Information Technology Reference
In-Depth Information
the new patient. Otherwise, it would use difference links to suggest an alternate
classification to try. Classification is a task for which other AI approaches, includ-
ing neural networks and decision trees, are often used. Such techniques, however,
work best when there are more exemplar cases, fewer descriptive features per
case, and less missing data per case than PROTOS had to deal with. The case-
based approach was effective given the nature of the data in the audiological
disorders domain.
CASEY [23] diagnosed heart failure patients by comparing them to earlier
cardiac patients with known diagnoses. The features of a case were the measures
and states of the patient. Measures were quantitative clinical data, like heart rate,
while states represented qualitative features, like high or low arterial pressure, or
the presence or absence of a particular disease. A solution, or diagnosis, was the
cause of the patient's heart failure. It was represented as a graph, showing the
causality among the measures and states of the patient. In comparing past and
present cases, CASEY determined how the cases differed, and how significant
the differences were. For example, a new case might be missing a feature that
was present in the old case; however, if some other feature could cause the same
states as the missing feature, the cases could still match. A past diagnosis was
adapted to fit a new case by modifying the graph with the new patient's data
and causal relations.
CASEY was able to build upon and integrate an earlier model-based system,
which diagnosed heart failures based on a physiological model of the heart. When
it could not find a close enough match in its case base, CASEY would invoke the
earlier system to produce a diagnosis. CASEY was originally cited for producing
comparable diagnoses to the model-based system with an order of magnitude
speed-up. It has since been recognized for its early integration of two different
reasoning modalities. Today, applications of CBR in the Health Sciences are fre-
quently integrated, not only with other AI approaches, but with other computer
systems, such as health information management systems, and with laboratory
and medical equipment.
Diagnosis and classification are tasks that have since been the focus of CBR
systems in many different domains. The other types of tasks most commonly
performed by CBR systems in the health sciences are therapy planning and
education. An early therapy planning system was ICONS [24,25]. ICONS rec-
ommended antibiotic therapy for patients with bacterial infections in a hospital
intensive care unit (ICU). Had it been possible to accurately and quickly diag-
nose the cause of an infection, selecting a treatment plan would have been a
much easier task. However, it took two days for laboratory analysis to defini-
tively identify the responsible bacterial pathogen, and delaying therapy for so
long would endanger the patient.
Physicians would resort to calculated antibiotic therapy, in order to quickly
prescribe the treatment most likely to combat the infection while taking into
account any contraindications or side-effects. ICONS was built to provide deci-
sion support for physicians performing this task. Its case base contained experi-
ences with ICU patients who received calculated antibiotic therapy. When a new
Search WWH ::




Custom Search