Page 28 - BH-2-2

Brain & Heart Predictive modeling using electroencephalogram

environments. In general, this issue can be fixed by effective phase to accurately classify the training signals into the
pre-processing techniques. With potential applications in appropriate classes if the EEG features retrieved from the
14
the diagnosis and treatment of sleep disorders, the study data are not pertinent to the associated neurophysiological
offers a promising method for enhancing the classification event. It is advisable to use an effective feature extraction
of sleep states using deep learning models and sophisticated technique to enhance the speed and efficiency of the BCI
signal processing techniques. 15 rather than directly applying classification steps to the
raw signals, which could produce results but would take
In EEG research, segmenting signals into their
component frequency bands is crucial because each band time. The EEG data can be subjected to various simple
is associated with a different type of cognitive process. measurements to glean the necessary information, and
To isolate the signal components related to the studied transformations can be applied to the signal to provide a
physiological action, temporal filters such as low-pass and new perspective. Machine learning (ML) finds applications
band-pass filters are used. In the course of signal processing, in diverse fields, such as spam filtering, computer vision,
the alpha and beta rhythms are typically extracted for tasks and weather forecasting. Effective feature extraction
involving motor imagery. Temporal filtering is facilitated separates the extracted features into discrete classes from
training datasets, enabling the categorization of test signals
by tools such as the discrete Fourier transform and filters according to the various physiological tasks performed. 19
based on the finite impulse response (FIR). Spatial filters
serve a similar purpose, removing unwanted information Most ML systems force you to make a tradeoff between
from EEG signals. However, to obtain a comprehensive bias and variance due to the inverse relationship between
understanding of the physiological task, recording data the two. A large proportion of stable classifiers have low
from a small number of electrodes in close proximity variance and high bias, while the unstable ones have high
is necessary. Taking into account a larger number of variance and low bias. Given this fact, straightforward
nodes could lead to complications such as redundancies, models may outperform their complex counterparts. The
channel correlations, features, and increased training data experimental results must be validated by addressing bias
requirements. Noise-free copies of the original signals can and variance issues, which requires an understanding of
be easily generated using spatial filters by defining linear the significance of cross-validation. Model selection and
combinations of the original signals. 16 performance estimation inspire validation methods. The
success of pattern recognition and classification methods
EEG signal processing and BCI implementations have 20,21
greatly benefited from the application of independent typically relies on one or more ad hoc parameters.
Table 1 below compares and contrasts various ML
component analysis (ICA). ICA separates the information approaches in terms of their respective levels of difficulty.
coming from various parts of the brain, increasing the
likelihood that only signals from the desired areas will Following the model selection process, the model’s
be retained, while components likely to be artifacts or performance can be estimated by calculating its true error
noise can be discarded. Subsequently, EEG signals can be rate, which is the classifier’s error rate across the entire
reconstructed based on the chosen features. One of the dataset. Optimizing the system involves raising the model’s
most popular spatial filtering methods in BCI studies is complexity and introducing some variance error while
common spatial patterns. This technique entails mapping simultaneously lowering the bias error to lower the training
17
EEG readings onto spatial patterns, with the patterns error. However, there is no guarantee that adding new test
8,22
selected to increase the dissimilarity between the studied data will improve the learned model’s performance.
groups, aiding in data categorization. Dimension reduction In this setting, access to validation data, in addition to
and appropriate techniques for identifying differences in traditional training and test data, is crucial. Model selection
signals belonging to different classes are crucial, as different
physiological actions can produce distinct signal patterns. Table 1. Comparison of machine learning methods
In some cases, it is not possible to observe the differences Method Difficulties
simply by inspecting or using classifying techniques on the
original signals. Feature extraction is frequently employed Nonlinear regression Levels of complexity in polynomials
to specify an interesting signal and illustrate the similarities Decision trees The number of levels can be customized
and differences between signals belonging to the same class Regularized models Modifying the regularization parameter
and those belonging to different classes. 18 in various ways
K-nearest neighbors Different choice of k
A BCI’s classification phase is critical, with feature
extraction and identification being two of its most Support vector machine Different choices of hyperparameter
crucial components. The BCI will struggle during the test Kernel-based methods Different choices of kernels

Volume 2 Issue 2 (2024) 4 doi: 10.36922/bh.2819