Digital Signal Processing Reference
In-Depth Information
Independent components
5
0
5
0
5
10
15
20
25
30
35
4
0
5
0
5
0
5
10
15
20
25
30
35
4
0
5
0
5
0
5
10
15
20
25
30
35
4
0
5
0
5
0
5
10
15
20
25
30
35
4
0
5
0
5
0
5
10
15
20
25
30
35
40
Diagnosis
1.5
1
0.5
0
0.5
0
5
10
15
20
25
30
35
40
Figure 6.19
ICA components using FastICA with symmetric approach and pow3-nonlinearity
after whitening and PCA dimension reduction to 5 dimensions. Below the
components, the diagnoses (0 or 1) of the patients are plotted for comparison. The
covariances of each signal with the diagnoses are
0
.
16, 0
.
27, 0
.
25, 0
.
04 and 0
.
43,
and visual comparison already confirms bad correspondence of one of the ICs with
the diagnosis signal.
−
holds true only with other values being small.
All in all, however, we note that the linear ICA model applied to
the given immunology data did not hold very well when trying to find
diagnosis patterns. Of course we did not have such nice linear models as
EEG data; altogether, not many medical models describing connections
of these immunology parameters have been found. Therefore we will try
to model the parameter-diagnosis relationship using supervised learning
in the next section.
Neural network learning
Having used the two unsupervised learning algorithms from above, we
now use supervised learning in order to approximate the parameter-
diagnosis function. We will show that the measured parameters are
Search WWH ::
Custom Search