Database Reference
In-Depth Information
Table 7.8.
Cutoff values.
GOT
GPT
LDH
GGT
BUN
MCV
MCH
TbiL
CRT
1
40
40
350
60
13
85
30
1.0
0.8
2
100
100
500
100
18
95
32.5
2.0
1.2
3
200
200
700
300
25
105
35
3.0
1.5
The diagnoses are largely dependent on the doctor's experience. Therefore, the
diagnostic rule is not necessarily represented by a single rule. We applied ADG to
the diagnosis of hepatobiliary disorders. The number of groups in ADG
corresponds to the number of extracted rules.
(
) (
) (
) (
)
Rule 1
(8 agents):
>
1
∧
=
0
∧
>
1
∧
>
1
GGT
BUN
MCV
MCH
(
) (
) (
) (
)
Rule 2
(8 agents):
=
∧
=
∧
>
∧
>
GOT
0
GPT
0
GGT
0
MCH
0
(
)
∧
CRTNN
>
0
(
) (
) (
) (
)
Rule 3
(8 agents):
LDH
=
0
∧
GGT
>
0
∧
MCV
>
1
∧
TBIL
=
0
(
)
∧
CRTNN
>
0
(
) (
) (
)
Rule 4
(5 agents):
LDH
=
0
∧
MCH
>
0
∧
CRTNN
=
0
(
) (
)
Rule 5
(4 agents):
LDH
=
0
∧
MCH
=
3
(
) (
) (
)
Rule 6
(4 agents):
GGT
=
3
∧
BUN
<
2
∧
MCV
=
0
(
) (
) (
) (
)
Rule 7
(4 agents):
GPT
Fig.7.8.
The acquired rules by ADG.
=
0
∧
LDH
=
0
∧
MCH
=
0
∧
CRTNN
=
2
We applied ADG to the diagnosis of alcoholic liver damage. As a result, 50
agents in the best individual are divided into 12 groups. Figure 7.8 shows a part of
the acquired rules that corresponds to tree structural programs in the best
individual. Rules are arranged according to the number of agents that support each
rule. A rule with more agents means a frequently adopted rule. A rule with fewer
agents means a rule for exceptional data.
In this section, we treat data containing multiple rules. We propose a new
method using ADG to extract multiple rules. As a result, we extracted some
effective rules for medical diagnosis.
7.4 Immune Multiagent Neural Networks
General NN learning aims to learn only good training cases, and it can realize high
classification. However, we meet real databases, such as medical databases, where
the records have noise or some contradictory cases. If we train the network with
noise or some contradictory cases, the classification capability of the network will
be reduced. Jacobs et al. proposed module nets, which can handle a subset of the
complete set of training cases. Reflective NNs, proposed by Ichimura, are a kind
of module nets [15].
The architecture of the proposed reflective NN is based on the network
module concept shown in Fig. 7.9. There are two kinds of network modules: an
Search WWH ::
Custom Search