Information Technology Reference
In-Depth Information
or any other better suited statistical method (as described previously) may be used
to further sort out and select the best candidate features (significant energy per time,
band and electrode), in terms of their task discriminating power.
3.3 Results
The proposed methodology is tested on simulated data, where there exist well-
defined spatiotemporal differences in frequency content between the target and the
control tasks, as discussed in the following section. In addition, the performance
of the proposed approach, as well as its results on actual experimental dataset, is
discussed in [14].
3.3.1 Simulation Test
Two different tasks are simulated by two different groups of signals. The first group
(control task) consists of 10 simulated spatiotemporal signals, each one comprised
of five channels. The idea is to reflect 10 participants virtually registered with a
5-channel-EEG system each. All the channels of the control task are randomly gen-
erated quasi-white noise signals, approximately 9-s-long (500 Hz sampling rate -
4,608 samples). The second group (target task) comprises of three channels (chan-
nels 1, 3, 4) reflecting white noise and two channels (channels 2, 5) encoding
frequency-modulated signals mixed again with quasi-white noise. Channel 2 con-
sists of a time-varying theta EEG signal occurring at a fixed latency, linearly mod-
ulated (5-7 Hz) and varying in length randomly between 512 and 1,024 samples
among subjects, and a gamma EEG signal, linearly modulated (30-90 Hz) and vary-
ing in length randomly between 1,024 and 2,048 samples among subjects, all mixed
with quasi-white noise. In a similar manner, channel 5 consists of an alpha band lin-
early modulated signal (9-12 Hz) varying in length randomly (768-1,536 samples)
and a gamma linearly modulated signal (30-90 Hz) varying in length randomly be-
tween 512 and 1,024 samples, mixed with quasi-white noise. Quasi-white noise
covers the interval between the modulated signals. Such a generated signal (channel
2) together with the wavelet time-frequency representation is depicted in Fig. 3.4.
Theta and gamma bands are apparent at different latencies. The tabulated channels
in Table 3.2 are the significant ones extracted with the proposed approach from the
six (most widely studied) frequency bands (delta, theta, alpha, beta, gamma1, and
gamma2). The channels listed in the first column are the selected ones after the first
statistical test (Step 2), whereas the channels listed in the second column are the
refined ones after the second statistical selection (Step 4). Although the first stage
can identify both channels (2 and 5) with the pre-specified frequency content, it is
not able to discriminate correctly the activated frequency bands because of leakage
effects between bands, as illustrated in Fig. 3.4. In contrast, the second stage focuses
