Computational Intelligence Techniques for Classification in Microarray Analysis - Computational Intelligence in Healthcare: Advanced Methodologies - page 289

rate for the ESOINN network is determined as a function of the pre-classified cases.
The equivalence index is defined as indicated in (10):
{ }

E
j

P
j

# i ∈ I ' / i ∈ c ∧ i ∈ c

eq =

(10)

# I '

j through the ESOINN

E

j c represents the set of meshes belonging to class

Where

j through the

c represents the set of individuals belonging to class

P

network and

PAM algorithm.

4.2.4 Classification
Once the meshes are generated by the clustering process, previously unclassified
individuals are now classified by selecting the nearest mesh. When the mesh has been
selected, the case is assigned to the class of the mesh selected. The assignment is
defined as (11).

i ∈ G

E
k

→ i ∈ C

(11)

s u j

c

j

where i is the unclassified individual, i is the individual closest to the individual
i calculated using the Euclidean distance.
As shown in Figure 1, the reuse stage receives the filtered and not-filtered data re-
sulting from the retriever stage as inputs. The input is used for both the ESOINN
neural network and the PAM technique. The ESOINN neural network generates a set
of groups assigned to different classes. These groups are composed of meshes con-
taining different elements together with the information of the previous classification.
The PAM technique repeats the same project concurrently and generates the groups
for each of the classes. The groups generated by PAM contain the individuals and
their previous classification, but do not consider sub-groups. Finally, the equivalence
index is calculated and the new patient is classified. The error rate for the ESOINN
network is made through (8) to determine the class for each of the groups.

4.3 Knowledge Extraction

The knowledge extraction phase detects anomalous classifications, since it accounts
for the existence of probes with irrelevant information, or those that were decisive for
the misclassification. Sometimes, the existence of certain probes causes a classifica-
tion of patients based on erroneous criteria, such as the distinction between men and
women. Such a classification, without being wrong, is irrelevant to the problem,
which is why the probes that can cause these classifications are analyzed at this stage.
If the human expert notes that the probes contain irrelevant information, they are
marked as irrelevant and not taken into account in the next iteration of the CBR cycle.
The extraction of knowledge that is presented to the human expert is carried out us-
ing the RIPPER [43]. There are other alternatives for the generation of decision rules
which operate similar to the decision trees, including J48 [17] and PART [44]. These