Elevating Neuro-Linguistic Decoding: Deepening Neural-Device Interaction with RNN-GRU for Non-Invasive Language Decoding

Full text

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
312 | P a g e
www.ijacsa.thesai.org
Elevating Neuro-Linguistic Decoding: Deepening
Neural-Device Interaction with RNN-GRU for Non-
Invasive Language Decoding
V Moses Jayakumar1*, Dr. R. Rajakumari2, Ms. Kuppala Padmini3, Dr. Sanjiv Rao Godla4,
Prof. Ts. Dr. Yousef A.Baker El-Ebiary5, Dr. Vijayalakshmi Ponnuswamy6
Department of English and Foreign Languages, Saveetha School of Engineering, SIMATS, Chennai, India1
Associate professor, Department of English and Foreign Languages, Saveetha School of Engineering, SIMATS, Chennai, India2
Assistant Professor, Computer Science and Science Department,
Sreenidhi Institute of Science and Technology, Hyderabad, Telangana3
Professor, Department of CSE (Artificial Intelligence & Machine Learning),
Aditya College of Engineering & Technology - Surampalem, Andhra Pradesh, India4
Faculty of Informatics and Computing, UniSZA University, Malaysia5
Professor, Department of Artificial Intelligence and Data Science, Koneru Lakshmiah Educational Foundation (KL Deemed to be
University), Green Fields, Vaddeswaram, Guntur District, Andhra Pradesh, India6
Abstract—Exploring innovative pathways for non-invasive
neural communication with language interfaces, this research
delves into the interdisciplinary realm of neurolinguistic
learning, merging neuroscience and machine learning. It
scrutinizes the intricacies of decoding neural patterns associated
with language comprehension. Leveraging advanced neural
network architectures, specifically Deep Recurrent Neural
Networks (RNN) and Gated Recurrent Units (GRU), the study
aims to amplify the landscape of neuro-device interaction. The
focus of Neurolinguistic Learning lies in extracting language-
related brain signals without resorting to invasive procedures.
Employing cutting-edge non-invasive methods and deep learning
techniques, the research aims to elevate the capabilities of neural
devices such as brain-machine interfaces and neuroprosthetics. A
distinctive approach involves crafting a sophisticated Deep RNN-
GRU model designed to capture intricate brain patterns linked to
language processing. This architectural innovation, implemented
in the Python software environment, harnesses the strengths of
RNNs and GRUs to enhance language decoding. The study's
outcomes hold promise for advancing non-invasive brain
language decoding systems, contributing to the expanding
knowledge base in neurolinguistic learning. The remarkable
accuracy of the proposed RNN-GRU model, boasting a 90%
accuracy rate, signifies its potential application in critical real-
world scenarios. This includes assistive technologies and brain-
machine interfaces where precise decoding of cerebral language
signals is paramount. The research underscores the efficacy of
deep learning methodologies in pushing the boundaries of
neurotechnology. Notably, the model outperforms established
techniques, surpassing alternatives like CSP-SVM and EEGNet
by an impressive 30.4% in accuracy. The model's proficiency in
deciphering topic words underscores its ability to extract
intricate language patterns from non-invasive brain inputs.
Keywords—Recurrent Neural Networks (RNN); Gated
Recurrent Units (GRU); neurolinguistic learning; neural devices;
brain machine interfaces
I. INTRODUCTION
Within the quickly developing field of neurotechnology,
the goal of creating a seamless interface between the human
brain and external devices has spurred innovative research
efforts [1]. Neuro technology is advancing by developing
neural-device interaction, an interdisciplinary field that
combines neuroscience and engineering to improve two-way
communication between neural systems and external devices,
aiming to create a seamless interface [2]. Addressing
fundamental issues and opening up new avenues for human-
machine interfaces are the driving forces behind the
advancement of neural-device interaction [3]. Conventional
means of communication between neural devices and the brain
frequently struggle with issues of signal integrity, bandwidth
of information, and procedure invasiveness [4]. It is becoming
increasingly important to overcome these obstacles as
technology develops in order to improve our comprehension
of neural processes and to use this knowledge for useful
applications that help people with neurological disorders,
disabilities, or those looking to enhance their cognitive
abilities.
The understanding of neural signaling’ s complexity and
the need for advanced models capable of real-time signal
interpretation and deciphering are at the core of this research
endeavor [5]. One promising approach is the use of deep
reinforcement learning networks (DNRNNs) and GRUs. The
dynamic information embedded in neural signals linked to
different cognitive functions can be decoded by these models,
which are excellent at capturing temporal dependencies and
sequential patterns [6]. Learning more about neural-device
interaction is important not only for academics and
researchers, but also for a wide range of applications in
human-computer interaction, rehabilitation, and healthcare [7].
More innovative assistive technologies, tailored therapeutic
interventions, and more successful neuroprosthetics can all be
made possible by improved neural-device interfaces [8].

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
313 | P a g e
www.ijacsa.thesai.org
Furthermore, these developments pave the way for
revolutionary discoveries in areas like brain-machine
interfaces, neuromodulation, and cognitive augmentation by
facilitating a more intuitive and natural interaction between
people and machines [9]. This research explores various
methods for data collection and neural network architecture
creation, emphasizing non-invasiveness. It describes a
workflow for deep RNN-GRU-based neurolinguistic learning
to improve neural-device interaction. The goal is to advance
brain functions and foster a new era of human-machine
cooperation [10].
A growing field identified as neurolinguistic learning has
emerged from the dynamic intersection of neuroscience and
artificial intelligence in an effort to understand the neural basis
of language [11]. In an effort to uncover the mysteries buried
in the neural code that underpins our capacity for language
comprehension and production, this research explores the
complex relationship between neural activity and language
processing. Neurolinguistic learning aims to directly access
the neural substrate of language, in contrast to traditional
linguistic analyses, which rely on external behavioral
measures. This approach provides a more nuanced and direct
understanding of the cognitive processes involved. The
realization that language, a distinguishing feature of human
cognition, is not limited to observable behaviors or linguistic
outputs is what spurred researchers to explore the field of
neurolinguistic learning [12]. Rather, it is firmly anchored in
the intricate and dynamic neural activity patterns that emerge
inside the brain.
Specifically, non-invasive neural language decoding is the
emphasis of this research, which is an important application of
neurolinguistic learning [13]. Using invasive techniques like
brain electrode implantation, the traditional methods for
deciphering neural language patterns are frequently applied.
Concerns about safety, ethics, and the need to create more
widely available technologies, however, drive the search for
non-invasive alternatives. Understanding neural language
processes can be gained without invasive procedures by using
non-invasive techniques like functional magnetic resonance
imaging (fMRI) and electroencephalography (EEG). This
work aims to apply deep learning models, namely Deep RNN
and GRU, to advance the state-of-the-art in non-invasive
neural language decoding. These architectures are especially
well-suited to modeling the dynamic nature of language
processing because they are good at capturing sequential
patterns and temporal dependencies. Through the use of these
sophisticated neural network architectures, the research hopes
to shed light on the complexities of neural language
representation and, as a result, improve our comprehension
and decoding skills for the ideas encoded in neural language
[3].
Non-invasive neural language decoding has potential to
revolutionize assistive technology, neurorehabilitation, and
communication technology. It can help people with
communication impairments, offer new perspectives on
cognitive processes, and create more user-friendly interfaces.
This project combines linguistics, artificial intelligence, and
neuroscience, transforming our understanding of language and
the human mind. [14].Improving the smooth connection
between neural devices and the complex processes of
language expression and comprehension is one of the main
issues in this field [1]. The need to overcome the drawbacks
of the invasive procedures that are typically used in neural
interface development is what drives this research [15]. Even
though they work well, invasive techniques like implanting
electrodes directly into the brain come with risks, such as
tissue damage and infections. As a result, the search for non-
invasive substitutes has taken center stage in the development
of neural-device interfaces [16]. This project specifically
focuses on leveraging advanced neural network architectures,
namely Deep RNN and GRU, to decode neural language
signals without resorting to invasive interventions.
Our main focus is on the field of neurolinguistic learning,
which investigates the complex connection between language
processing and brain activity. The complexity of language
patterns is a challenge for traditional neural interfaces because
of the difficulties in decoding the rich and dynamic
information contained in neural signals. In this work, the
author explore the potential of deep learning—more
especially, RNN-GRU models—to non-invasively decipher
the complex patterns related to language. RNN-GRU models
were specifically chosen because of their demonstrated ability
to handle sequential data and capture temporal dependencies.
These architectures offer a sophisticated understanding of how
neural signals encode linguistic information over time, making
them well-suited to simulate the dynamic nature of language
processing. Through these advanced neural network
architectures, we hope to open up new possibilities for neural
devices and usher in a new era of non-invasive neural
language decoding.
The practical applications of this research have
transformative potential and go beyond the domain of
neuroscience. A successful implementation could transform
augmentative communication technologies and make it
possible for people with disabilities or communication
disorders to express themselves with never-before-seen ease.
Furthermore, our method's non-invasiveness reduces related
health risks and encourages accessibility and broad
acceptance.
The key contributions of the article is,
 The work proposes a non-invasive method for
neurolinguistic learning that harvests language-related
brain signals without necessitating invasive procedures.
This is achieved by utilizing the most advanced deep
learning algorithms, specifically Deep RNN-GRU.
 The study increases the possibility of neuro-device
interaction by using complex neural network
architectures, notably Deep RNN and GRU. This
technology has significant promise for non-invasive
neuro-communication applications in both ethical and
helpful situations. The incorporation of these
topologies facilitates the capture of complex brain
patterns associated with language processing in the
creation of neurotechnological interfaces.
 The real contribution is the development and use of the
Deep RNN-GRU model, which is done using the

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
314 | P a g e
www.ijacsa.thesai.org
Python programming language. This well-designed
architecture plays to the strengths of RNNs and GRUs
while showcasing an advanced tool for improved
language decoding, encouraging transparency and
reproducibility within the scientific community.
 The work offers novel and analytical techniques for
deciphering language-related brain signals, which
significantly advances the rapidly expanding field of
neurolinguistic learning. The exceptional accuracy and
performance of the suggested RNN-GRU model
demonstrate its potential as a revolutionary tool in the
ongoing advancement of non-invasive neural language
decoding systems.
The remainder sections of the article includes related
works, problem statement, methodology and results in
Sections II, III, IV and V respectively. The paper is concluded
in Section VI.
II. RELATED WORKS
Dash et al. [17] proposed neural interpretation of speech
in amyotrophic lateral sclerosis. A motor neuron-related
illness identified as ALS can result in locked-in syndrome,
which is total paralysis with awareness. Through brain
computer interfaces, such as EEG spellers, which have a low
communication rate, these locked-in patients can converse.
Neural speech decoding paradigms that could lead to normal
communication rates have been the focus of recent research.
However, the focus of current neural decoding research is on
typical speakers, and it is unclear how far these findings can
be applied to a target population (such as those with ALS).
The study examined the decoding of spoken and imagined
phrases from non-invasive magnetic resonance imaging
signals of individuals with ALS using seven machine learning
decoders and multiple spectral characteristics (band-power of
neural signals: delta, theta, alpha, beta, and gamma frequency
ranges). The outcomes of the experiment showed that while
ALS patients' decoding performance is considerably higher
than chance, it is still lower than that of healthy individuals.
For five imagined phrases and five spoken phrases from ALS
patients, the best scores were 75% and 88%, respectively. As
far, this is the first instance of neural speech decoding for a
population with speech disorders. The disadvantage is that in
order to confirm the study's effectiveness, analysis involving a
greater number of individuals with more severe ALS and
multiple sessions are required. Moreover, improved
neurolinguistic comprehension of the imagining of speech
would facilitate the development of algorithms for improved
imagined speech decoding performance.
Cooney et al. [13] proposed an EEG-fNIRS bimodal deep
machine learning design for overt and imagining speech
decoding. Research on brain-computer interfaces is
increasingly utilizing various characteristics of multiple signal
modalities at the same time. The bimodal gathering
procedures that integrate the temporal and spatial resolutions
of electroencephalography and near-infrared spectroscopy
require new decoding techniques. Present an EEG-fNIRS
hybrid BCI that utilizes a unique bimodal in nature deep
neural network design consisting of two convolutional sub-
networks to decode both overt and imagined speech. Each
subnet's features are fused before being further extracted and
categorized. Classification accuracy using the hybrid approach
showed substantial gains on EEG used independently for
imagined speech (p = 0.02) and a tendency towards a
significance for overt speech .The classification accuracy was
46.31% and 34.29%. Bimodal decoding produced
significantly better results for both speech types when
compared to fNIRS .While stimulus affected overt and
imagined words in significantly different ways, deeper subnets
improved performance. The bimodal approach performed
significantly better than the unimodal results for several tasks.
The results imply that neural signal decoding could be
enhanced by multi-modal deep learning. With this novel
architecture, speech deciphering from bimodal in nature neural
signals can be enhanced.
Llanos et al. [18] proposed peripheral stimulation of
nerves without invasive procedures improves speech in adults
category learning. In animal models, vagus nerve stimulation
has been demonstrated to prime adult sensory-perceptual
systems towards plasticity. Accurate temporal integrating with
auditory stimuli can significantly improve the specificity of
auditory cortical representations. Here, the study investigated
whether adult speech category learning is improved by sub-
perceptual thresholds transcutaneous stimulation of the vagus
nerve in conjunction with non-native speech sounds. To
recognize non-native Mandarin tone categories, twenty-four
native English speakers received training. The tVNS was
matched with the tone groups that were either easier or harder
to learn for each of the two groups. While receiving no
stimulation, the control group used the same thresholding
process as the intervention groups. Our findings showed that
tVNS significantly improved learning and retention of
accurate stimulus-response associations for speech categories,
but only when stimulus was combined with categories that
were simpler to learn. This effect manifested quickly,
generalizing to new exemplars, and differed qualitatively from
the typical individual variability seen in hundreds of learners
completing the same task in the absence of stimulus. Before
and after training, electroencephalography recordings showed
no signs of tVNS-induced modifications to the sensation
representations of auditory stimuli. According to these
findings, paired-tVNS selectively improves both perception
and consolidation of memories of intuitively salient categories
by inducing a temporally exact neuromodulatory signal.
Feng et al. [19] proposed brain and language semantic
alignment: a curriculum contrastive approach for
electroencephalography-to-text generation. The tremendous
potential for brain-computer interfaces has led to a growing
interest in Electroencephalography-to-Text creation, which
attempts to produce natural text from EEG signals. But a
significant obstacle to this task is the striking difference
between the semantic-dependent representation of text and the
subject-dependent EEG representation. In order to address
this, the study develops a Curriculum Semantic-aware
Contrastive Learning approach that reduces the discrepancy
by effectively recalibrating the subject-dependent EEG
representation to the semantic-dependent equivalent. More
precisely, semantically similar EEG representations are pulled
together by our C-SCL, while dissimilar ones are pushed

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
315 | P a g e
www.ijacsa.thesai.org
apart. Furthermore, carefully utilize curriculum learning to
both craft and make the learning progressively meaningful
contrastive pairs in order to incorporate more meaningful
contrastive pairs. Numerous experiments on the ZuCo
benchmark, and our approach, when combined with various
models and architectures, achieve the new state-of-the-art
while demonstrating steady improvements through three types
of metrics. Additional research demonstrates not just its
advantages in low-resource and single-subject settings, but
also its strong generalizability in zero-shot scenarios.
Lee et al. [20] proposed deciphering language-specific
imagined speech neural correlation through EEG signals.
Degenerative diseases and brain lesions can cause devastating
speech impairments. For people with severe speech deficits,
the use of imaginary speech in brain-computer interfaces has
proven to be an urging hope for reestablishing speech
production nerve impulses. However, due to low signal-to-
noise ratio and high variation in both temporal and spatial
information, studies in the EEG-based simulated speech
domain still have some limitations. In this work, the author
examined the neural signals of two native speaker groups
performing two tasks in separate languages like English and
Chinese. The study postulated that the tonal and ideogram-
based Chinese language and the non-tonal and phonogram-
based English language would differ spectrally in how their
brains computed speech. The results showed that, in some
frequency band groups, Chinese and English had significantly
different corresponding power spectral densities. Furthermore,
native Chinese speakers in the theta band demonstrated
distinct spatial evaluation during the imagination task. In order
to decode the brainwaves of speech, this paper will therefore
propose the essential the spectral and spatial data of word
creativity with specialized language. The main flaw is that
while the experiment's imagination tasks were designed to
categorize words using machine learning algorithms, there
hasn't yet been any evaluation of the classification
performance.
Jensen et al. [21] proposed MVPA analysis of intertribal
phase coherence of neuromagnetic responses to words reliably
classifies multiple levels of language processing in the brain.
One of the least understood aspects of the human brain is
language's neural processing, yet a number of circumstances
call for an objective, participant-friendly, and noninvasive
assessment of the language function's neurocognitive state. A
brief task-free recording of MEG reactions to a series of
spoken language contrasts was suggested as a basis for a
solution to this problem. Spoken stimuli with differences in
lexicon, semantics, were used. The multivariate pattern
analysis to investigate intertribal phase coherence in five
canonical bands based on beam former source reconstruction
is utilized. By employing this method, effectively distinguish
between the brain responses to real words and pseudo words,
between proper and improper syntax, and between semantic
variations. The most accurate classification results showed
dispersed activity patterns that were augmented by other
regions while being dominated by the core temporofrontal
language circuits. The neurolinguistic properties varied across
frequency bands; broad γ was used to classify lexical
processes,  and  was used to classify semantic distinctions,
and low γ feature patterns were used to classify syntax.
Importantly, every kind of processing started almost
simultaneously 100 milliseconds after the auditory data made
it possible to distinguish between spoken and written input.
This demonstrates that distinct neuronal networks operating at
different frequency bands are involved in individual
neurolinguistic processes, which occur simultaneously. This
gives rise to even greater hope that neurolinguistic processes
in a variety of populations can be objectively and
noninvasively evaluated using brain imaging. The
disadvantage is that in order to determine whether this method
can be used to identify linguistic abnormalities in different
populations, it is necessary to fully comprehend the
relationship between time courses, frequency bands, neuronal
substrates, and neurolinguistic properties.
Time courses, frequency bands, neuronal substrates, and
neurolinguistic properties interact in a way that necessitates a
thorough comprehension of the approach being considered for
detecting linguistic abnormalities in different populations.
Although this has great potential, a major limitation is that the
classification performance of the word categorization tasks
created with machine learning algorithms is not evaluated.
Nevertheless, more recent studies demonstrate the method's
strong generalizability in zero-shot scenarios in addition to its
benefits in low-resource and single-subject settings. Notably,
despite subnets not being specifically designed for different
data types and suboptimal fNIRS data timing, the dual
network enhancement in the majority of subjects' results is a
promising result. However, for wider application, resolving
the method's drawbacks and carrying out a comprehensive
assessment of its overall performance are still essential.
III. PROBLEM STATEMENT
Despite considerable progress in the development of
neural-device interfaces, the seamless and efficient
communication between the human brain and external
technologies remains a formidable challenge. The limitations
of current approaches, particularly the invasive nature of many
brain interfaces, pose significant risks and hinder widespread
adoption. This study addresses this pressing issue by
proposing an advanced methodology employing Deep
Recurrent Neural Networks (RNN) and Gated Recurrent Units
(GRU) for neurolinguistic learning, aiming to provide non-
invasive alternatives. The primary objective is to decode
cerebral language signals in a non-intrusive manner,
representing a crucial initial step towards enhancing the safety
and usability of neural interfaces in applications such as
assistive technologies, neuroprosthetics, and brain-machine
communication. The existing landscape of non-invasive neural
language decoding struggles to capture the intricate sequential
patterns inherent in language processing. The complexity and
dynamism of language-related brain signals pose challenges
for conventional techniques. Consequently, the research
advocates for the integration of deep RNN-GRU architectures,
renowned for their proficiency in capturing sequential
dependencies, into the neurolinguistic learning framework.
The central challenge lies in designing and optimizing deep
learning models to advance our understanding of non-invasive
neural language decoding, thereby facilitating more effective
and user-friendly neuro-device interactions [21].

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
316 | P a g e
www.ijacsa.thesai.org
IV. PROPOSED DEEP RNN-GRU BASED NEUROLINGUISTIC
LEARNING
The methodology advances non-invasive communication
between brain devices and language interfaces by utilizing a
multidisciplinary approach based in neurolinguistic learning.
The study explores the complexities of deciphering language-
related brain patterns, with a focus on the interface between
neuroscience and machine learning. By utilizing cutting-edge
neural network topologies, particularly Deep RNN and GRU,
the study seeks to improve the capabilities of neuro-device
interaction. Because the process is non-invasive, there is no
need for intrusive procedures, which ensures both practical
and ethical viability. A Deep RNN-GRU model is
painstakingly built in Python to capture intricate brain patterns
related to language processing. The model's ability to decipher
complex language patterns, particularly for subject words,
indicates its potential for practical uses such as brain-machine
interfaces and assistive technologies. This represents a major
advancement in the integration of neurolinguistic learning and
neurotechnology. The proposed methodology is shown in Fig.
1.
A. Data Collection
Eleven healthy volunteers within the ages of 20 and 34
were recruited for this study, six of them were male and five
of them were female. Respondents were made aware of the
methods, frameworks, and goals before to the study. Every
participant provided written permission in accordance with the
Declaration of Helsinki, and all research methods were
approved by Korea University's Institutional Review Board.
Eight terms that represent the subject, verb, and object parts of
the sentence were selected for the experimental setting based
on their applicability to natural human-machine interaction,
particularly with neural mechanical arm control. The
fundamental language was made up of these words, which
included subjects like "I" and "partner," verbs like "move,"
"have," and "drink," and object terms like "box," "cup," and
"phone." Every phrase was said by those taking part 25 times,
and their audio cues were captured. Respondents wore 64-
channel EEG actiCaps during the EEG monitoring session,
and MATLAB 2020a software's BrainVision Recorder was
used to record EEG signals. Respondents in the study
completed speech imaging tasks for every single one of the
three sub sessions that focused on subject, verb, and object
terms, correspondingly. High signal quality was maintained
during the entire trial by providing students with pauses to
preserve their physical and mental health and by displaying
illustrations on a monitor [22].
One method for transforming EEG signals into a format
that is easier to analyze and understand is called spectrogram
embedding. EEG data, which show the brain's electrical
activity over time, are frequently intricate and provide
important insights into cognitive functions. By converting the
EEG signal into a spectrogram—a graphic depiction of the
signal's frequency content across time spectrogram embedding
is achieved. The first step in the procedure is to divide the
EEG signal into smaller temporal chunks, or epochs. By doing
this, the EEG signal is converted from the time domain to the
frequency domain, displaying the various frequency
components that are present. Following that, the data is
usually shown as a two-dimensional picture with time on one
axis and frequency on the other. The shading or color intensity
of the image indicates the amplitude of each frequency at a
certain moment in time.
B. Preprocessing using Bandpass Filter
The role of a Bandpass Filter is paramount in signal
processing, serving to selectively permit a specified range of
frequencies while attenuating frequencies outside this
designated band. This filter is instrumental in various
applications where isolating specific frequency components
from a signal is crucial. In fields such as telecommunications,
audio processing, and biomedical signal analysis, Bandpass
Filters help extract relevant information by allowing only the
desired frequency range to pass through. In the context of
communication systems, Bandpass Filters aid in frequency
division multiplexing, enabling multiple signals to coexist
without interference. Moreover, in biomedical applications,
Bandpass Filters are essential for isolating physiological
signals of interest, such as detecting heartbeats in an ECG.
Their versatility in isolating and enhancing specific frequency
components makes Bandpass Filters indispensable tools in
signal processing, facilitating accurate and targeted analysis
across diverse domains.
Fig. 1. Proposed methodology.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
317 | P a g e
www.ijacsa.thesai.org
Applying a Bandpass filter to EEG signals during
preprocessing is a crucial step in improving the specificity and
quality of brain information derived from the raw data. By
selectively allowing some frequencies and attenuating others,
the bandpass filter helps to separate the brain oscillations of
interest from possible noise and artefacts. One common option
for EEG data linked to language activities is to apply a
bandpass filter in a certain frequency range, this range has
been deliberately selected to include the brain frequencies
associated with cognitive functions such as language
comprehension and speaking. Unwanted elements, including
muscular artefacts or outside interference, are reduced by
using the bandpass filter, which makes it possible to analyze
the brain activity related to the experimental task more
narrowly. Bandpass filtering is important for EEG
preprocessing because it can increase the signal-to-noise ratio,
which guarantees that the underlying brain signals are more
accurately represented in the studies that follow. This specific
stage is critical for reliably extracting features for applications
requiring nuanced brain patterns, such as language decoding.
By helping to improve the overall quality of the EEG data,
bandpass filtering advances our knowledge of the brain
mechanisms underlying language and communication by
enabling more precise interpretations and insights into the
neural dynamics linked to cognitive activities.
C. Feature Extraction using Time Domain Analysis
Feature Extraction using Time Domain Analysis serves a
crucial role in the neuro-linguistic decoding framework
presented in this article. It involves the identification and
extraction of relevant features from temporal data patterns
associated with neural language signals. By delving into the
time domain, this technique enables the model to capture
subtle variations and temporal nuances inherent in the non-
invasive brain signals. This process enhances the
discriminative power of the features fed into the subsequent
RNN-GRU model, contributing to the accurate decoding of
complex linguistic patterns. Essentially, Feature Extraction
using Time Domain Analysis acts as a critical pre-processing
step, facilitating the comprehensive representation of temporal
information and thereby augmenting the overall effectiveness
of the neuro-linguistic decoding system proposed in the study.
Time-domain analysis feature extraction turns out to be a
crucial step in deciphering the temporal complexities of EEG
signals related to language processing, which is important in
the quest to advance neural-device interaction through deep
RNN-GRU based neurolinguistic learning for non-invasive
neural language decoding. Time-domain features provide a
way to describe the dynamic interaction between language
components and brain activity throughout the experimental
tasks. These features are produced directly from the timing
and amplitude information of EEG data. An important
temporal aspect of this research is the examination of Event-
Related Potentials (ERPs). ERPs are the mean brain responses
that are time-locked to certain events, such words being
presented in speech-imaging tasks. Researchers can learn
more about how the brain responds to language inputs over
time by extracting ERPs. The characteristics of ERP
components, such as their peak amplitudes, latencies, and
durations, offer a thorough description of the brain dynamics
connected to various language components. The Mean
Absolute Value (MAV) is given below,
MAV = 
 

 (1)
The length of the sample is denoted by M.
When analyzing EEG data, zero crossing is an essential
time-domain feature extraction technique, especially when
trying to comprehend the temporal dynamics of brain activity.
Finding the locations in the EEG signal where the amplitude
crosses the zero axis is the goal of this approach. Zero
crossing analysis offers important insights into the frequency
and pattern of oscillatory variations in the EEG signal,
providing information about the underlying brain processes
connected to language-related activities in the context of
neurolinguistic learning. Researchers can extract features that
describe the frequency of transitions between positive and
negative voltage values by measuring the number of times the
EEG waveform crosses zero within a certain time interval.
This characteristic is particularly relevant for identifying
rhythmic neural patterns and can enhance the effectiveness of
non-invasive neural language decoding techniques by
providing a thorough grasp of the temporal dynamics of brain
activity during language processing tasks.
{< 0 and  > 0} or { > 0 and  < 0} (2)
The consecutive samples are denoted as  and .
To better clarify the timing elements of brain responses
during language activities, the study may also concentrate on
temporal features including signal length, rise time, and fall
time. These behavioral characteristics add to our sophisticated
knowledge of the brain's real-time processing of language
data. Time-domain analysis is applied to both neural language
pattern decoding and deep RNN-GRU model training, where
it captures the sequential dependencies present in EEG data
related to non-invasive language decoding. This
methodological approach is in line with the overall objective
of improving neural-device interaction, which will aid in the
creation of more efficient and user-friendly brain-machine
interfaces for a range of applications in assistive technologies,
rehabilitation, and communication.
D. Deep RNN-GRU-based Neurolinguistic Learning for Non-
Invasive Neural Language Decoding
RNN-GRU plays an important role in sequential data
processing tasks, exhibiting distinct advantages in capturing
and understanding temporal dependencies within input
sequences. The GRU architecture, a variant of traditional
RNNs, introduces gating mechanisms that enable more
effective handling of long-range dependencies and mitigate
issues like vanishing gradients. This makes RNN-GRU
particularly well-suited for applications such as natural
language processing, time series analysis, and speech
recognition, where contextual information across different
time steps is crucial. The inherent ability of RNN-GRU to
selectively update and forget information, combined with its
parallel processing capabilities, enhances its efficiency in
modeling complex temporal patterns. These networks have
proven instrumental in tasks requiring nuanced understanding

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
318 | P a g e
www.ijacsa.thesai.org
of sequential data, making them a valuable asset in advancing
various fields.
The development of deep RNNs and GRUs has led to
major breakthroughs in neurolinguistic learning, a cutting-
edge discipline at the nexus of neuroscience and linguistics.
By non-invasively decoding cerebral language patterns, this
novel method seeks to open up new avenues for
comprehension of the complex interplay between language
processing and brain activity. Deep RNN-GRU models are an
advanced type of neural networks that are very useful for
language decoding tasks since they are made to collect and
analyze temporal connections in sequential input. Because of
the GRU's capacity to store and update information selectively
across long periods, the design makes it possible to represent
language-related brain signals' fluctuations in time in a
sophisticated manner.
The ability of deep RNN-GRU models to handle variable-
length sequences present in natural language is a significant
benefit in neurolinguistic learning. The network can learn
hierarchical characteristics of language representation, from
intricate syntactic patterns to subtle phonetic variations, thanks
to its hierarchical structure. It is ideally suited for deciphering
brain signals linked to different language processes because of
its versatility. These models are very useful for non-invasive
neural language decoding. Conventional approaches
frequently entail intrusive techniques like brain electrode
implantation, which restricts their application and raises
ethical questions. However, non-invasive neuroimaging data,
like electroencephalography (EEG), may be used to train deep
RNN-GRU models, making this method more generally
applicable and morally sound.
During the training phase, the model is exposed to
language stimuli while brain activity is being recorded. The
deep RNN-GRU continuously improves its capacity to
decipher language-related information from brain signals by
learning to associate particular patterns in the input data with
matching linguistic qualities. The model will get more and
more adept at capturing the complex links between brain
activity and language representation thanks to this iterative
learning process. Deep RNN-GRU-based neurolinguistic
learning has a wide range of significant applications. In
addition to basic studies on the neurological underpinnings of
language, this method has applications in therapeutic
situations. It may, for example, aid in the creation of assistive
technology for people with communication impairments or
function as a tool for tracking alterations in language-related
brain activity in response to treatment measures.
Even with the advancements, deep neurolinguistic learning
still faces several obstacles. Further work is needed to address
ethical issues with permission and privacy, interpretability of
learnt representations, and generalization of models across
different populations. Interdisciplinary cooperation among
neuroscientists, linguists, and machine learning specialists is
becoming more and more important as the field develops in
order to overcome these obstacles and realize the full potential
of deep RNN-GRU-based neurolinguistic learning.
Eq. (3) represents the hidden state update  at time t in the
RNN. Here, is the input at time  is the hidden state
from the previous time step,  is the weight matrix
associated with the hidden state, and tanh is the hyperbolic
tangent activation function. The tanh function introduces non-
linearity, allowing the network to capture complex
relationships and patterns in the data.
 = tanh ( + 󰇜 (3)
 =  (4)
Fig. 2. RNN-GRU architecture.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
319 | P a g e
www.ijacsa.thesai.org
Fig. 2 shows the architectural diagram of the RNN-GRU
model. The above equations form the basis of a GRU, a kind
of RNN architecture intended to effectively capture and
handle sequential data. The update gate  and reset gate,
which are both triggered by the sigmoid function σ, are
defined by Eq. (3). By deciding what to keep from the prior
hidden state   and the current input , these gates control
the flow of information. Eq. (4) uses the tanh function to
generate the candidate hidden state 
and integrates the reset
gate  to update the hidden state selectively. To provide a
seamless transition between the past and current states, Eq. (5)
finally combines the update gate with the former hidden
state and the candidate hidden state. All together, these
formulas describe the complex dynamics of a GRU, which
allows it to efficiently recognize and learn sequential patterns
in a variety of contexts, including natural language processing
and maybe neurolinguistic learning.
= σ ( + ) (5)
= σ ( + ) (6)

 = tanh ( + (* 󰇜) (7)
 = (1 - ) *  + * (8)
RNN-GRU Algorithm
Load and preprocess data // Bandpass filter
Feature Extraction // Time Domain Analysis
Define RNN-GRU model architecture
Split data into training and testing sets
Train the RNN-GRU model
Evaluate the model on the test set
Make predictions on new data
Visualize results
V. RESULTS AND DISCUSSION
With a foundation in neurolinguistic learning, the
methodology advances non-invasive communication between
language interfaces and brain devices through a
multidisciplinary approach. Situated at the nexus of
neuroscience and machine learning, the research delves into
the complexities involved in deciphering brain patterns linked
to language. The goal of the project is to improve neuro-
device interface capabilities by utilizing cutting-edge neural
network topologies, including Deep RNN and GRU. Because
the approach is non-invasive, it ensures both ethical and
practical feasibility by removing the need for intrusive
operations. A Deep RNN-GRU model that is carefully
designed to capture intricate brain patterns related to language
processing is created using Python. The model represents a
major advancement in the fusion of neurolinguistic learning
and neurotechnology because of its ability to decode complex
language patterns, particularly for subject words. This shows
the model's potential for use in assistive technologies and
brain-machine interfaces.
A. Model Loss
The model loss is a key metric of the model's performance
during training. It is commonly expressed as a mathematical
measure of the dissimilarity between expected and real neural
language patterns. When the loss trend is trending downward,
the model is doing a good job of reducing mistakes and
modifying its parameters to better suit the training set. A
steady decline in loss values across epochs indicates that the
non-invasive brain signals have been successfully learned to
recognize and adjust to. On the other hand, variations or
plateaus in the loss trajectory call for further examination and
may indicate that the model needs its hyper parameters
adjusted or that overfitting or underfitting occurred. Moreover,
comprehending the relationship between decoding accuracy
and loss offers a thorough grasp of the model's generalization
capabilities and clarifies how resilient it is when decoding a
variety of neural language patterns. It is depicted in Fig. 3.
Fig. 3. Model loss.
B. PVC Performance
A statistic called Percent Valid Correct (PVC)
performance is employed, especially in cognitive or
behavioral studies, to measure the precision and dependability
of a classification or prediction system. It shows the
proportion of accurate answers or forecasts among all valid
cases that were taken into account for a task or experiment.
This statistic only looks at how well the system performs
when a legitimate answer or forecast can be made; it ignores
incorrect or ambiguous data items. This statistic offers a more
focused evaluation of the system's effectiveness by
highlighting its accomplishments particularly in situations
where a significant answer or forecast is anticipated.
Fig. 4. PVC performance.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
320 | P a g e
www.ijacsa.thesai.org
A thorough assessment of the model's capacity to decipher
brain language patterns is provided by Fig. 4, which shows
PVC Performance across many linguistic aspects, namely
subject words, verb words, and object words on the x-axis.
The way that PVC performance is distributed among various
linguistic components provides information on how well the
model can identify and anticipate different sentence structure
components. Differences in the PVC performance of subject,
verb, and object words might be a sign of various brain
representations for these linguistic components or of varying
degrees of complexity. Understanding the model's complex
reactions to many aspects of language requires analyzing the
PVC performance across these categories. Doing so may
reveal brain activity patterns that alter according to
grammatical functions. Furthermore, it offers useful data for
adjusting the architecture and training strategies of the model
to improve decoding accuracy across various linguistic
components, which helps to improve neurolinguistic learning
techniques in non-invasive neural language decoding
paradigms.
C. Decoding Accuracy over Time
A statistic called decoding accuracy over time is used to
evaluate how well a neural decoding model performs and
changes over the course of an experiment or activity. This
statistic assesses how well the model can predict and
understand neural patterns linked to certain cognitive
processes or stimuli throughout time. The decoding accuracy's
dynamic nature over time offers valuable insights into the
model's flexibility and learning dynamics, demonstrating its
ability to grasp temporal variations in brain activity.
Researchers can identify patterns, trends, or fluctuations in the
model's performance by analyzing decoding accuracy at
various time intervals. This provides a thorough knowledge of
the model's ability to detect and adapt to temporal variations
in cognitive or language processing. This measure is
especially useful for research using time-series data, such
EEG signals, since it offers a detailed assessment of the
model's performance in real-time and its possible applications,
such as brain-machine interfaces and neurolinguistic learning.
Fig. 5. Decoding accuracy over time.
A more comprehensive illustration of how the model's
accuracy changes throughout the course of the task or
experiment is given in Fig. 5. Decoding accuracy trajectory
tracking over time can show learning, adaptation, or
stabilization tendencies in response to changing cognitive
demands. Accuracy peaks or troughs at particular times might
be related to different stages of the experiment, such when
stimuli are presented or when language tasks are performed.
Determining the model's sensitivity to temporal changes in
brain activity and maybe identifying crucial intervals for
optimal performance require an understanding of the
oscillations in decoding accuracy. Furthermore, this temporal
analysis provides useful insights for improving the model,
helping scientists adjust parameters or add adaptive techniques
to improve accuracy at critical times. In the end, this helps
develop more efficient and temporally-aware neural decoding
systems for use in neuroscience and brain-machine interfaces.
D. PVC Distribution across Different Word Types for Various
Methods
The pattern or spread of PVC performance across several
categories or classes of words within a given dataset is
referred to as the PVC distribution across various word kinds.
This metric measures the precision of a classification or
decoding system and evaluates its performance over a range of
linguistic aspects, especially in the context of neurolinguistic
learning or non-invasive brain language decoding. The
distribution analysis seeks to identify any differences in the
model's ability to decode various word kinds, including verb,
object, and subject terms. Gaining an understanding of the
PVC distribution allows one to assess the model's
performance in a more complex way by gaining insight into
how sensitive and flexible it is to different linguistic elements.
TABLE I. PVC DISTRIBUTION ACROSS DIFFERENT WORD TYPES FOR
VARIOUS METHODS
Methods
Subject Word
Verb Word
Object Word
CSP-SVM [23]
0.60
0.52
0.48
EEGNet [24]
0.78
0.56
0.53
Proposed RNN-GRU
0.90
0.72
0.70
Table I presents the decoding accuracy ratings for the
various techniques (CSP-SVM [23], EEGNet [24], and the
suggested RNN-GRU model) for various linguistic
components (verb, object, and subject words). Prominently,
the suggested RNN-GRU model outperforms the other
techniques in every category, with exceptional accuracy of
0.90 for subject words, 0.72 for verb words, and 0.70 for
object words. This shows that in terms of collecting and
interpreting neural patterns associated with various linguistic
components, the RNN-GRU architecture—which was created
for neurolinguistic learning in non-invasive neural language
decoding performs better than more conventional techniques
like CSP-SVM and EEGNet. The suggested RNN-GRU
model's effectiveness in comprehending and decoding
complex linguistic representations from non-invasive neural
signals is highlighted by the notable accuracy improvement,
especially in the decoding of subject words. This highlights
the model's potential to advance the fields of neural-device
interaction and neurolinguistic learning. It is depicted in Fig.
6.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
321 | P a g e
www.ijacsa.thesai.org
Fig. 6. PVC distribution across different word types for various methods.
E. Discussion
The study's findings, which are represented in the
decoding accuracy scores for various techniques across
subject, verb, and object words, offer important new
information on the effectiveness of applied neurolinguistic
learning strategies for non-invasive brain language decoding.
Remarkably, the suggested RNN-GRU model demonstrates
significant accuracy gains over conventional techniques like
CSP-SVM [23] and EEGNet [24], especially in the decoding
of topic words. This indicates how well the model is able to
represent and decipher intricate brain patterns linked to
various language components. The observed distribution of
PVC performance over various word kinds clarifies the
model's subtle competency and provides a thorough grasp of
its flexibility to various language processing components. The
area of neuro-device interaction has benefited greatly from
these discoveries, which highlight the promise of deep
learning techniques more especially, the suggested RNN-GRU
model in improving the precision and usability of non-
invasive neural language decoding systems.
The observed distribution of PVC performance over
various word kinds clarifies the model's subtle competency
and provides a thorough grasp of its flexibility to various
language processing components. The area of neuro-device
interaction has benefited greatly from these discoveries, which
highlight the promise of deep learning techniques more
especially, the suggested RNN-GRU model in improving the
precision and usability of non-invasive neural language
decoding systems. Overall, the results suggest potential
directions for applications in neurotechnology and human-
computer interaction, as well as advancing neurolinguistic
learning approaches and laying the groundwork for future
advancements in non-invasive cerebral language decoding.
VI. CONCLUSION AND FUTURE SCOPE
This research underscores the advancement possibilities in
non-invasive neural language decoding through the
application of a deep RNN-GRU-based neurolinguistic
learning technique, thereby augmenting the capabilities of
brain-device interfaces. The findings presented illustrate the
superior aptitude of the proposed RNN-GRU model in
capturing intricate linguistic nuances from non-invasive brain
signals, outperforming traditional methods like CSP-SVM and
EEGNet, particularly in decoding topic terms. The model's
adaptability to diverse linguistic components is evident in the
nuanced distribution of PVC performance across different
word types, emphasizing its potential to enhance the accuracy
and robustness of non-invasive neural language decoding
systems. The flexibility of the model to various linguistic
elements highlights its potential to improve the precision and
resilience of non-invasive neural language decoding systems.
For responsible implementation, it is imperative to handle
constraints including generalizability, interpretability, and
ethical issues. Neural patterns associated with language
comprehension can vary across individuals, languages, and
contexts. Thus, the model's performance might differ when
applied to different populations or languages.
In order to further increase decoding performance, future
research could concentrate on optimizing hyper parameters
and fine-tuning the model for the proposed RNN-GRU
architecture. Expanding the dataset to include more real-world
scenarios and language components might improve the
model's applicability and generalizability. Enhancing the
model for real-time decoding and dynamic language
processing tasks could increase its usefulness in applications
like assistive technology and brain-machine interfaces.
Furthermore, examining the interpretability of the model's
learnt representations may yield further insights into the
neurological underpinnings of language processing. It is still
essential for responsible implementation to address ethical
issues, such as participant privacy and the moral use of brain
data.
REFERENCES
[1] L. A. Jorgenson et al., ‘The BRAIN Initiative: developing technology to
catalyse neuroscience discovery’, Philosophical Transactions of the
Royal Society B: Biological Sciences, vol. 370, no. 1668, May 2015,
doi: 10.1098/rstb.2014.0164.
[2] J. Kumar et al., ‘Innovative Approaches and Therapies to Enhance
Neuroplasticity and Promote Recovery in Patients with Neurological
Disorders: A Narrative Review’, Cureus, vol. 15, no. 7, Jul. 2023, doi:
10.7759/cureus.41914.
[3] N. G. Hatsopoulos and J. P. Donoghue, ‘The Science of Neural Interface
Systems’, Annual review of neuroscience, vol. 32, p. 249, 2009, doi:
10.1146/annurev.neuro.051508.135241.
[4] U. Salahuddin and P.-X. Gao, ‘Signal Generation, Acquisition, and
Processing in Brain Machine Interfaces: A Unified Review’, Frontiers in
Neuroscience, vol. 15, 2021, doi: 10.3389/fnins.2021.728178.
[5] L. Zhao et al., ‘When Brain-inspired AI Meets AGI’. arXiv, Mar. 28,
2023. Accessed: Dec. 07, 2023. [Online]. Available:
http://arxiv.org/abs/2303.15935.
[6] C. Loriette, J. L. Amengual, and S. B. Hamed, ‘Beyond the brain-
computer interface: Decoding brain activity as a tool to understand
neuronal mechanisms subtending cognition and behavior’, Frontiers in
Neuroscience, vol. 16, 2022, doi: 10.3389/fnins.2022.811736.
[7] X. Li and Y. Xu, ‘Role of Human-Computer Interaction Healthcare
System in the Teaching of Physiology and Medicine’, Computational
Intelligence and Neuroscience, vol. 2022, 2022, doi:
10.1155/2022/5849736.
[8] D. O. Adewole et al., ‘The Evolution of Neuroprosthetic Interfaces’,
Critical reviews in biomedical engineering, vol. 44, no. 1–2, p. 123,
2016, doi: 10.1615/CritRevBiomedEng.2016017198.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 2, 2024
322 | P a g e
www.ijacsa.thesai.org
[9] M. J. Young, D. J. Lin, and L. R. Hochberg, ‘Brain-computer interfaces
in neurorecovery and neurorehabilitation’, Seminars in neurology, vol.
41, no. 2, p. 206, Apr. 2021, doi: 10.1055/s-0041-1725137.
[10] J. Iwry, D. B. Yaden, and A. B. Newberg, ‘Noninvasive Brain
Stimulation and Personal Identity: Ethical Considerations’, Frontiers in
Human Neuroscience, vol. 11, 2017, doi: 10.3389/fnhum.2017.00281.
[11] C. Surianarayanan, J. J. Lawrence, P. R. Chelliah, E. Prakash, and C.
Hewage, ‘Convergence of Artificial Intelligence and Neuroscience
towards the Diagnosis of Neurological Disorders—A Scoping Review’,
Sensors (Basel, Switzerland), vol. 23, no. 6, Mar. 2023, doi:
10.3390/s23063062.
[12] E. Rossi and M. T. Diaz, ‘How aging and bilingualism influence
language processing: theoretical and neural models’, Linguistic
approaches to bilingualism, vol. 6, no. 1–2, p. 9, 2016, doi:
10.1075/lab.14029.ros.
[13] C. Cooney, R. Folli, and D. Coyle, ‘A Bimodal Deep Learning
Architecture for EEG-fNIRS Decoding of Overt and Imagined Speech’,
IEEE Trans. Biomed. Eng., vol. 69, no. 6, pp. 1983–1994, Jun. 2022,
doi: 10.1109/TBME.2021.3132861.
[14] J. A. Chandler, K. I. V. der Loos, S. Boehnke, J. S. Beaudry, D. Z.
Buchman, and J. Illes, ‘Brain Computer Interfaces and Communication
Disabilities: Ethical, Legal, and Social Aspects of Decoding Speech
From the Brain’, Frontiers in Human Neuroscience, vol. 16, 2022, doi:
10.3389/fnhum.2022.841035.
[15] J. Peksa and D. Mamchur, ‘State-of-the-Art on Brain-Computer
Interface Technology’, Sensors, vol. 23, no. 13, Art. no. 13, Jan. 2023,
doi: 10.3390/s23136001.
[16] J. Lobo-Prat, P. N. Kooren, A. H. Stienen, J. L. Herder, B. F. Koopman,
and P. H. Veltink, ‘Non-invasive control interfaces for intention
detection in active movement-assistive devices’, Journal of
NeuroEngineering and Rehabilitation, vol. 11, 2014, doi: 10.1186/1743-
0003-11-168.
[17] D. Dash, P. Ferrari, A. Hernandez, D. Heitzman, S. G. Austin, and J.
Wang, ‘Neural Speech Decoding for Amyotrophic Lateral Sclerosis’, in
Interspeech 2020, ISCA, Oct. 2020, pp. 2782–2786. doi:
10.21437/Interspeech.2020-3071.
[18] F. Llanos, J. R. McHaney, W. L. Schuerman, H. G. Yi, M. K. Leonard,
and B. Chandrasekaran, ‘Non-invasive peripheral nerve stimulation
selectively enhances speech category learning in adults’, npj Sci. Learn.,
vol. 5, no. 1, p. 12, Aug. 2020, doi: 10.1038/s41539-020-0070-0.
[19] X. Feng, X. Feng, B. Qin, and T. Liu, ‘Aligning Semantic in Brain and
Language: A Curriculum Contrastive Method for
Electroencephalography-to-Text Generation’, IEEE Trans. Neural Syst.
Rehabil. Eng., vol. 31, pp. 3874–3883, 2023, doi:
10.1109/TNSRE.2023.3314642.
[20] L. Kw, L. Dh, K. Sj, and L. Sw, ‘Decoding Neural Correlation of
Language-Specific Imagined Speech using EEG Signals’, Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society. IEEE Engineering in Medicine and Biology Society.
Annual International Conference, vol. 2022, Jul. 2022, doi:
10.1109/EMBC48229.2022.9871721.
[21] M. Jensen, R. Hyder, and Y. Shtyrov, ‘MVPA Analysis of Intertrial
Phase Coherence of Neuromagnetic Responses to Words Reliably
Classifies Multiple Levels of Language Processing in the Brain’,
eNeuro, vol. 6, no. 4, p. ENEURO.0444-18.2019, Jul. 2019, doi:
10.1523/ENEURO.0444-18.2019.
[22] J.-H. Jeong et al., ‘Multimodal signal dataset for 11 intuitive movement
tasks from single upper extremity during multiple recording sessions’,
GigaScience, vol. 9, no. 10, p. giaa098, Oct. 2020, doi:
10.1093/gigascience/giaa098.
[23] ‘A feature extraction and classification algorithm for motor imagery
EEG signals based on decision tree and CSP-SVM’. Accessed: Dec. 11,
2023. [Online]. Available: https://www.spiedigitallibrary.org/conference
-proceedings-of-spie/11900/119002K/A-feature-extraction-and-
classification-algorithm-for-motor-imagery-
EEG/10.1117/12.2601842.short?SSO=1.
[24] ‘Q-EEGNet: an Energy-Efficient 8-bit Quantized Parallel EEGNet
Implementation for Edge Motor-Imagery Brain-Machine Interfaces |
IEEE Conference Publication | IEEE Xplore’. Accessed: Dec. 11, 2023.
[Online]. Available: https://ieeexplore.ieee.org/abstract/document/
9239647.