You are on page 1of 12

KING SAUD UNIVERSITY

COLLEGE OF COMPUTER & INFORMATION SCIENCES


DEPARTMENT OF COMPUTER SCIENCE
Second Semester 1427/1428 AH

Medical Expert Systems


HDP and PUFF

Course:
CSC 562

Prepared By:
Amani Al-Ajlan
427220111

April - 2007

1
ABSTRACT
It seems that very early on, scientists and doctors alike were captivated by the potential such
a technology might have in medical. With intelligent computers able to store and process vast
stores of knowledge, the hope was that they would become as a perfect doctor, assisting or
surpassing clinicians with tasks like diagnosis. Medical artificial intelligence is primarily
concerned with the construction of AI programs that perform diagnosis and make therapy
recommendations.

INTRODUCTION
Expert system is a knowledge intensive computer program that captures the
expertise of a human in domains of knowledge. Medical Expert systems are the common
type of Artificial Intelligence in medical in routine clinical use. They contain medical
knowledge, usually about a very specifically defined task, and are able to reason with
data from individual patients to come up with reasoned conclusions.
Expert systems can be applied in various tasks of medical domains:

1. Diagnostic assistance. When a patient's case is complex, rare or the person making the
diagnosis is simply inexperienced, an expert system can help come up with likely diagnoses
based on patient data.

2. Expert laboratory information systems. Report receive from a laboratory was generated by a
computer system that has automatically interpreted the test results

3. Therapy critiquing and planning. Systems can either look for inconsistencies, errors and
omissions in an existing treatment plan, or can be used to formulate a treatment based upon a
patient's specific condition and accepted treatment guidelines.

4. Agents for information retrieval. Software 'agents' can be sent to search for and retrieve
information relevant to a particular problem. The agent contains knowledge about its user's
preferences and needs, and may also need to have medical knowledge to be able to assess the
importance and utility of what it finds.

5. Generating alerts and reminders. An expert system scan laboratory test results or drug orders
and send reminders or warnings through an e-mail system.

2
6. Image recognition and interpretation. Many medical images can now be automatically
interpreted.

The aim of this paper is to understand two medical expert systems. In section 1 I present
HDP, Heart Disease Program, which is a diagnostic assistance expert system; in section 2 I
present PUFF, Pulmonary Function System which is Expert laboratory information system; I
describe their knowledge acquisition, knowledge representation, how they act when uncertainty
arises and evaluation of them and I conclude my work in section 3.

Example of Medical Expert Systems:

1. Heart Disease Program (HDP):


The Heart Disease Program (HDP) is a large diagnostic program covering most areas of heart
disease. The physician can enter patient information about the history, physical examination, and
laboratory tests, and then the program generates detailed explanations of differential diagnoses
indicating the clinical data items which support each diagnosis.
Its' Purpose is to assist physicians in the diagnosis of patients with heart disease [7].

The Heart Disease Program (HDP) can be divided into 3 main components:
1. A user interface. Figure.1
2. The knowledge base and inference mechanisms.
3. Mechanisms to summarize and explain diagnoses Figure.2 and Figure.3

3
Figure.1

Figure.2

4
Figure.3

Knowledge Acquisition:
Knowledge acquisition is the most important aspect of the expert system development. It is
referred to as the process of transfer and transformation of expertise from knowledge sources to a
program. HDP gets knowledge from cardiologists and from books, also it develops the
knowledge base from clinical expertise incrementally by adding proved cases (After the diagnosis
had been done by HDP the cardiologists were asked to comment on which diagnostic elements
were most relevant, which were not relevant, and what was missing to update knowledge base).

Knowledge Representation:
Knowledge must be modeled or represented in a way that a computer can process. The
model of knowledge used by expert system is called knowledge base. The knowledge base of first
version of HDP was rule based; a program has large number of interconnected and nested IF-
THEN statements, or rules, that are the basis for knowledge in the system. However, it soon
became apparent that uncertainty plays large part in the diagnosis and management of heart
disease that the program had to deal with probabilities directly. For this reason, the knowledge
base of second version of HDP is implemented as a network of 200 nodes, analogous to a
Bayesian belief network, representing common potential heart diseases. The nodes are linked by
probabilities; the probability on the link may be fixed or dependent on patient data. However,
nodes in the HDP can represent different severity levels of diseases, and feedback loops are
permitted. Mechanisms to reason about the time course of symptoms and diseases are
incorporated [4].

5
The Algorithm for the Differential Diagnosis:

1. The facts about the patient are entered via a menu system. These facts are turned into
finding objects that become the terminal nodes for the diagnostic.
2. The probabilities of the paths are computed, using the patient information to adjust the
probabilities on the links.
3. The input is searched for definite consequences.
4. All of the findings from the input list that potentially reflect abnormal states are selected
to guide the hypothesis generation process.
5. For each of the abnormal findings that need causes, a search is made to find all of the
diagnostic nodes with causal paths to the finding.
6. The diagnostic nodes are ordered by the number of findings they account for. The top
node and any others that account for almost as many findings (almost is four fewer) are
used as a seeds for generating hypotheses.

7. From these, seeds are generated that can cover the findings.

For each of the cover sets, a hypothesis is built. The hypothesis starts with the cover set and the
known true nodes.

1. First a search is made for states that would be true no matter which causal paths were
chosen for the findings. These are added to the hypothesis.
2. The findings and unaccounted states are sorted in decreasing order of the difference in
probability.
3. For each finding the best path is found from any state in the accumulating hypothesis and
the states on that path are added to the hypothesis.
4. When the hypothesis is complete, the links among the nodes in the hypothesis are
checked to make sure there are no circularities in the hypothesis.
5. The hypothesis is then pruned of unnecessary causes and unnecessary paths that decrease
the probability.
6. The probability of the total hypothesis is determined by multiplying the probabilities of
each node in the hypothesis.
7. The hypotheses are ordered by their probability. The differential diagnosis list consists of
the best hypothesis and any others. [5]

6
The output summary contains complete hypotheses about the patient including all the
different possible diseases and physiological states. Each hypothesis includes explanations of how
the clinical data items justify the various diagnoses.

Uncertainty:
Medical diagnosis is a task that almost involves uncertainty, because it never has access to
whole truth about the environment. To solve this problem HDP use Bayesian belief network that
capture uncertain knowledge in natural and efficient way.
Bayesian belief network is a directed graph and has the following specification:
1. A set of variables makes up the nodes of the network.
2. A set of directed links or arrows connects pair of nodes.
3. Each node annotated with conditional probability information that quantifies the effect of the
parents on the node. [3]

Evaluation of the HDP:

The diagnoses used for comparison came from the hospital physicians at the New England
Medical Center [2]. Clinical data on 127 patients were used in comparison. The physicians'
diagnosis and the HDP diagnosis were compared to the cardiologist diagnosis. Then the
sensitivity was calculated by counting how many diagnoses in the second list are present in the
first. For example, if the comparison is between HDP and the cardiologist, and the cardiologist
has 6 diagnoses, 3 of which are present in the HDP diagnosis, then sensitivity is 3/6 = 0.5.

Sensitivity%
HDP & Physician 64
HDP alone 58
Physician 34

These results show that the Heart Disease Program had sensitivity significantly greater than
the physicians entering the cases did when compared to the cardiologist's diagnoses. Combining
the diagnoses of the HDP and the physicians gave a further benefit. This is likely to be a realistic
way of using the program's output, but intervention studies are required to confirm it.

7
Advantage of HDP:

HDP is a web based diagnostics system that available from any point on the Earth 24 hours
per day, so it spreads the expert knowledge of a few highly skilled advanced doctors leading
experts in their fields to a much broader medical staff.

Problem of HDP:

• Time required entering data, expecting physicians to enter a lot of clinical data like
physical examination, and laboratory tests. Survey of participating physicians indicates
that these difficulties stem from lack of familiarity with computer systems. For this
problem, I suggest instead of enter all these information extract data directly from patient
record in hospital information system.

2. Pulmonary Function System (PUFF):


Puff expert system diagnoses the results of pulmonary function tests. PUFF was initially
developed on the SUMEX computer, a large research machine at Stanford University, using
EMYSCIN and was later rewritten in a production version to run on the hospital's own
minicomputer [1]. Its task to interpret set of pulmonary function (PF) test results, like volume of
the lungs, the ability of the patient to move air into and out of the lungs, and the ability of the
lungs to get oxygen into the blood and carbon dioxide out, then it produces a set of interpretation
statements and a diagnosis for the patient.
Its' Purpose is to generate reports from a set of interpretation statements for pulmonary function
diagnosis and saving the staff a great deal of tedious work. The staff themselves would not be
displaced by this tool because their expertise still would be necessary to verify PUFF’s output, to
handle unexpected complex cases, and to correct interpretations that they felt were inaccurate
prior to printing the final report for physician signature and entry into the patient record.
Approximately 85% of the reports generated are accepted without modifications. PUFF’s
performance is good enough that it is used daily in clinical service. [1]

8
Knowledge Acquisition:
A knowledge engineer sat down with an expert pulmonary physiologist at the Pacific
Medical Center in San Francisco and developed rules, 100 cases (previously diagnosed patients)
were selected. The cases were chosen to span the variety of known disease states. The pulmonary
function expert posed hypothetical rules for diagnosing the illness and the knowledge engineer
encoded the rules (in LISP) and tested them with the test cases. The expert reviewed the test
results and modified or added rules to handle the cases that were incorrectly diagnosed. It
continued until the expert system was satisfied.

Knowledge Representation:
The knowledge base of PUFF was rule based; consists of a set of 64 production rules dealing
with the interpretation of pulmonary function tests (A more recent version had about 400 rules).

Example of a PUFF rule:

RULE31
IF:
1. The severity of obstructive airways
disease of the patient is greater than or equal
to mild, and
2. the degree of diffusion defect of the
patient is greater than or equal to mild, and
3. the TLC observed/predicted of the patient
is greater than or equal to 110, and
4. the observed/predicted difference in
RV/TLC of the patient is greater than or
equal to 10
THEN:
1. There is strongly suggestive evidence
(0.9) that the subtype of obstructive airways
disease is emphysema, and
2. It is definite (1.0) that "OAD, Diffusion
Defect, elevated TLC, and elevated RV
together indicate emphysema" is one of the
findings

Figure.4

The PUFF control structure is primarily a goal-directed, backward chaining of production


rules. The goal of the system at any time is to determine a value for a given clinical parameter. To
conclude a value for that clinical parameter, it tries a pre-computed list of rules whose actions

9
conclude values for the clinical parameter. If the rules fail to conclude a value for a parameter, a
question is then asked of the user in order to obtain that value.

Uncertainty:

In a rule-based approach, it was very obvious that there was a need to handle probabilistic
statements. PUFF uses EMYSCIN, so it deals uncertainty as the same way in EMYSCIN. It uses
probability theory to define the meaning of the values, which assign to each sentence a numerical
degree of belief between -1 and 1.

Evaluation of PUFF Expert System:

The diagnose used for comparison came from 144 cases at Pacific Medical Center [1]. The
144 cases was selected and interpreted by the system, then interpreted separately by two
pulmonary physiologists (the expert working in developing PUFF and a physician from a
different medical center).

Agreement on Diagnosis%
MD-1 and MD-2 92
MD-1 and PUFF 96
MD-2 and PUFF 89

The table shows that that the overall rate of agreement between the two physiologists
(MD-1 and MD-2 in the table) on the diagnoses of disease was 92%. The agreement between
PUFF and the physician who served as the expert to develop the PUFF knowledge base (MD-1 in
the table) was 96%. Finally, the agreement between PUFF and the physician who had no part in
the development of the PUFF knowledge base (MD-2) was 89%. So PUFF’s performance is good
enough that it is used daily in clinical service.

10
Advantages of PUFF:
1. Many medical expert systems had required large amounts of time for data input, PUFF
produced PF interpretations automatically without the necessity for user interaction. The
data needed to interpret patient status are available in a computer from the patient’s
history and from measurements taken in a laboratory. Other information is not required in
order to produce accurate diagnoses of pulmonary disease in the patient.
2. The amount of domain-specific knowledge involved in pulmonary function testing is
limited enough to make it feasible to acquire, understand, and represent that knowledge.
3. It is easy to use PUFF system and the clinical staffs in the PF lab were already receptive
to the use of computers within their clinical routines.

3. Conclusion:

In this report I review two medical expert systems HDP and PUFF. They assist physicians in
daily work of the diagnosis of patients and they spread the expert knowledge of a few highly
skilled advanced doctors leading experts in their fields to a much broader medical staff. I think
medical expert system provide a challenging domain in artificial intelligent and need more study
and research especially of it accuracy effect human life.

Acknowledgment:
In the beginning, I thank Allah for making me accomplish this report. Actually I enjoy prepare
this report. It gave me the opportunity to work in a professional and sophisticated manner. And I
give great thanks to Dr. Hassan for his cooperation.

11
Reference:
1. Aikens J S. PUFF: An expert system for interpretation of pulmonary function data.
Computers and Biomedical Research;16: 199 208.(1983)
2. Hamish S F Fraser, William J Long, Shapur Naimi: Differential Diagnoses of the Heart
Disease Program have better Sensitivity than Resident Physicians. Chute CG, ed. Proc
AMIA Annu Fall Symp;622 - 626.(1998)
3. Stuart Russell, and Peter Norvig . Artificial Intelligence a Modern Approach. 2nd ed.
Prentice Hall, 2003. 492-493.
4. William J. Long, Hamish S. F. Fraser, S. Naimi: Reasoning requirements for diagnosis of
heart disease. Artificial Intelligence in Medicine 10(1): 5-24 (1997)
5. William J. Long: Medical Diagnosis using a probabilistic causal network. Applied
Artificial Intelligence 3(2-3): 367-383 (1989)
6. < http://www.openclinical.org>.
7. <http://groups.csail.mit.edu/medg/projects/hdp>

12

You might also like