Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 207

We extracted If-Then rules from the trained IMANN of the ICU database.
The output signal in the ICU database was “dead” or “alive” so that each subgroup
in Fig. 7.23 outputted “dead” or “alive.” The If-Then rules extracted from each
subgroup were associated with the corresponding output signal. For example, the
seventh subgroup outputted “dead.” The If-Then rules extracted from the seventh
subgroup by the search algorithm in Fig. 7.23 are:
z If “Ventilator” is true, then “dead” is very true.
z If “Urgent operation” is rather true, then “dead” is very true.
z If “Urgent operation” is rather true and “Ventilator” is true, then “dead” is
very true.
7.5 Conclusion and Discussion
In this chapter, we described three methods of extracting If-Then rules from neural
networks: KBANN with SLA, ADG, and IMANN. KBANN represents the
knowledge structure of experts in a network structure. ADG combines multiagent
system and GP techniques and extracts If-Then rules by generating cooperative
behaviors of agents. IMANN combines the ideas of multiagent system and neural
networks and makes two-stage classification by PLNN and DNN. All of the three
methods were applied to medical databases, and they extracted If-Then rules from
them successfully.
However, we may face the following dilemma. After giving all possible
patterns of input vectors, the search algorithm may be able to extract not only
explicit knowledge, but also new unknown knowledge from the network. However,
the extracted rules may be contaminated by some meaningless rules. On the other
hand, if we use an expert's explicit knowledge to prevent such contamination, we
will acquire only ordinary knowledge from databases.
To develop effective data-mining methods for medical databases, we should
begin by finding explicit knowledge from the network. After that, we should try to
develop new techniques for extracting new unknown knowledge from databases.
Our proposed methods will help develop effective data-mining methods for
medical databases.
Acknowledgment
This research was performed with partial support of a Hiroshima City University
Grant for Special Academic Research (General Studies) and supported by a
Grant-in-Aid for Scientific Research (Grant-in-Aid for Scientific Research (B)
13470099, Grant-in-Aid for Exploratory Research 14657106, and Grant-in-Aid for
Young Scientists 15790306) from the Japanese Ministry of Education, Culture,
Sports, Science, and Technology and the Japanese Society for the Promotion of
Science.