DISTRIBUTED DATA-BASE SYSTEM FOR CHRONIC DISEASES

_ Dr. Robert Nordyke
Pacific Health Research Institute

Dr. Franklin F. Kuo
University of Hawaii

Dr. Casimir Kulikowski
Rutgers University

A. Introduction

We propose to establish a resource sharing project for the development
of computer systems for consultation and research to be made available to
clinical facilities from a set of distributed data bases. Our goal is to
investigate the relationships between patient data and computer models of
selected chronic diseases, with major emphasis on the problems of sequen-
tial decisions in patient management. We will build upon the HEW supported
work at Straub Clinic and Hospital on chronic disease clinics. We propose
to design a system that provides the clinician, nurse and paramedic with
information at a level of detail and sophistication appropriate for the
users' needs and suited to the particular patient's characteristics. To
do this we intend to design data acquisition and medical decision protocols
to be run on small, local computers. These local minicomputers are linked
via computer-communication networks (ALOHA and ARPANET) to the SUMEX computer
at Stanford, as shown in Fig. 1: The ALOHANET can make these consultation
programs available via radio-links and satellite to other Hawaiian islands
and remote areas of the Pacific basin, where nurse/paramedic protocols for
the management of chronic diseases could have a significant effect on the

quality of health care delivery.
Page 2

One of the principal objectives of our proposed research program is
the development of sophisticated computer programs based upon methods of
artificial intelligence to aid in the diagnosis and treatment of specific
chronic diseases. Since these chronic disease consultation programs are
of immediate usefulness in general health care, we propose to have these
programs resident in SUMEX so that they can be accessed from any terminal
in TYMNET or ARPANET. Thus we propose to develop these chronic disease
programs as a central resource at SUMEX to be shared by clinical investi-
gators in the HEW and NIH health care community.

The consultation programs within SUMEX will provide: multiple models
of disease, varying levels of resolution, different modes of interpretation
(causal, logical, taxonomic, associative, probabilistic, etc.), facilities
for explanation, instruction and querying of a data base of existing cases.
A control program at the central resource (SUMEX) will receive requests
from the local clinical control program and decide on the appropriate
level and scope of the response. The local minicomputer will limit the
type of information transfered to the resource machine to maintain security
and confidentiality of the medical information. Local mass data will be
on disc and on a large time-sharing computer, the BCC 500 at the University
of Hawaii.

In summary we are proposing to develop a prototype of a distributed
data base system for chronic diseases. Our general reasons for choosing
chronic diseases as an area of investigation are that, because of their

numbers, even a small increase in efficiency would have a major absolute
Page 3

impact in health care delivery. As a group they take up a sizeable
fraction of both primary care and specialist physicians' time (perhaps

one third to one half), often inappropriately. The protracted maintenance
periods inherent in the care of patients with chronic diseases, in addition
to their ever-accumulating mumbers, renders coordinated, individualized,

quality-controlled care difficult.

B. Clinical Objectives in Chronic Diseases
The main objectives of a chronic disease clinic are to maximize the
well-being of patients with chronic diseases, minimize the effects of
the disease upon the individual, with the most appropriate use of person-
nel and other resources at a reasonable cost, by providing:
a. an efficient and complete evaluation of the patient
b. confirmation of diagnosis or problem
Cc. maximum patient involvement in long term care
d. continuing training and education of personnel, including physicians
e. patient education regarding the disease, treatment, side effects
f. extension of the roles of nurses and allied health personnel
g. observation and comparison of patient's condition and course
throughout therapy
h. consultation on complex.problems - diagnostic or therapeutic
i. a standardized data pool for quality control including PSRO and
study of the disease entities and related factors

j. periodic automated peer review (cost, process, outcome)
Page 4

At Straub Clinic and Hospital a number of nurse-managed, physician-
Supervised chronic disease clinics are in full operation today. These
clinics were developed through an HEW grant to the Straub Medical Research
Institute (now the Pacific Health Research Institute) and include hyper-
tension, thyroid, Parkinson's disease, diabetes, gout, cancer chemotherapy,
and multiple diseases. The nurse-manager is a registered nurse who has
been trained by a supervising physician for assisting with the management
of patients with the chronic disease in an uncontrolled state and for the
primary management of patients with stabilized chronic disease. The
present clinics meet many of the objectives, but falls short in part on
objectives a, g, h, i, and j.

Thus far, the Straub chronic disease clinics (excepting the thyroid
clinic) have been operating without the aid of a digital computer. Much
of the work of these clinics involve management protocols based upon
patient data taken during three stages --evaluation, stabilization and
chronic maintenance. It is easy to see how useful a computer can be in
just maintaining, analyzing and summarizing the large volume of data that
is collected during the protracted period of chronic disease maintenance
(usually the lifetime of the patient).

The supervising physician's role in the present day clinics comprise of
the following:

1. Review, clarify, and amplify if necessary the nurse and self-admin-

istered computer stored and retrieved history and physical examina-
tion

2. Confirm data as necessary
Page 5

Review lab studies and order additional tests where necessary

Pf WY

Make diagnoses

wn
°

See patients for evaluation and maintenance as indicated

6. Order medication and be responsible for necessary changes
Since the physician's load for chronic diseases is extremely high (his
patients tend to accumulate with time and the data per patient grows
steadily), a set of computer programs which will aid him in items 1 - 5
can be of great value. One of the primary objectives in this proposal
is to develop more sophisticated management protocols and consultative
aids within the framework of the chronic disease clinics as described
above.

The computer based systems we propose to develop will be different
from most designed to date in that we must deal with protracted follow-up
management, which may involve complex treatment decisions in some cases,
but which in others may be reasonably stable and easy to handle.
Repetitive observations and comparison to previous data is required in
both cases, with management by nurse or allied health personnel as a
natural outcome of the system once standard "profiles" have been identified
for prognosis and treatment.

Our experience to date in the existing chronic disease clinics have
indicated that this approach leads to improvements in measurable health
outcomes, improved cost efficiencies, and satisfaction on the part of patients,
nurses, and physicians even in non-automated settings. A computer based

system specifically designed to bring in and analyze data at a high level
Page 6

of sophistication should only improve these benefits further.

: In our development program we will start by investigating 3 specific
diseases--thyroid, hypertension, and diabetes, building a generalizable
system which will later encompass other important chronic diseases such as
chronic heart failure, follow-up of myocardial infarction, arthritis,

gout, peptic ulcer and others. Thyroid disease is chosen as one of the
starting clinics because we have accumulated six years of experience with an
interactive computer aided system for consultation, summary and long term
management. Hypertension has been chosen because the clinic is also well
established (with off-line records) and because it is such a common and
important chronic condition. Diabetes is chosen for the same reasons and
also because it often occurs in the same individual as hypertension, giving

us a first example of a common combination of chronic conditions.
Page 7

C. Decision-Making Methods: Multiple Models

In the past four years artificial intelligence approaches and tech-
niques have inspired the development of several rather different programs
for medical consultation.

The application of traditional statistical methods, decision tree
techniques and various heuristics have shown their value in quite a number
of medical areas, but they have fallen short.of satisfying either the
computer science designers or the clinical users. To the former it often
represents a direct application of known techniques with little scope
for innovation; to the latter, the physician, the task of providing reli-
able statistics or unique decision logic often proves difficult enough,
but problems of interpreting unfamiliar forms of inference is a hindrance
to acceptability and transferability. But the main drawback of traditional
approaches has been that they impose some over-simplified structure on the.
medical knowledge needed to make a decision and as a result much that is
commonplace and obvious to the clinician does not fit the structure and is
lost in the computer's decision-making.

What gradually became apparent to several workers in this field is
that if the computer is to become a worthwhile consultant we must develop
flexible and sophisticated means for representing medical knowledge so an
expert can pour his experience and knowledge into a program without it
rapidly becoming an unwieldly and often undecipherable conglomerate of

special purpose rules and information.
Page 8

It is oversimplistic to assume that a single method of representa-
tion of knowledge or a single method of decision-making is appropriate
for any disease.

Current work on artificial intelligence in medicine has led to
several different approaches, each possessing advantages specific to the
diseases that were investigated. The principal types of models used to
describe knowledge relevant to diseases are:

* causal-associated process models with related structure of observa-

tions [Kulikowski & Weiss, 1971, Amarel § Kulikowski, 1972, Weiss, 1974]

+ logical consequent rule (premise-action) models [Shortliffe, et. al.,

1974]

* hierarchical taxonomic models [Pople, 1974]

* frame-based models [Gorry, et. al., 1974]

The logical rule-based and frame-based systems are very general,
information processing models that can encompass a wide variety of medical
information. The causal process models require that mechanisms of disease
be reasonably well known, as do the taxonomic models if they are to be use-
ful in decision-making. These last provide a structure that can be used to
link a large set of diseases, while the former describe the fine structure
of the course of related diseases.

The set of chronic diseases we are investigating encompasses sufficient
variety within a class of similar, selected problems. We can foresee the
need to apply several of the above kinds of models depending on the extent

of knowledge in each field. For example, a causal model can be used to
Page 9

describe those aspects of thyroid dysfunction directly related to the produc-
tion of thyroid hormone, whereas a rule-based system may prove more effect-
ive in handling the great variety of loosely related peripheral findings

that result from the metabolic imbalance. Similarly, a causal analysis

can be useful in distinguishing essential hypertension from the ''curable"
causes.

The problem of narrowing down a diagnosis from among many hypotheses
can be approached by using discrimination procedures over a taxonomic
model [Pople, 1974]; by the selection of "relevant" contexts guided by a
structure of suggestion pointers [Kulikowski and Trigoboff, 1974], or by
the selection of "active'' frames [Gorry, et. al., 1974].

We propose to compare some of these approaches within the chronic
diseases and those diseases directly related to them. Patients seen by a
primary care group physician or nurse, and those coming from the automated _
health appraisal center will need to be evaluated by a general consultation
program to detect the presence of a chronic disease before the detailed
chronic disease program is called in. This provides a realistic setting
for the design and evaluation of a general consultation program with
capabilities of "focusing-in" on specific diseases.

Within each chronic disease we will study its detailed time-course,
its causal structure, the effect. of different therapies (as a function of
progression) and the periodicities of re-examination and re-assessment. This
is designed to recognize ''characteristic" profiles of progression, control

and management of the disease.
Page 10

We propose to investigate strategies of decision-making in the multi-
ple-model Situation. Several different measures for the weight of evidence
towards or against hypotheses have been proposed: direct probabilistic,
measures of belief, incremental probability ratios, causal weights and
discriminating heuristic scores.

In some cases a probability measure can be defined and used profitably
from accumulated statistics, but problems usually arise in considering
the aggregation of multiple findings, for which statistics are usually
incomplete, unreliable or unavailable. Doubts have also been raised about
the appropriateness of probability-based reasoning in evaluating hypotheses
of disease [Shortliffe, 1974]. We will investigate the consistency among the
different types of measures of evidence that have been proposed in an
attempt to determine the appropriate context for each.

One of the issues related to measures of evidence is the evaluation
and justification of the disease models when new cases are tested on the
system. We propose to design protocols for establishing the presence of
patho-physiological states within a causal network, for determining the
validity of the transition weights between states. Most of the programs we
plan to develop will require the interaction of the expert to advise on
changes in the model of disease, but we expect to move gradually towards
automatic adjustment of weights.

To obtain the greatest benefits from the distributed data system we
propose to investigate the sharing of control decisions about the types

of consultation programs to be called, the degree to which explanation or
Page 11

instruction is to be provided, the levels of summarization, details of
‘reporting, etc. At one extreme all control could be left in the hands of the
user to select the desired mix between local and central processing
capabilities. More realistically, however, the local node programs ought

to have built-in decisions about summarization and reporting procedures
suited to local needs, while the central node ought to have control over

the degree of resolution of explanations, the types and complexity of models

to be invoked, and the rights of different users to access the programs.

D. Distributed Data Bases in Clinical Health Care

The proposed research program will make use of several existing
computer facilities--the Straub Clinic and Hospital Automated Health
Appraisal Center PDP 11/40 computer; the University of Hawaii BCC 500
computer, which is a large time-sharing, virtual memory system capable of _
supporting up to 100 simultaneous users, and with a disk storage capacity
of 3X10" bits, and the Stanford University SUMEX computer system which is
a DEC PDP 10 KI machine running the TENEX operating system. We propose to
acquire a separate PDP 11/40 system for the primary work of the chronic
disease clinics. The proposed system is shown in Fig. 1.

Why do we need to use four different computers and three geographically
separated sites? The reasons are: availability, size of primary and
secondary memory, transferability of data and programs, and ease of access.
We have already stated why we want to use SUMEX--to develop a set of chronic

disease consultative programs which can be a central resource to the NIH>
Page 12

and HEW community. On the input end, the Straub Health Appriasal Center
PDP 11/40 is an existing resource which provides a medical profile of the
patient which is the raw input data for the chronic disease programs. The
existing Straub PDP 11/40 is I/O bound and does not have any primary memory
Space available. Thus it is necessary for the chronic disease clinics to
acquire its own PDP 11/40.

The BCC 500 by virtue of its large secondary storage capacity will be
used for on-line storage of the patient data master file. We have considered
the possibility of using the BCC 500 for the chronic disease processing tasks.
But the principal reason for deciding against its use is that it is a one-
of-a-kind machine, and programs developed for it are not easily transferable.
The detailed break-down of the kinds of programs we intend to develop on the
PDP-11 local node is illustrated in Fig. 2.

Data acquisition protocols, summarization and reporting programs govern
the interaction with the user. These, together with the consultation
programs and intercomputer I/O handling programs are all subject to the
local control program. This program communicates with control programs at
the BCC 500 (to input and retrieve data), and at the SUMEX PDP-10 to proceed
into the more comprehensive programs. Some of these are illustrated in
Fig. 3, where it is indicated that the control program can transfer com-
mand to consultation, explanation and instruction programs accessing the
multiple and detailed chronic disease models. The explanation and instruc-

tion programs ought also to be accessed directly from the consultation

program.
Page 13

The ALOHANET (see Fig. 4) extends to the other islands of the Hawaiian
chain via ground radio links. It also operates on the NASA satellite ATS-1
which commmicates to the entire Pacific basin. The system that we propose
to develop has important implications for the delivery of health care in remote
areas such as the Hawaiian islands of Lanai and Molokai and to the U. S.

Trust Territories. Straub Clinic has a remote care center on Lanai and it
is envisioned that we will place one of the entry ports to the chronic
disease computer on Lanai,where nurse-paramedic care protocols are being

developed.

The idea of distributed programs and data accessible via computer
networks is one that is very relevant to the NIH commmity. Heretofore
the decision-making diagnosis programs have been developed at geographic-
ally distributed institutions and were not easily transferable. Now, with
SUMEX accessible from both TYMNET and ARPANET we can develop programs
which need not be transferred but can be accessed directly from anywhere
in the country. Costly duplications of effort can be avoided. Duplicate
computer resources can now be avoided. In most sites only a minicomputer
is needed for preprocessing and local data storage. Large computers can
be centralized and software resources can be pooled to develop extremely
sophisticated clinical aids which will be available to all the NIH commun-
ity, not just a select few. One of the principal goals of our research

project is to develop an operational model of remotely accessed chronic

disease clinics based upon the above premises of computer resource sharing.
Page 14

REFERENCES

Amarel, S. and Kulikowski, C.; 'Medical Decision-Making and Computer Model-
ing,'' Proc. Subconference on Computers in Biomedicine, Fifth Hawaii
International Conference on Systems Science, 1972, p. 173.

Gorry, G. A.; Sussman, G. J.; et. al.; "Description of a Frame Representa-
tion for Medical Knowledge," presented by G. J. Sussman at the Inter-
active Knowledge Systems Workshop, SRI, June 1974.

Kulikowski, C. A. and Weiss, S.; "Computer-Based Models of Glaucoma," Computers
in Biomedicine, Technical Report No. 5, Department of Computer Science,
Rutgers University, November 1971.

Pople, H.; "The DIALOG System," Seminar of the Rutgers Research Resource on
Computers in Biomedicine, May 1974.

Shortliffe, E. H.; Axline, S. G.; Buchanan, B. G.; Merrigan, T. C. and
Cohen, S. N.; "An Artificial Intelligence Program to Advise Physicians
Regarding Anti-microbial Therapy,'' Computers in Biomedical Research, 6,
pp. 544-560, 1973.

Shortliffe, E. H.; 'Mycin: A Rule Based Computer Program for Advising
Physicians Regarding Antimicrobial Therapy Selection," Ph. D. Dissertation,
Stanford University, October 1974.

Weiss, S. M.; "A System for Model-Based Computer-Aided Diagnosis and Therapy,"'
Technical Report No. 27 (thesis), Rutgers Research Resource on Computers
in Biomedicine, June 1974.
STRAUB CLINIC UNIVERSITY OF HAWAI//

 

 

 

 

PDP 11/40 PDP 11/40
. ALOHA PA
MULTIPHASICPe] CHRONIC fe 7710.4 5, wea] PANET Lest acc 500
DISEASE MENEHUNE TIP
SCREENING | |PROCESSING

 

 

 

 

 

 

 

 

 

 

 

 

 

 

f satellite
and
f A ground
lnk

 

 

Physical Lab Patient Follow-up
Exams Tests History Tests and t
Questions
ARPANET SUMEX
He — >
TIP PDP 10

 

 

 

 

 

 

STANFORD UNIVERSITY

Proposed Computer-Communications System For Chronic Disease Clinics

Figure |
 
    

AUTOMATED
HEALTH
TESTING

PDP 11/40

 
   
 

  

 

DATA ACQUISITION

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

SUMMARIZED
CHRONIC
DISEASES
MODELS

 

 

 

 

 

 

 

 

 

 

; PROTOCOLS
Patient
Input a
Data SUMMARIZATION |g _
Output -—&~ |_AND REPORTING
Date GENERAL
R
CONSULTATION [*_
CHRONIC DISEASES
CONSULTATION
_ HYPERTENSION Y
MULTIPLE lg¢——>| DIABETES
DISEASES le! Tivpoip
mS OTHER

 

 

Local Clinical Minicomputer Node

Figure 2

 

LOCAL
CONTROL
PROGRAM

 

 

 

 

I/O HANDLING

 

|

 

 

ENCOUNTER
FILE

 

 

(PDP 11/40)

PDP 10

BCC 500
POP 11740

CONTROL
PROGRAM

 
 
 
   

 

HYPERTENSION

 

 

 

 

 

 

 

 

 

 

 

 

 

MULTIPLE
DISEASES

 

\Al/

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

BCC 500

 

 

 

 

aL |

MULTIPLE
COMPREHENSIVE >| _ DIABETES
DISEASES
> THYR

MODELS NO a

OTHER

DIAGNOSIS

CONSULTATION (——»!__ PROGNOSIS

I iN THERAPY
EXPLANATION
INSTRUCTION

 

I/O HANDLING

 

 

Shared Resource Node (PDP IO / SUMEX)

Figure 3

 
KAUAI

      

ATS-1 ‘Satellite INTELSAT -1¥
(PACNET) (ARPANET)

     

HAWAIL

Legend:
-—-- projected
—— existing
Ed central station (MENEHUNE)
Mm, repeater
@ user node
€ computing resource

FIGURE 4 The ALOHANET