Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 184

Fig.7.5. The network structure for the model of the occurrence of hypertension.

Using BP learning, the network was trained with a real medical database
containing 1024 patient records. In this investigation, we used a training set
consisting of 100 occurrence data and 100 no-occurrence data selected by random
sampling. Table 7.3 shows a few examples from the training set corresponding to
the occurrence of hypertension.
Medical information, such as laboratory tests, lifestyles, and chief complaint,
is often ambiguous, and it is difficult to distinguish between normal and
pathological values by a crisp threshold. In the database of the occurrence of
hypertension, each term was given two cutoff values, as shown in Table 7.4; these
provide gray zones between normal and pathological values.
The developed system had correctly classified 95.0% of all patient records.
We detected a neuron with large WD in the 2nd hidden layer related to blood
pressure. According to the neuron generation algorithm, this neuron was split into
two and the attributes of the old neuron were inherited by the new neurons as
shown in Fig. 7.6. After we trained the network again using a new network
structure, the newly developed system correctly diagnosed 96.7% of patient
records. In this work, the neuron annihilation process was not applied, as there
was no neuron to be annihilated.
As a result, a hidden neuron was added to the initial network structure. The
initial network structure was constructed to represent the knowledge structure of
(forbreaking)

Table 7.4. Cutoff values for the occurrence of hypertension.

Value 1 Value 2

Age 40 50

Obesity index 20 24

60 100

ˠ -GTP

Volume of consumed alcohol 7 15

Systolic blood pressure 85 90

Diastolic blood pressure 130 140