Radial Basis Function Neural Network for Classification of Quantitative EEG in Patients with Advanced Chronic Renal Failure - Foundations on Natural and Artificial Computation - page 413

Table 2. qEEG extracted parameters used for training of ANN

Parameter Frequency range

MDF 1-26 Hz

MDF 4-12 Hz

Spectral entropy
Sample entropy
δ relative power

0.54 Hz

low δ relative power 0.52 Hz

high δ relative power 24 Hz

θ relative power 4-8 Hz

α relative power 8-12 Hz

β relative power 15-25 Hz

γ relative power 25-40 Hz

following frequency bands: δ (0.75-4 Hz), θ (4-8 Hz), α (8-12 Hz), β (15-25 Hz), γ
(25-40 Hz). Linear parameters derived from power spectral density (mean dom-
inant frequency (MDF) and spectral entropy) and non-linear (sample entropy)
were also calculated. MDF was calculated as the mean of frequency weighted by
power spectral density. Spectral entropy was calculated as the Shannon entropy

H ( X )= p ( x )log 2 p ( x ))

−

x∈X

of spectral density. Sample entropy, introduced in [9], measures the probability
that two subsequences similar in their first m points (with distance <r )are
still similar for m + 1. It was calculated using free downloadable subroutine
sampen ( N, m, r ): N is the number of points of the sample (length segment=1
second, 500 points), m =3, r =0 . 2[9].
In figure 2, we show the symmetric scatter plot matrix containing all pairwise
scatter plots for the calculated qEEG parameters and the corresponding values
of the pearson coecient. Each scatter plot was generated by pooling feature
values for all subjects. Perfect agreement between features i and j would show
up as a straight line in row i, column j. If there were no correlation between the
two features, the corresponding scatter plot would display a “cloud” of points.
MDF, spectral entropy and β power seem to be highly correlated. α power has
a high negative correlation with δ power. On the other hand, sample entropy,
appears to be less correlated to any other parameter. So, the chosen subset of
parameters seems to explore orthogonal aspects of signal.

2.4 Artificial Neural Network Analysis

After confirming the absence of significant differences between left and right
side, parameters obtained for each of the 19 EEG derivations were pooled in five
groups grouped by antero-posterior situation of electrodes: frontal (F3-Fz-F4),
central (C3-Cz-C4), parietal (P3-Pz-P4), temporal (T3-T4) and occipital (O1-
Oz-O2). 11 characteristic parameters were calculated for the recording in each