Database Reference
In-Depth Information
9. Discovery of Positive and Negative Rules
247
9.7 Rule Discovery as Knowledge Acquisition and
Decision Support
9.7.1 Expert System: RH
Another point of discovery of rules is automated knowledge acquisition from
databases. Knowledge acquisition is referred to as a bottleneck problem in
development of expert systems [9.2], which has not fully been solved and is
expected to be solved by induction of rules from databases. However, there are
few papers that discuss the evaluation of discovered rules from the viewpoint
of knowledge acquisition [9.12].
For this purpose, we have developed an expert system, called RH (rule-
based system for headache) using the acquired knowledge. The reason for
selecting the domain of headache is that earlier we developed an expert system
RHINOS (rule-based headache information organizing system), which makes
a differential diagnosis in headache [9.3]. In this system, it takes about six
months to acquire knowledge from domain experts. RH consists of two parts.
First, it requires inputs and applies exclusive and negative rules to select
candidates (focusing mechanism). Then, it requires additional inputs and
applies positive rules for differential diagnosis between selected candidates.
Finally, RH outputs diagnostic conclusions.
9.7.2 Evaluation of RH
RH was evaluated in clinical practice with respect to its classification ac-
curacy by using 930 patients who came to the outpatient clinic after the
development of this system. Experimental results about classification accu-
racy are shown in Table 9.5. The first and second rows show the performance
of rules obtained using PRIMROSE-REX2; the results in the first row are
derived using both positive and negative rules and those in the second row
are derived using only positive rules. The third and fourth rows show the
results derived using both positive and negative rules and those by positive
rules acquired directly from medical experts. These results show that the
combination of positive and negative rules outperforms positive rules and
gains almost the same performance as those by experts.
Table 9.5. Evaluation of RH (accuracy: averaged).
Method
Accuracy
PRIMEROSE-REX2(positive and negative)
91.4% (851/930)
PRIMEROSE-REX2(positive)
78.5% (729/930)
RHINOS (positive and negative)
93.5% (864/930)
RHINOS (positive)
82.8% (765/930)
 
Search WWH ::




Custom Search