Pattern Recognition Techniques - Biomedical Signal Analysis: Contemporary Methods and Applications - page 210

14

12

10

8

6

4

2

0

0 2 4 6 8 10 12 14

Figure 6.20
PCA Dimension reduction versus perceptron reconstruction error.

values for the original network w ,w 0 , as shown in figure 6.21.
The single-layer neural network was trained with our measured
immunological parameters to reveal the diagnoses of our patients. Since
the bearing and the interdependencies of our measured parameters are
not fully understood, it is dicult to ascribe importance to certain
parameters. Six measured parameters were found to be essential for
the ANN learning process to assign the diagnosis CB or ILD to the
individual data samples.
A point of interest is the distances of the patient samples from the
ANN separation boundary line (figure 6.21). The ANN showed three
outliers in the assignment of the samples to the diagnoses CB and ILD,
leading to wrong diagnosis assignments. Under these three outliers, two
turned out to be CB patients with bronchial asthma, representing a
distinct subgroup of the CB patient group. Two CB patients had the
greatest distance to the separation boundary; those were identified as
patients with a severe clinical course of CB. Similarly, three patients
with ILD showed a distinct separation distance. These patients were
identified as those with a severe course of the disease. Thus the ANN
showed a graduated discrimination specificity for the diagnoses CB and
ILD.