Page 31 - BH-2-2

Brain & Heart Predictive modeling using electroencephalogram

movements of their left or right hand, foot, or tongue
in response to randomized cues. The sequence of the
indicators was random. This experiment comprised
multiple runs of 60 trials each. Each trial began with an
auditory stimulus presented at t = 5 s, followed by the
appearance of a cross “+” on the screen, which remained
visible throughout the trial. At t = 10 s, participants were
shown an arrow pointing left, right, up, or down for 2 s,
during which they were instructed to imagine moving their
left hand, right hand, tongue, or feet. Finally, at t = 12 s, the
cross vanished. This experimental paradigm formed the
basis for in-house data acquisition trials conducted with
the Emotiv EEG headset.
The stimulus presentation software included with
BCI2000 serves various purposes, primarily facilitating Figure 3. Electrode locations.
real-time data acquisition from any headset and the
optional integration of this data with feedback applications.
In this study, cues were generated with this stimulus
presentation software, and EEG data and event markers
were recorded in real-time to reflect their association
with the cues. Successful data collection requires running
a sufficient number of independent experiments. Fifteen
normal subjects have been used to collect data for this
study through an EEG headset. The BCI2000 stimulation
presentation program used visual cues to instruct
participants to perform a left motor imagery task, a right
motor imagery task, or nothing at all. This resulted in a
total of 360 trials, with 60 trials allocated to each of the
three activities (left, right, and rest).
Subjects were instructed to maintain a state of activity Figure 4. Frequency plot (Subject K).
during data collection, visualizing themselves opening or
closing the corresponding hand at a rate of about once per neighbors (KNN), principal component analysis (PCA),
second when the left or right arrow was on screen. The and dual-layer neural networks (NN), are selected based
oscillatory nature of the EEG signals has been observed by on their unique advantages and disadvantages. KNNs are a
researchers during motor imagery tasks. popular non-parametric algorithm that is easy to use, simple
to understand, resilient to noisy data, and appropriate for
Figure 4 displays the frequency distribution for subject K, tasks involving multiple classes. However, its computational
providing evidence that the analyzed signals are acquired requirements can be problematic, particularly when
during motor imagery tasks. Each individual exhibits a dealing with big datasets. PCA, on the other hand, offers
unique “sweet spot” in terms of the frequency at which a useful method for reducing dimensionality, which helps
their power output is maximized. In frequency plots for to improve computational efficiency and lessen the impact
Subject K, it is clear that there are not very many frequencies of dimensionality. However, it makes the assumption
with significant power. This is because the subject’s accuracy that the principal components with the highest variance
on motor imagery tasks is lower. Understanding the timing contain the most relevant information, which may limit
and frequency characteristics of these signals can be its application in capturing subtle aspects of EEG signals.
valuable, particularly concerning event-related potential Dual-layer NNs, on the other hand, are particularly good
latencies and the synchronization/desynchronization of at capturing complex nonlinear relationships in the data,
EEG signals after/during motor imagery tasks. making them an attractive option for tasks where complex
patterns call for sophisticated modeling. However, given
4.2. Predictive models used their computational demands, overfitting proneness, and
The BCI IV Competition-I dataset presents an EEG signal decreased interpretability, they might need to be carefully
classification task. Various ML models, including k-nearest considered, especially in scenarios where resources are

Volume 2 Issue 2 (2024) 7 doi: 10.36922/bh.2819