This application of SRI's evidential
reasoning technology highlights its ability to simultaneously
reason from any subset of some 60 different inputs, of varying reliability
and relevance, to automatically diagnose a patient complaining of chest
When a patient complaining of chest pains reports to a physician, a large
number of possibly relevant inputs must be considered during the diagnosis.
Some of these inputs are subjective, being observations made by the
attending physician (e.g., What is the patient's skin color?) or the
patient's answers to the physician's questions (e.g., Where does the pain
originate?). Other inputs are the objective results of laboratory tests or
measurements (e.g., What is the patient's temperature?). We have identified
some 56 different inputs that are relevant to selecting among 12 possible
pathologies. The relative diagnostic significance of these inputs depends
on both the degree of certainty with which the inputs are known (if at all)
and their native strength as discriminators among the pathologies under
The first step in constructing this application was to conduct a literature
search for studies that reported on the diagnostic relevance of various
indicators for those pathologies associated with chest pains. For some
indicators, the appropriate clinical data was available to statistically
estimate conditional probabilities; for others, expert estimates of these
conditionals had to suffice; for still others, no reasonable basis for
estimating the conditionals was available, but the presence or absence of
the indicator was known to logically eliminate some pathologies. In all
cases, probability bounds were substituted for point probabilities where
they were not precisely known.
The next step was to establish Gister's gallery that
delimits the possible pathologies and indicator values. It consists of a
frame of discernment for the pathologies (lower left) and
one for each indicator, and a compatibility relation
connecting each indicator frame to the pathology frame. These relations
capture which pathologies are consistent with each indicator value.
Finally, we established a Gister analysis (lower right)
that records the probabilistic relationships, delimited in the gallery, and
the patient-specific data recorded by the attending physician. Once
recorded, the analysis functions as a dataflow model of the diagnostic
procedure; as data progresses through the analysis, evidential reasoning
operations calculate the relevance of each of the given indicators and
fuse them to arrive at a
consensus based on all of the data or restricted to particular subsets of
data (e.g., pertaining to the description of the pain). While entering the
available data, the physician also enters a credibility capturing the
physician's faith in the response. As more credible data is entered, the
diagnosis becomes more specific.
This application of evidential reasoning highlights its ability to reason
with data of varying reliability. Both subjective and objective
observations are combined with clinically validated data, expert opinion,
and logical possibilities. Each patient workup can be run even if
significant portions of the patient's history and physical exam are
missing. The physician fills out a multiple-choice questionnaire (upper
left), provides a credibility rating for each answer, and the system
determines the diagnosis. The physician can examine different places in the
analysis to see how various portions of the patient's workup may suggest
different diagnoses.This system has already been shown to provide accurate
diagnoses when run on actual patient cases with problems ranging from acute
myocardial infarction to anxiety and esophagitis.
John D. Lowrance,
Artificial Intelligence Center
Gister and Gister-CL are trademarks of SRI International.
Copyright © 1995 SRI International, 333 Ravenswood Ave., Menlo Park, CA 94025 USA.
All rights reserved.