford University

Two additional grants have been requested for the period 8/1/78 through 8/1780.

II. Interactions With The SUMEX-AIM Resource

All system development has taken place on the SUMEX-AIM facility. The
facility has not only provided excellent support for our programming efforts but
has served as a major communication link among members of the project. Through
the SUMEX-AIM facility, program development has taken place concurrently at
Stanford and New Mexico. Systems available on SUMEX-AIM such as INTERLISP, TV-
EDIT, and BULLETIN BOARD have made possible the project's programming,
documentation and communication efforts. The interactive environment of the
facility is especially important in this type of project development.

J. Lederberg & £. Feigenbaum 160
MOLGEN PROJECT Section 4.2.5

We have taken advantage of the collective expertise on medically-oriented
knowledge-based systems of the other SUMEX-AIM projects. In addition to
especially close ties with other projects at Stanford, we have greatly benefitted
by interaction with other projects at yearly meetings and through exchange of
working papers and ideas over the system.

The combination of the excellent computing facilities and the instant
communication with a large number of experts in this field has been a determining
factor in the success of the MOLGEN project.

III. Research Plans
A. Praject goals and plans

In exploring the three major motivations mentioned in section I.A. for
creating the MOLGEN project, there are many specific subproblems. We have
identified five for concentrated effort in the next two year period.

(1) Creating a more comprehensive genetics knowledge base. Expanding the
knowledge base within the area of DNA structural manipulation problems.

(2) Abstracting and Saving Plans. Recognizing when newly-created experiment
designs are worth saving and then generalizing those plans so they are
useful for more than the specific problem environment which caused their
generation.

(3) Making use of the process of hypothesis formation to help debug MOLGEN-
produced experiment designs. This process is especially important in a
domain like molecular genetics where incomplete knowledge about objects
and processes is the rule rather than the exception.

(4) Experiment planning by analogy. MOLGEN provides an excellent environment
for exploring various types of analogical reasoning. We integrate
problem-solving by analogy into the experiment design system as one of
the possible tools for solving subproblems.

(5) Performance evaluation as an integral part of the knowledge
representation and acquisition system. We view the process of evaluating
a system's performance and suggesting improvements as an Al problem
solving task. The strategies for this evaluation will be stored within
the same framework as all other MOLGEN planning strategies.

Each time the Heuristic Programming Project at Stanford has built another
large AI program, we have learned more about how to do it better and faster next
time. For example, the production rule interpreter in Heuristic Dendral (for
special-purpose rules) became the general rule interpreter of MYCIN. One of the
significant products of MOLGEN research will be the sets of ideas and programs
for encoding and manipulating large amounts of knowledge about a scientific
discipline. te have transferred some parts of the NGLGEN Units package to
another project interested in building a knowledge base about AI methods and
techniques. Making the tools used here available for use in neW programs is an
important aspect of our work and is generally important for cumulation of

161 J. Lederberg & E. Feigenbaum
Section 4.2.5 MOLGEN PROJECT

knowledge in the AI field. In order to do this we must reformulate the methods
so they are more generally applicable and more readily combined in diverse ways,

B. Justification and requirements for continued SUMEX use.

The MOLGEN project is completely dependent on the SUMEX facility. While we
have solved many of the original problems facing us in a manner useful to working
geneticists, we are just in the middie phase of building a planning system,
Without support from SUMEX to complete this system, many of the results of the
last two years will be ineffective. In the past six months our interactions with
geneticists outside of Professor Lederberg's laboratory have increased greatly.
The geneticists are excited about helping us with our knowledge base. Also, with
our help, they are finding useful ways to use the computing facility in their
current research. Thus the serendipity of supporting MOLGEN is the creation of
many useful computer research programs.

We are asked to state our requirements for continued SUMEX use. We project

that our usage of processor cycles and file storage will grow to twice the
current levels in the coming year.

J. Lederberg & E. Feigenbaum 162
MYCIN PROJECT Section 4.2.6

4.2.6 MYCIN PROJECT

MYCIN - Computer-based Consultation in Clinical Therapeutics

S. N. Cohen, M.D. (Genetics and Pharmacology) and
B. G. Buchanan, Ph.D. (Computer Science)
Stanford University

1) Summary of research

The MYCIN program is an outgrowth of nearly a decade of work on DENDRAL.
We are building on, and improving, many of the ideas from DENORAL about
representing large amounts of domain-specific knowledge for computer aided
problem solving. The representation, use and acquisition of knowledge for
computer programs has been called "knowledge engineering" [D.Michie, On Machine
Intelligence, New York: Wiley, 1974]. We represent that body of knowledge as a
collection of decision rules about diagnosis and therapy selection in infectious
diseases. These conditional sentences are called "production rules". The
production rule formalism provides an easily understood representation of facts
and relations.

The MYCIN knowledge base currently consists of approximately 400 such
rules. Each rule consists of a set of preconditions (called the "premise")
which, if true, justifies the conclusion made in the "action" part of the rule
(an example is shown below).

If 1} the gram stain of the organism is gram negative,
and 2} the morphology of the organism is rod,
and 3) the aerobicity of the organism is anaerobic,

Then there is suggestive evidence (.6) that the identity of
the organism is Bacteroides.

The Medical Problem

A number of recent studies (documented in Shortliffe, 1976) indicate a
major need to improve the quality of antimicrobial therapy. Almost one-half of
the total cost of drugs spent in treating hospitalized patients is spent on
antibiotics, and if results of a number of recent studies are to be believed, a
significant part of this therapy is associated with serious misuse. Some of the
inappropriate therapy involves incorrect selection of a therapeutic regimen,
while another serious problem is the incorrect decision to administer any
antibiotic. One recent study concluded that one out of every four people in the
United States was given penicillin during a recent year, and nearly 90% of these
prescriptions were unnecessary. Other studies have shown that physicians will
often reach therapeutic decisions that differ significantly from the decisions
that would have been suggested by experts in infectious disease therapy
practicing at the same institution.

Initial culture reports from a microbiological laboratory may become
available within 12 hours from the time a clinical specimen is obtained from the

163 J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

patient. While the information in these early reports often serves to classify
the organism in general terms, it does not often permit precise identification.
It may be clinically unwise to postpone therapy until such identification can be
made With certainty, a process that usually requires 24 to 48 hours, or longer.
Thus it is commonly necessary for the physician to estimate the range of passible
infecting organisms, and to start appropriate therapy even before the laboratory
is able to identify the offending organism and its antibiotic sensitivities. In
this setting MYCIN plays tuo roles: (a) providing consultative advice that will
assist the physician in making the best therapeutic decision that can be made on
the basis of available information, and, (b) by its questioning of the physician,
pinpointing the items of clinical data that are necessary to increase the
validity of the clinical decision.

Progress Summary

In the past year our work has been guided by three fundamental objectives,
roughly corresponding to use, explanation, and acquisition of medical knowledge
in a symbolic reasoning program. These are discussed in the next three sections.

Use of Knonledge

A major objective of the MYCIN system has been to provide a computer-based
therapeutic tool designed to be useful in both clinical and research
environments. This requires development of a system that has a medically and
scientifically sound knowledge base, and that displays a high level of competence
in its field.

Expansion of Knowledge Base

Work on improvements to the knowledge base included both expansion into new
areas and further development of existing areas of expertise. A start has been
made, for instance, on rules for dealing with cystitis infections, a very common
source of bacterial infection. In addition, a urinary tract infection therapy
grid has been developed. This is a large table containing basic information
about drug and dose selection, which functions as a basic information source for
the therapy selection routines in the program.

We have also developed a facility that plots the steady state blood levels
of antibiotics, based on a range of patient-specific parameters (body surface
area, level of kidney function, etc.). This facility presents a very clear
picture of the consequences of selecting various dose levels and intervals, and
Will be very useful in helping a clinician select a therapy regimen that is
maximally effective without endangering the patient due to toxic levels of drug
in the blood.

Extensions to the system included improving the design of rules so that
they now include "justifications" and literature references. That is, rather
than simply indicating that "gram negative rods in a nosocomial infection are
likely to be pseudomonas", the rule now has a "justification" which is a further
explanation of the reasoning behind the rule. This justification can be printed
out at the users request, and may indicate, in this case, that pseudomonas is

J. Lederberg & E. Feigenbaum 164
MYCIN PROJECT Section 4.2.6

more common in the hospital setting, and hence a likely causative organism in
nosocomial infections. The literature references are pointers to published
articles that give further information Cincluding original clinical studies)
which serves to illustrate the reasoning behind the rule.

Finally, as a result of our recent evaluation of the system's performance
on meningitis cases, we have made several improvements to the body of rules
dealing with this area.

Evaluation of Competence

A formal evaluation of Mycin's performance in recommending treatment for
patients with infectious meningitis was begun in May of 1977. The study design
enabled unbiased comparisons to be made of Mycin's therapy recommendations with
those of infectious diseases specialists and physicians of varying degrees of
medical expertise. These comparisons provided a means to establish the level of
competence of the Mycin system in prescribing therapy for patients with
meningitis.

Ten patients with infectious meningitis of variable etiologies were
selected for the study by a physician not familiar with the Mycin system. A
detailed clinical summary was compiled for each patient. These summaries were
presented to five faculty members in the Division of Infectious Diseases at
Stanford, one senior research fellow in infectious diseases, a senior resident in
medicine and a senior medical student, with a request to select therapy for each
case. A Mycin consultation was performed far each patient using the same clinical
summary data. As a result, a total of ten therapy recommendations, including the
actual therapy the patient received, were compiled for each case.

The patient summaries were then given to eight prominent infectious
diseases specialists at institutions other than Stanford. They were asked to
select therapy for the patients and to evaluate the recommendations of the
Stanford physicians, the student, Mycin and the actual therapy the patient
received. The therapies were listed in random order for each case and the
national evaluators were not informed of the identities of the ten prescribers.
The evaluators used the following criteria to rate the prescribers'
recommendations: "equivalent", the recommendation was identical or equivalent to
his oun; "close call", the recommendation was different but the evaluator was
willing to rate it as an acceptable alternative; and "not acceptable", the
recommendation was inappropriate.

The results of the evaluation, as summarized in the table beloun,

demonstrate that Mycin's competence in selecting antimicrobial therapy for
meningitis is comparable to that of the infectious disease faculty at Stanford.

165 J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

Prescribers Acceptability
Rating
MYCIN | 52
Faculty-1 | 50
Faculty-2 | 48
ID Fellow 1 48
Faculty-3 | 46
Actual therapy l 46
Facul ty-4 l 44
Resident J 26
Faculty-5 | 34
Student | 24

The acceptability rating is the cumulative total of therapy selections rated as
"equivalent" or a "close call". Since there are eight national evaluators and

ten cases to be evaluated a perfect score would have been 80. Mycin's therapy

recommendations received the largest number of acceptable ratings.

Use of Prototypes to Guide the Reasoning

A consultation program is currently under development which uses a set of
domain-specific prototypes of typical situations to guide the invacation of
production rules. This pregram has been implemented in the domain of pulmonary
physiology in which the prototypes represent the various pulmonary diseases.
Fach prototype consists of a group of situation-specific components, each having
plausible values, likely trap values and a possible default value.

The addition of prototypes to the rule-based system allons us to:

1) Control the consultation using information given in the prototypes. The
determination of values of parameters in the original consultation system
(which in turn causes rules to be tried and questions to be asked) is
replaced by the determination of values of components in a prototype,

i.e., "filling in" the prototype. This "search" for information is more

focused since it occurs within a more limited context, namely, the context

defined by the prototype.

In addition, once a prototype is chosen, a "control element" associated
with the prototype is invoked. This control element explicitly states
actions to be performed in "filling out" the prototype,

2) Analyze the result of the consultation when a value is unknown vs. the
result when a default value is used. That is, if we provide a default
value for a component, does it make a difference in the result? For
example, if it is found that using a default value for a missing lab test
results in a different therapy for a patient, then the user can be
informed of the significance of that lab test.

3) Detect inconsistencies in the data using the Plausible Values associated
with the component in the prototype.

J. Lederberg €& E. Feigenbaum 166
MYCIN PROJECT Section 4.2.6

4) Detect error conditions and suggest actions to correct them using the
Likely Trap values associated with the component in the prototype.

5) Give explanations of system performance in terms of the prototypes. A
trace of system actions is provided by noting which prototype is invoked
at any time.

6) Classify consultations in terms of the prototypes, allowing indexing and
retrieval of consultations for uses such as testing a new rule relevant to
some situation.

Compilation of the Rule Base into a Decision Tree

In an effort to overcome the inherent slowness of interpreting an
increasingly large set of production rules, work is underway to "compile" the
rule base. The block of rules which make conclusions about a particular clinical
parameter implicitly constitutes a program to conclude the parameter. Under the
present control structure, that program is trivial; the rules are simply
interpreted one at a time; if a predicate appears in 50 rules, it will be tested
50 times. The rule compiler, however, can construct an explicit LISP program
from those rules, and, by making compile-time inferences, restructure it to
eliminate redundant computation of related premise clauses. The result is an
optimal decision tree, which effectively tries several rules in parallel; a
single test in the compiled program Ce.g. "what is the infection?") can cause a
large set of rules to fail at once.

The rule compiler will thus allow the consultation program to use an
efficient deductive mechanism, even as the rule base expands, while the flexible
rule format is still available for explanation and debugging.

Development of a Tutoring System

The objective of this part of the project is to implement and test a
computer program for tutoring medical students and physicians in infectious
disease diagnosis and therapy. The problem-solving and dialogue capabilities of
this tutor are distinct so that arbitrarily complex case histories may be
discussed with the student or physician.

The tutor itself has the capability to discuss the problem and the expert
program's solution with a student. Therefore, the teaching capability is
distinct from the problem-solving expertise of the consultation program. This
will enable us to study alternative teaching methods.

This work proceeds by formalizing the conventional rhetorical patterns one
finds in tutoring dialogues. The resultant formalization is, as in the
consultation program, independent of any particular case, so the tutor will have
the capability to discuss any case presented by a student or selected from a
computerized library of cases. This project is similar in style to current
research in the area of "intelligent computer aided instruction." The power of
these programs derives from 1) separation of problem solving ability from
teaching strategies, 2) construction of a student model which describes the

167 J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

student's understanding with respect to the decisions of the expert program, and

3)

consideration of explanation techniques that make a point clearly and capture

what the student needs to knon.

iD

2)

3)

4)

J.

The initial design of the tutor has involved three major steps:

Modification of the control structure of the expert program to leave behind
detailed traces for use by the tutorial program. These traces indicate
specifically which information was necessary to apply each rule; they are
significantly more complex than the traces used by the current question-answer
program in MYCIN.

Formalization of tutorial discourse procedures. Proceeding from an extensive
hand simulation, patterns in the tutorial dialogue were formalized inte
procedures that constitute a sequence of action options for the tutor. For
example, the most complex procedure details how to discuss a topic with the
student: it captures the exchange of initiative as the student requests more
data and the tutor quizzes him about new conclusions he can draw. Another
procedure captures the situation in which the tutor mentions rules that are
related to one that the tutor and student have just discussed. These
procedures have been written in stylized INTERLISP so they are translatable
into an English form. This makes it easy for us to show other researchers and
educators the specific rhetoric patterns, domain knowledge, and teaching
strategies used by our program. There are about 20 such procedures in the
current formalization of the case method dialogue.

Formalization of teaching strategies and the communication model. Again using
stylized code, rules have been written to decide what the student knows during
the dialogue, based on his behavior and past history. For example, given a
student hypothesis, the. tutor applies communicaticn model rules to determine
which of the expert (domain) rules have probably been considered by the
student. The teaching strategies use this belief as a starting point and make
an appropriate response. A range of strategies have been formulated,
consisting mainly of ways to state, hint or quiz about a rule. A taxonomy of
question types based on difficulty allows the tutor to choose a strategy
appropriate to the student's knowledge of the (domain) rule. Thus, the
communication model and teaching strategies serve to control] the rhetorical
options that have been specified in the discourse procedures.

Augmentation of the expert knowledge base. Several forms of meta-level
knowledge and annotations have been added to the (domain) knowledge base.
These include: a) rule schema, or templates that describe a type of rule and
tts import (Centered by a domain expert); b) key factors, or distinguished rule
clauses that are derived by the tutor from the rule schema and indicate the
key piece of information being taken into account by the rule; and c) factor
interrelationships, such as an indication that when one factor is known other
factors are not relevant.

Lederberg & E. Feigenbaum 168
MYCIN PROJECT Section 4.2.6

Explanation of Reasoning

One of the important contributions of the MYCIN work to artificial
intelligence has been our persistent emphasis on an intelligent system with the
ability to explain the reasoning behind its decisions. It should be able to do so
in terms that suggest to the physician that the program approaches the problem in
much the same way that he does. This permits the user to validate the program's
reasoning, and modify (or reject) the advice if he believes that some step in the
decision process is not justified. It also gives the program an inherent
instructional capability that allous the physician to learn from each
consultation session.

The use of a rule-based representation of knowledge makes it possible for
the system to explain the basis for its recommendations. For example, if asked
"How did you determine the identity of the organism?" the program answers by
displaying the rules which were actually used, and explaining, if requested, how
each of the premises of the rules was established. This is something which
people readily understand, and it provides a far more comprehensible and
acceptable explanation than would be possible if the program were to use a simple
statistical approach to diagnosis.

Extension of Question Types

Additions to MYCIN's question-ansuering system within the last year fall
into three categories: general improvements to the parsing mechanism, refinements
to answers that are given, and new types of questions that can be recognized and
answered.

The first major improvement to the parsing mechanism is that each word in a
particular question is given a single interpretation. In particular, a word may
not implicate both a context and a parameter, or more than one parameter. When
dictionary pointers alone are not enough to distinguish among ambiguous
interpretations of a word, the routine that generates an answer to the question
will select the interpretation that best answers the question. For example, ina
rule-retrieval question where a single word implicates several different
parameters, the parameter that is used in the most rules will be chosen.

Another change is that it is no longer necessary to explicitly mention the
name of some context in a question that is specific to the consultation. Rough
semantic analysis is often successful in determining that a question is about the
consultation, and in selecting the appropriate context.

A final improvement to interpreting the question is in the analysis of
questions that mention more than one context from the consultation. In the past,
one single context was chosen, and the selection was often arbitrary. Now,
heuristics are used to choose among the contexts, and in certain questions, more
than one will be used Ce.g. in "DID YOU USE THE SITE OF CULTURE-1 TO DETERMINE
THE IDENTITY OF ORGANISM-2?").

Questions that the system has always been able to answer now get answers

that are closer to what the user wants. One example of this is that the system
now differentiates between the conjunction and the disjunction of parameters in a

169 J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

question. In the past "HOW DO YOU USE THE FACT THAT AN ORGANISM IS A GRAN
NEGATIVE ROD" would have been answered by listing the rules that use either
negative gramstain or marphology rod instead of those that use both. In addition
to simply listing the rules, the system non also summarizes how they are used Cin
the question above, it would add that most of the rules that use those two
factors (gram negative, and rod) are used to conclude about the identity of the
organism). When printing rules in answer to a question, the system now includes
a brief explanation of when the rule will be used, and gives the user the option
of seeing the rule's author, medical justification, and relevant literature
references.

Some of the "holes" in the question-answerer's repertoire have been filled
in. The system can now explain how it calculated the dosage of the drugs that it
prescribed and it can explain how it decided to treat for particular infections
and organisms.

Development of Automatic Testing System

We have added a mechanism for automatically testing the question-answering
system which will be used to debug the system and to maintain it in a working
state as the consultation system changes around it. Testing in the past has been
tedious due to the necessary interaction with a person asking questions and
verifying the validity of the answers.

The new testing program consists of two phases. The first allows a MYCIN
expert to run QA and save information on sample questions that were answered
correctly. The information includes the question itself, intermediate parses,
and an encoding of the answer that was given. This information is stored in a
file.

The second phase is undertaken after substantive changes have been made to
the system. A batch job is submitted which will run a consultation, then ask the
questions that were stored on the file created during the first phase of testing.
This job writes a report of the comparison of the new parse and answer with what
is stored in the file. The discrepancies that are noted in these reports can be
used to pinpoint the sources of bugs in the QA system, if any.

Acquisition of Knowledge

A third major objective has been to provide the program with capabilities
that enable augmentation or modification of the knowledge base by experts in
infectious disease therapy, in order to codify knowledge in the domain, as well
as to improve the validity of future consultations. The system therefore
requires some capability for acquiring knowledge by interacting with experts in
the field, and for incorporating this knowledge into its knowledge base.

Substantial effort was placed in updating sections of Teiresias, the
knowledge acquisition system developed by R. Davis [Davis, 1976]. In particular
the post-consultation review procedure was modified to utilize the existing
question-answering facilities to describe how particular conclusions were made.
This allows the user direct access to the full capabilities of the question-

J. Lederberg & £. Feigenbaum 170
MYCIN PROJECT Section 4.2.6

answering system while the system aids the expert in isolating the problems in
the reasoning process by unfolding the reasoning tree using the methods of the QA
module to describe conclusions of rules.

The actual rule-acquisition procedure has been improved by the development
of a new parser to perform the actual translation of English to the internal LISP
representation.

During the design and implementation of another knowledge base for an
EMYCIN system, a number of issues dealing with the initial acquisition of a
knowledge base were identified. They are primarily the result of having to elicit
and encode substantial amounts of domain-specific knowledge from an expert who is
a novice at explicating and representing this knowledge in the form of rules and
parameters. We are pursuing the design of facilities which will aid the new
expert during the initial stages of this knowledge explication process.

MYCIN PUBLICATIONS

Shortliffe, E H, Axline, S G, Buchanan, BG, Merigan, T C, Cohen, SN. An
artificial intelligence program to advise physicians regarding
antimicrobial therapy, Computers and Biomedical Research, 6:544-560
(1973).

Shortliffe, E Hs Axline, S$ G, Buchanan, B G, Cohen, S N, Design considerations
for a program to provide consultations in clinical therapeutics,
Presented at San Diego Biomedical Symposium 1974 (February 6-8, 1974).

Shortliffe, E H, MYCIN: A rule-based computer program for advising physicians
regarding Antimicrobial therapy selection, Thesis: Ph.®. in Medical
Information Sciences, Stanford University, Stanford CA, 409 pages,
October 1974. Also, Computer-Based Medical Consultations: MYCIN, American
Elsevier, New York, 1976.

Shorttliffe E H MYCIN: A rule-based computer program for advising physicians
regarding antimicrobial therapy selection Cabstract only Proceedings of
the ACM National Congress (SIGBIO Session), p. 739, November 1974.
Reproduced in Computing Reviews 16:331 (1975).

Shortliffe E H, Rhame F S, Axline S G, Cohen S N, Buchanan B G, Davis R, Scott A
C, Chavez-Pardo R, and van Melle W J MYCIN: A computer program providing
antimicrobial therapy recommendations Cabstract only). Presented at the
28th Annual Meeting, Western Society For Clinical Research, Carmel, CA, 6
Feb 1975. Clin. Res. 23:107a €1975). Reproduced in Clinical Medicine,
p. 34, August 1975.

Shortliffe, &— H and Buchanan, BG, A Model of Inexact Reasoning in Medicine,
Mathematical Biosciences 23:351-379, 1975.

171 J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

Shortliffe, EH, Davis, R, Axline, S G, Buchanan, B G, Green, C C, Cohen, SN,
Computer-based consultations in clinical therapeutics: explanation and
rule acquisition capabilities of the MYCIN system, Computers and
Biomedical Research, 8:303-320 (August 1975).

Shortliffe E H, Axline S, Buchanan BG, Davis R, Cohen S, A computer-based
approach to the promotion of rational clinical use of antimicrobials,
International Symposium on Clinical Pharmacy and Clinical Pharmacology,
Sept 1975, Boston, Mass. Cinvited paper)

Shortliffe EH, Judgmental knowledge as a basis for computer-assisted clinical]
decision making, Proceedings of the 1975 International Conference on
Cybernetics and Society, pp 256-7, September 1975.

Davis R, King J J, An Overview of Production Systems, Machine Intelligence 8:
Machine Representations of Knowledge (eds E H Elcock and D Michie), John
Wylie, April 1977.

Davis R, Buchanan B G, Shortliffe E H, Production rules as a representation for a
knowledge-based consultation system, Artificial Intelligence, Vol 8, No 1
(February 1977).

Shortliffe EH, Davis R, Some considerations for the implementation of knowledge-
based expert systems, SIGART Newsletter, 55:9-12, December 1975.

Scott AC, Clancey W, Davis R, Shortliffe £ H, Explanation capabilities of
knowledge based production systems, American Journal of Computational
Linguistics, Microfiche 62, 1977.

Wraith S, Aikins J, Buchanan B G, Clancy W, Davis R, Fagan L, Scott AC, van
Melle W, Yu V, Axline S, Cohen S, Computerized consultation system for
the selection of antimicrobial therapy. American Journal of Hospital
Pharmacy 33:1304-1308 (December 1976).

B.G. Buchanan, R. Davis, V. Yu and S. Cohen, "Rule Based Medical Decision
Making by Computer," Proceedings of MEDINFO.77, 1977).

Davis R, Applications of Meta Level Knowledge to the Construction, Maintenance,
and Use of Large Knowledge Bases. Memo HPP-76-7, Stanford University,
June 1976.

Davis R, Meta rules: content directed invocation, to appear, Proc ACM Canf. on Al
and Programming Languages, August 1977.

Davis R, Knowledge acquisition in rule-based systems: knowledge about
representations as a basis for system construction and maintenance, to

appear, Proc. Conf. on Pattern-directed Inference Systems, May 1977.

Davis R, Interactive transfer of expertise: acquisition of new inference rules,
to appear, Proc. Fifth IJCAI, August 1977.

Davis R, A decision support system for medical diagnosis and therapy selection,
in "Data Base" (SIGBDP Nenstietter), 8:58 (Winter 1977).

J. Lederberg & E. Feigenbaum 172
MYCIN PROJECT Section 4.2.6

Davis R, Buchanan B G, Meta level knowledge: overview and applications, to
appear, Proc. Fifth IJCAI, August 1977.

Clancey, W, The Structure of A Case Method Dialogue, accepted for publication in
The International Journal of Man-Machine Studies, fall '78.

FUNDING

KNOWLEDGE-BASED INTELLIGENT SYSTEMS
THE MYCIN PROJECT

Bruce G. Buchanan, Principal Investigator
Adjunct Professor, Computer Science
Stanford University

National Science Foundation
Grant M€S77-02712

Budget for Year 1. from 671777 through 5730778
Total Direct Costs $ 32,357.

2-Year Budget Summary from 671/77 through 5730779
2~Year Total Direct Costs 64,396.

II) Interactions with the SUMEX-AIM Resource
Collaborative Efforts
University of Rochester
Professor Charles Odoroff again requested access to MYCIN to use in his
medical school class on medical computing. In the past, the students who
examined MYCIN have provided valuable criticisms on which further developments
can be based.
Carnegie-Mellon University
Dr. John McDermott, working with Professor Allan Newell, visited Stanford
and has had extensive network communication with us about MYCIN's production rule
formalism. He is translating MYCIN's infectious disease rules into their oun OPS

formalism in order to study the advantages and disadvantages of writing "fine
grained"" rules, as in OPS, or "coarse grained" rules, as in MYCIN.

173 ‘J. Lederberg & E. Feigenbaum
Section 4.2.6 MYCIN PROJECT

ISI

Ors. William Mann and James Moore of ISI visited Stanford to discuss, in
part, their use of MYCIN to study man-machine communication. Because MYCIN has
been developed for physicians to interact easily with the program , it was
thought to be one of the most appropriate objects for their study.

EMYCIN

Part of our concern is to generalize the methods used in MYCIN, and make
them available to others. We have developed a prototype of an "empty MYCIN"
consultation system, called EMYCIN, that embodies most of the control structure,
and none of the specific medical knowledge, of MYCIN. That is, for domains that
are structured similarly to MYCIN's primary domain, the existing mechanisms for
offering consultations can be coupled with knowledge (rules) of the new domain.
Substantial progress has been made on providing a clear, consistent package for
use by other groups.

Pulmonary Function Testing and Intensive Care Unit Monitoring

The MYCIN program and techniques have been adapted for tno new
medical/Artificial Intelligence applications. The first domain is the
interpretation of the measurements made in the Pulmonary Function Laboratory from
the Pacific Medical Center at San Francisco. This application uses the MYCIN
program directly, but with a new rule set designed to diagnose the type and
severity of pulmonary disease, and some measure of the reliability of the tests
used in the laboratory. The second domain is the continuous interpretation
problem for patients with mechanical breathing assistance in the Intensive Care
Unit. This program uses the basic MYCIN architecture and concepts with
extensions to handle the change in patient condition from time to time. Both of
these applications are explained in more detail in the section on the PUFF
project.

HEADMED

We have continued strong collaboration with Dr. J.Heiser at U.C. Irvine who
has developed the HEADMED program from NYCIN. See HEADMED section for details.

Medical Consultation in Rheumatology

Preliminary design of a consultant for rheumatologic diseases was begun by
Dr. R. Blum. This system would arrive at a diagnostic conceptualization of a
case using a production rule type formalism as 1s used by MYCIN; however,
fundamental revisions to the control structure would guide the rule invocation.
It is anticipated that rule invocation would be in a data-driven Cforward)
direction rather than MYCIN's current goal-driven method. Rules would be
embedded in a hierarchy of diagnostic hypotheses, conceptually similar to the
INTERNIST Project at the University of Pittsburgh.

J. Lederberg & E. Feigenbaum 174
MYCIN PROJECT Section 4.2.6

The presence af a large clinical databank would permit the addition of
quantitative methods not presently utilized in the MYCIN formalism. The ARAMIS
databank, a collection of clinical data on thousands of patients with
rheumatologic diagnoses, will be used.

The major advantages of a consultant system compromised of both knowledge
derived from clinicians and from a clinical databank would be 1) the capability
for comparison of the two sources of Knowledge, 2) the advantage over a pure
databank approach of being able to provide advice when the data alone are
insufficient, and 3) the advantage over a purely rule-based system of possessing
quantitative precision when the data allow it in determining probabilities of
symptom/disease association or in predicting outcome parameters.

DENDRAL

Many problems of developing complex reasoning programs in INTERLISP are
common to both DENORAL and MYCIN. We have continued close association with the
DENDRAL project and are jointly working on solutions to the problems of managing
large programs.

Critique of Resource Management

Management of this resource remains as professional as ever. The service
from the staff, as well as from the computer, is superb.

As the computer becomes more heavily loaded we encounter more and more
reluctance from physicians to use the program more than once because of the
response delays. This inhibits our progress, needless to say. We are working on
software modifications to alleviate the problem; we are also looking forward to
acquisition of new system hardware, specifically the DEC 2020, to provide same
relief.

III. Research Plans
Long Range Goals

We intend to build the MYCIN rule-based consultation program into a tool
for physicians in limited areas of medicine. Or. E.H. Shortliffe will again take
over the leadership of the project and intends to become principal investigator
on new grant applications that we submit.

Because of our difficulty in securing funding for a phase of the project
which is neither basic research nor immediate health care delivery, we may be
forced to give up the idea of moving MYCIN out of the research environment with
federal support. There are many challenging research questions left to be
explored both in artificial intelligence and in medical computing, while ne seek
a solution to the problem of funding development of health care technology, as
opposed to either research or application.

175 J. Lederberg & E. Feigenbaum