Database Reference
In-Depth Information
If
VA
is zero for neuron
j
, this neuron does not generate any additional
information. In other words, this neuron does not perform any signal-processing
functions. Therefore, we eliminate the neuron whose
j
VA
is smaller than a
j
threshold value.
Criterion (B) can be identified by monitoring the dependence between the
output values of neurons in the network. If two neurons are totally dependent, that
is, given the output of one neuron the output of the other can be decided with
probability one, one of the neurons can be annihilated without affecting network
performance.
7.2.3 Application to a Medical Database
To verify the effectiveness of the KBANN with structure-level adaptation of the
NN, we develop a model of the occurrence of hypertension. In this chapter, the
prior knowledge is given by medical experts, who decided 16 input terms and 1
output term. There are two kinds of intermediate assumptions between the input
terms and the output term. Among the input terms, 10 terms are categorized as
biochemical tests related to the measurement of blood pressure for past five years;
the remaining terms are “gender,” “age,” “obesity index,” “
ˠ
-GTP,” and “volume
of consumed alcohol.” The output term represents whether the patient developed
hypertension during the five years. Such information is transferred into a network
using the concept illustrated in Table 7.3. Figure 7.5 shows the network structure
for the model of the occurrence of hypertension.
Table 7.3.
Example records for the occurrence of hypertension.
No.
Occur
Gender
Age
Obesity
index
Consumed
alcohol
ˠ
-GTP
1
+
m
48
25.86
54
13
2
+
f
45
25.56
13
0
3
-
f
47
21.05
10
0
4
-
m
43
26.61
17
0
5
+
m
49
28.73
40
0
No.
Systolic blood pressure
(one per year)
Diastolic blood pressure
(one per year)
1st
2nd
3rd
4th
5th
1st
2nd
3r
d
4th
5th
1
120
122
137
116
153
76
73
83
80
93
2
132
141
120
151
139
84
74
82
85
79
3
120
113
136
111
126
70
58
78
64
69
4
120
136
120
130
136
66
80
80
80
83
5
138
140
134
138
140
78
80
88
88
90
Search WWH ::
Custom Search