Fractional Hybrid Election Based Optimization with DRN for Brain
Thoughts to Text Conversion Using EEG Signal
Jyoti Prakash Botkar
1
, Virendra V. Shete
2
and Ramesh Y. Mali
3
1
Department of Electronics and Communication, School of Engineering, MIT ADT University, Rajbaugh Loni Kalbhor,
Solapur Highway, Pune - 412201, Maharashtra, India
2
School of Engineering, MIT ADT University, Rajbaugh Loni Kalbhor, Solapur Highway, Loni Kalbhor Railway Station,
Pune - 412201, Maharashtra, India
3
Department of Electric and Electronics Engineering, MIT School of Computing, MIT ADT University, Rajbaugh Loni
Kalbhor, Solapur Highway Railway Station, Pune - 412201, Maharashtra, India
Keywords: Electroencephalography, Brain Computer Interface, Maximum A Posteriori Probability, Deep Recedual
Netork, Gaussion Filter.
Abstract: A Brain-Computer Interface (BCI) system based on electroencephalography (EEG) permits users to
interconnect to contact the outer world using devices, like intelligent robots and wheelchairs by interpreting
their brain EEG signals. Translating brain dynamics into natural language is critical to BCI, which has seen
considerable development in recent years. This work proposes a novel method for brain thoughts-to-text
conversion utilizing EEG signals. At first, the Brain EEG signal is taken from a database, which contains EEG
signals related to tried imaginary thoughts of questions and statements. Afterward, signal preprocessing is
done by exploiting a Gaussian Filter to reduce noise in EEG signals. Consequently, signal segmentation is
done by the Maximum A posteriori probability (MAP) estimator. Later, feature extraction is done. After, word
recognition is accomplished using Deep Residual Network (DRN) tuned by the proposed optimization
technique called Fractional Hybrid Election-Based Optimization (FHEBO). Here, the FHEBO is developed
by the amalgamation of Hybrid Election-Based Optimization (HEBO) and Fractional Calculus (FC). Further,
language modelling is accomplished by utilizing the Gaussian Mixture Model (GMM). Moreover, the
proposed approach is observed to record maximal text conversion accuracy at 91.765%, precision at 92.765%,
F-measure at 94.241%, recall at 95.765%, and minimal error rate at 8.235%.
1 INTRODUCTION
Brain-Computer Interface (BCI) systems are progressed to
decode an individual’s intention, state of mind, and
emotions, by observing person’s brainwaves through
sensors placed externally or internally on the human brain
(Ullah & Halim, 2021). As a significant pathway across the
human brain and the outside world, BCI systems allow
users to interact or communicate with external devices like
service robots or wheelchairs by their brain signals (Zhang
et al., 2017) . In recent times, it is feasible to the availability
of specific states of the electromagnetic field and neurons
produced inside the brain. This is feasible due to the
availability of various modalities, such as functional
Magnetic Resonance Imaging (fMRI),
Magnetoencephalography (MEG), and EEG. Among these
technologies, EEG has benefits over others due to wireless
connectivity, low cost, portability, easy handling, and
portability (Kumar et al., 2018) . EEGs are a vital section
of all BCI systems and are used for recording brain signals.
The voltage fluctuations created through the movement of
ions inside neurons in the brain in response to certain
incitements are determined using EEG. There are two
techniques of
measuring voltage fluctuations utilizing
EEG sensors, such as non-invasive and invasive. Sensors
are surgically kept under the topmost part of the skull.
Therefore, these techniques are challenging to implement,
unsafe, and expensive, sensors are kept outside of the scalp
in the non-invasive BCI. This makes non-invasive
techniques straightforward, economical, handy, and safe
(Ullah & Halim, 2021).
BCI technology engages the communication of data
from one person’s brain to external devices utilizing a
wireless medium. A significant advantage of this
technology is that it is non-invasive with no complications
in utilizing bulky devices such as exoskeletons and comes
at a low cost (Rajesh et al., 2020). BCI receives electrical
signals and changes them into control commands such as a
biological communication channel without compromise in
a natural way (Junwei et al., 2019) . An EEG-based BCI
was used as a user identification method. They also have
many applications in the medical field. Emotion
852
Botkar, J. P., Shete, V. V. and Mali, R. Y.
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal.
DOI: 10.5220/0013606100004664
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 852-861
ISBN: 978-989-758-763-4
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
recognition has been applied in several domains such as
commerce, security, education, and others. Words
associated with the detected emotion are first modified
utilizing a correlation finder among emotion and words.
Next, sentence correctness is checked using a language
model depend on Long Short Term Memory (LSTM). The
LSTM networks are widely used in speech recognition,
time series prediction, grammar learning, handwriting
recognition, etc (Gupta et al., n.d.). Visual/
mental imagery
(VI/MI) is the processing of visual data from the memory
(rather than perceptual). Deep neural networks are utilized
for EEG related tasks which includes emotion recognition,
mental pressure detection and sleep analysis in addition to
the MI-EEG task classification (Ullah & Halim, 2021).
Deep Learning (DL) algorithms were introduced for
classifying EEG signals. A convolutional neural network
(CNN) has been effectively used in EEG-based BCIs for
classification and end-to-end feature extraction as well as
speech recognition and computer vision (Ahn & Lee, 2021).
Brain
thought to text conversion is effectuated
utilizing EEG signal in this work., Initially, Input Brain
EEG signal is taken from the dataset, which contains EEG
signals associated to attempted imaginary thoughts or
questions/statements. Afterward, signal preprocessing is
effectuated utilizing a Gaussian Filter to lessen noise in the
EEG signals. The signal segmentation is done with the help
of MAP estimator. Feature extraction is done based on
frequency-based features. Later, word recognition is carried
out using DRN trained using the proposed optimization
technique called FHEBO. The devised FHEBO is
developed by the amalgamation of FC and HEBO. The
HEBO is established by the integration of the Election
Optimization Algorithm (EBOA) and Hybrid Leader Based
Optimization (HLBO) .
2 LITERATURE REVIEW
(Ullah & Halim, 2021) designed Deep convolutional
neural network (DCNN) for brain thoughts to text
conversion. The technique recorded a superior
recognition for all the alphabets in English language
even with fewer parameters. Although, the
performance was affected by the interference from
other parameters. (Willett et al., 2021) introduced
Recurrent Neural Network (RNN) for brain-to-text
communication via handwriting. The method worked
successfully with data-limited regimes and unlabelled
neural sequences. However, the method failed to
improve longevity and performance. (Zhang et al.,
2017) proposed a Hybrid DL model based on a
convolutional recurrent neural network for brain
thoughts-to-text conversion. This approach exhibited
high adaptability and worked well in multi-class
scenarios, although the training time needed was
high.(Kumar et al., 2018) designed a Random Forest
(RF) classifier for brain thoughts-to-text conversion.
The technique was robust and had a high accuracy.
The method failed to be applied in several BCI
applications like rehabilitation systems and artificial
telepathy and rehabilitation systems.(Rajesh et al.,
2020) designed a Novel tiny symmetric algorithm for
brain thoughts-to-text conversion. The technique was
lightweight and enabled the transfer of information
from the patient's brain to the caretaker securely. The
method failed to decrease the size of the system to be
easily handled by patients.
2.1 Challenges
The major difficulties met by the various techniques
in brain thoughts to text conversion are listed as,
• The DCNN-based method presented in
(Ullah & Halim, 2021) was not extended to
recognizing individual words or generating
complete sentences.
• In (Zhang et al., 2017) , the hybrid DL
model failed to focus on enhancing
accuracy in a person-independent situation,
where few subjects participate in training
and the rest of the subjects participate in
testing.
• The RF method (Kumar et al., 2018) failed
to excerpt the robust features from EEG
signals to enhance the system's recognition
achievement.
The key challenges in developing BCI systems
based on EEG signals are that the techniques suffer
from high computational complexity and fluctuations
in EEG signals due to the environmental noise
limiting their performance.
3 PROPOSED FHEBO-DRN FOR
BRAIN THOUGHTS TO TEXT
CONVERSION
This work proposes a new method for brain thoughts-
to-text conversion utilizing an EEG signal. The
proposed technique is implemented as follows.
Initially, input brain EEG signals is taken from the
dataset, which contains EEG signals corresponding to
attempted imaginary thoughts of questions/
statements. Next, signal preprocessing is carried out
utilizing a Gaussian Filter (Kopparapu & Satish,
2011) to reduce noise in the EEG signals.
Consequently, signal segmentation is utilizing MAP
estimator (Popescu, 2021). After that, feature
extraction is done based on frequency-based features,
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal
853
like spectral spread, power spectral density, total
power ratio, spectral flux, and spectral centroid. Next,
word recognition is carried out utilizing DRN (Chen
et al., 2019) trained utilizing the proposed
optimization technique called FHEBO. Here, the
proposed FHEBO is developed by the combination of
FC (Bhaladhare & Jinwala, 2014) and HEBO. The
HEBO is developed by the combination of HLBO
(Dehghani & Trojovský, 2022), and EBOA
(Trojovský & Dehghani, 2022). Further, the
Language modelling for speech recognition is carried
out using the Gaussian Mixture Model (GMM) (Afy
et al., n.d.) . Figure 1 displays the structural diagram
of the FHEBO-DRN for brain thoughts-to-text
conversion.
3.1 Data acquisition
The input EEG signal exploited in this work is taken
from the dataset, and the dataset is described by
𝐺=
𝐾
1
,𝐾
2
,𝐾
3
⋯𝐾
,𝐾
ℎ
(1)
where 𝐺 is the dataset,
h
specifies the number of
EEG signals, 𝐾
denotes the
th
i
EEG considered for
processing.
3.2 Signal Preprocessing
The EEG signal 𝐾
is forwarded to the signal pre-
processing phase for removing the inherent noise in
the signal. The Gaussian filter [9] effectively smooths
the signal by reducing high-frequency noise.
Gaussian filter convolves the signal with a Gaussian
function, which has a bell-shaped curve. The
probability distribution function of the filter is
expressed as,
()
()
−
−
=
2
2
2
2
exp
2
1
,,
k
k
p
kkk
p
pZ
μ
σ
π
μσ
(2)
where, 𝜎
indicates the mean, 𝜇
2
symbolizes
variance, p is the time interval. The output of this
phase is the denoised signal depicted as 𝑀
.
Figure 1: A systematic view of the FHEBO-DRN for brain
thoughts to text conversion
3.3 Signal segmentation
The pre-processed signal 𝑀
is subjected to signal
segmentation (Popescu, 2021), which is
accomplished using MAP estimation (Popescu, 2021)
as it is easy to implement and has a low computational
complexity. The conceptual description of the
segmentation algorithm utilizing MAP estimation is
provided below.
The segmentation issue is resolved by finding
sequence 𝑚
=𝑚
1
,𝑚
2
,..,𝑚
that minimize the
optimal criteria of the form,
𝑚
=𝑎𝑟𝑔 𝑚𝑖𝑛
1,0..

𝐽
(
𝑚
)
(3)
where, 𝑚
represents the groups of parameter
vectors, noise scaling, and jump times, 𝐽indicates the
sum of squared residuals.
To determine all segments. a linear regression
model is used. For the measurements associated with
the 𝑗
ℎ
segment, 𝑒
1
+ 1,…𝑒

=𝑒
1

+ 1 , the
least squares assessment of its sample parameters and
covariance matrix is determined:
() ()
1
1
1
l
j
m
hh
h
hm
j
Hj e
F
θφ
−
−
=+
=
(4)
(
𝑗
)
=
∑
𝜙
ℎ
𝐹
ℎ
1
𝜙
ℎ
ℎ
1
1
(5)
wherein, 𝜙
ℎ
symbolizes the regressor, 𝐻, 𝑇is the
transpose, 𝐹
ℎ
indicates nominal covariance matrix of
INCOFT 2025 - International Conference on Futuristic Technology
854
the noise, and 𝑒
ℎ
is considered to be Gaussian and 𝑚
specifies the time index.
For optimal segmentation algorithm, the
following parameters are utilized:
()
()
()
()
()
1
1
1
j
j
T
m
TT
hh hhh
hm
J
jejFej
φθ φθ
−
−
=+
=− −
(6)
()
log det ( )Gj Hj=−
(7)
1
()
jj
M
jmm
−
=−
(8)
Here, 𝑀 signifies the number of data in each
segment and 𝐺denotes the logarithmic value of the
determinant of the covariance matrix. The values in
the
l
m
segmentation has a degree of freedom 𝑙−1
and needs 2
segmentations, which makes the
process extremely complex.
The MAP estimator is used to address this high
dimensional complexity by considering the
assumptions on noise scaling 𝜆
(
𝑗
)
, as follows:
,
:1
h
D
ata Signal e h M=
(9)
1: Analyse every segmentation, with the jump
times
l
m
and the number of jumps 𝑚, for all cases.
2: For all segmentation results, the ideal model for
all segments is calculated in the form of the
covariance matrices 𝐻
(
𝑗
)
and least square estimates
𝜃
(
𝑗
)
.
3: For all segment calculate:
𝐽(𝑗) =
∑
𝑒
ℎ
−𝜃
ℎ
𝜃
(
𝑗
)
𝐹
ℎ
1
ℎ
1
1
𝑒
ℎ
−
𝜙
ℎ
𝜃
(
𝑗
)
(10)
() ()
log detGj Hj=−
(11)
𝑀
(
𝑗
)
=𝑚
−𝑚
1
(12)
4: To determine 𝑚
, the MAP estimator is used
based on three constraints on noise scaling, 𝜆
(
𝑗
)
, with
()
01tt<<
the transformation probability at every
time instant.
(i)Known
()
j
λ
λ
=
0
,
() ()()
t
t
ljJjGm
l
j
lm
l
l
−
++=
=
∧
1
log2minarg
1
,
(13)
where,
()
10 << tt
denotes the change
probability at all instants.
ii) constant and unknown
()
=j
𝑚
∧
=𝑎𝑟𝑔𝑚𝑖𝑛
,
∑
𝐺
(
𝑗
)
1
+
(
𝑀𝑓 − 𝑙𝑏 − 2
)
×
𝑙𝑜𝑔
∑
(
)
4
1
+ 2𝑙𝑙𝑜𝑔
1
(14)
iii)unknown and changing
()
j
𝑚
∧
=𝑎𝑟𝑔𝑚𝑖𝑛
,
∑
𝐺
(
𝑗
)
1
+
(
𝑀
(
𝑗
)
−𝑙𝑏−
2
)
×𝑙𝑜𝑔
∑
(
)
(
)
4
1
+ 2𝑙𝑙𝑜𝑔
1
(15)
Results: Number 𝑙 and locations 𝑚
, 𝑚
=
𝑚
1
,𝑚
2
,..,𝑚
Only one of the equations in step 4 is utilized to
compute 𝑚
, based on the hypothesis of noise scaling.
The preprocessed signal can be split into various
segments as given below,
𝐿
=𝐿
1
,𝐿
2
,…𝐿
,𝐿
(16)
where, 𝜆 is commonly selected as an independent
function, 𝜆
(
𝑗
)
is noise scaling based on segmentation.
The segmented output produced by the MAP is
signified as
r
.
3.4 Feature Extraction
The segmented signal 𝑟is subjected to feature
extraction. Feature extraction is done based on
frequency-based features. The frequency-based
features like spectral flux, spectral centroid, spectral
spread, power spectral density, and total power ratio
are explained below,
i) Spectral flux
The spectral components (Mannepalli et al., 2017)
of the signal are extracted with the help of the spectral
flux. Spectral components are a significant feature
due to the spectral contents of the signal change over
time, and the efficiency of recognition decreases. The
equation of spectral flux is defined by,
𝑏
1
=
∑(|
𝑍
(
𝑐
)
|
−
|
𝑍
1
𝑐
|)
2
2
1
(17)
where, 𝐷 specifies the vector length,
𝑍
(
𝑐
)
specifies the spectral magnitude of 𝑐
ℎ
instant
and 𝑏
1
is the spectral flux.
ii) Spectral centroid
The Spectral Centroid (SC) (Hassan et al., 2016)
indicates the center of mass of the spectrum. The SC
is determined using the following expression,
𝑏
2
=
∑

(
)
1
∑
(
)
1
(18)
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal
855
where the amplitude value of
th
x
bin is specified
as 𝑆
(
𝑥
)
and 𝑏
2
represents the spectral centroid.
iii) Spectral spread
Spectral Spread (SS) (Hassan et al., 2016) is the
spread of the spectrum around its centroid (ie) it
measures the standard deviation of the spectral
distribution. The spectral spread is defined by,
𝑏
3
=
∑(

)
2
(
)
1
∑
(
)
1
(19)
where
3
b
indicates the spectral spread.
iv) Power Spectral Density (PSD)
PSD (Hong et al., 2018) specifies the strength of
a signal as a function of frequency. The
autocorrelation sequence for a specified data set is
initially assessed for nonparametric techniques. The
PSD is expressed as,
𝑏
4
=𝑌
=
1
∑
𝑍

𝑒
2
2
1
(20)
where, 𝑍

is the 𝑛
ℎ
sequence magnitude,
n
k
Y
designates the power of the
th
n
data sequence’s
th
k
frequency band, 𝐷 is the sample count and 𝑏
4
denotes
the PSD.
v)Tonal power ratio
This is utilized to determine the tonalness of a
speech signal (Mannepalli et al., 2017) . This is the
percentage of the tonal power of the spectrum
components to the whole power, and is defined by,
𝑏
5
=
(
)
∑|
(
,
)|
2
2
0
(21)
where, 𝑂
(
𝑎
)
represents the tonal power that is
calculated through totaling every bin
z
that lies
above a threshold and is local maximal,
𝑍
(
𝑧,𝑎
)
represents the preprocessed signal spectrum
and 𝑏
5
indicates tonal power ration. The feature
vector is denoted by 𝑣=
𝑏
1
,𝑏
2
,𝑏
3
,𝑏
4
,𝑏
5
.
3.5 Word recognition
Word recognition utilizing Deep learning is generally
the simplest approach to speech recognition. Here,
the input feature
v
is applied to the DRN(Chen et al.,
2019) for establishing word recognition. Further, the
training process of the DRN is effectuated using the
FHEBO algorithmic approach. Here, FHEBO is the
combination of FC and HEBO.
Figure 2: Architecture of DRN
3.5.1 Architecture of DRN
The DRN is a Deeper Neural Network (DNN) with
low gradient vanishing or explosion and higher
training speed. The DRN (Chen et al., 2019) was
originally designed for complicated image
classification processes, and it consists of fully
connected layers and pooling layers, 2-dimensional
convolution layers. A DRN structure contains various
layers, such as (i) Convolutional layers (ii) pooling
layers (iii) Convolutional layers (iv) Residual blocks
(v) Linear Classifier. These layers in DRN are
described below.
1)Convolutional Layer: A typical two-
dimensional convolutional layer can significantly
reduce the free parameters in the training procedure
and improve performance to the benefits of the
weight sharing and local receptive filed. The
following equation is used to model the process
established in the convolutional layer.
𝑐𝑜𝑛𝑣1𝑑
(
𝑣
)
=
∑
𝑊

1
0
∗𝑣 (22)
where,
v
is the input applied to the DRN,
𝑊specifies the learnable kernel matrix, 𝐵

is the
input feature dimension and
∗
is the cross-
correlation operator.
2)Pooling layer: The function of this layer is
generally applied to subsequent convolution layers,
and is mostly exploited for handling overfitting and
decreasing the spatial extent of the feature maps.
3) Activation function: A activation function that
is non-linear is utilized to higher the linearity of the
extracted features. ReLU alleviate the vanishing
gradient problem and significantly accelerates
convergence.
INCOFT 2025 - International Conference on Futuristic Technology
856
4)Batch Normalization: Batch normalization
minimizes internal covariate variation by scaling
input layers and smoothing implementations, thus
enhancing training speed and reliability, when
mitigating gradient bursting/fading and overfitting
issues with better learning rates.
5)Residual Blocks: The residual block had
shortcut connection from input to output. The input is
attached directly to the outputs, when output and
input are of equal dimension, otherwise a dimension-
matching factor is exploited.
6)Linear Classifier: This layer encompasses a
Softmax function and fully connected (FC) layer. The
following illustrates the output of this layer,
𝑞=𝑇
×
𝑣
×
+𝑢
×
(23)
where, 𝑇
×
specifies the weight matrix of 𝑃×
𝑈dimension, 𝑣
×
represents input feature map of
𝑈×𝑉dimension and 𝑢 is the bias of dimension
𝑃×𝑉. The output is designated as 𝑞 . Figure 2
illustrations the architecture of DRN.
3.5.2 Training of DRN with FHEBO
The DRN is structurally optimized by using the
FHEBO developed by the combination of FC
(Bhaladhare & Jinwala, 2014) and HEBO. The
HEBO is developed by the combination of HLBO
(Dehghani & Trojovský, 2022) and EBOA
(Trojovský & Dehghani, 2022). A novel
optimization algorithm called EBOA (Trojovský &
Dehghani, 2022) is created that imitates the voting
procedure to elect a leader. The basic impetus of
EBOA is the procedure of voting, electing a leader
and the cause of degree of awareness on electing a
leader. The EBOA people are led by a search space
directed by an elected leader. The HEBO is developed
based on the process of guiding a solution to optimal
one under the supervision of a hybrid leader. Instead
of considering a specific member for updating the
population, three members are taken into
consideration for updating the solutions, thus
avoiding convergence to local minima. FC
(Bhaladhare & Jinwala, 2014) plays a significant role
in increasing the efficiency of many methods like
curve fitting, filtering, modeling, edge detection and
pattern matching. The integration of FC in HEBO
enhances the convergence speed and assists in
attaining a global optimal solution.
i)Initialization
All members of the population specify a solution
to the issue in FHEBO. In the mathematical
perspective, the population is specified through a
matrix utilizing the following equation,
𝐼=
⎣
⎢
⎢
⎢
⎡
𝐼
1
⋮
𝐼
⋮
𝐼
⎦
⎥
⎥
⎥
⎤
×
=
⎣
⎢
⎢
⎢
⎡
𝑡
1,1
⋯𝑡
1,
⋯𝑡
1,
⋮⋱⋮⋰⋮
𝑡
,1
⋯𝑡
,
⋯𝑡
,
⋮⋰⋮⋱⋮
𝑡
,1
⋯𝑡
,
⋯𝑡
,
⎦
⎥
⎥
⎥
⎤
×
(24)
The primary location of member is estimated
randomly as follows.
𝑡
,
=𝑏𝑝
+𝑛.𝑗𝑝
−𝑏𝑝
,𝑎 = 1,2 …,𝐸,𝑔=
1,2,…,𝑡, (25)
where, 𝑛 is a random number ranging in [0,1], and
𝑏𝑝
and 𝑗𝑝
represents the lower and upper limit of
the 𝑔
ℎ
variable. 𝐼
represents the 𝑎
ℎ
member, 𝐼is the
population matrix, 𝑡 indicates the count of decision
variables, 𝐸 specifies the population size,
𝑡
,
indicates the 𝑔
ℎ
problem variable value
represented by 𝑎
ℎ
population member.
ii)Fitness function
The solution is considered to be the optimal
solution with the minimal Mean Square Error (MSE),
and the MSE is proved as follows.
𝑀𝑆𝐸 =
1
∑
𝑞
∗
−𝑞
1
2
(26)
where,
j
q
∗
specifies the anticipated value,
𝑦 denotes the sample count, and 𝑞
characterizes the
recognized output by the DRN.
iii) Phase 1: Exploration
The members take part in the election depending
on their awareness and vote for a candidate. A
person’s awareness is considered in terms of the
goodness and quality of the objective function value.
The updating process in this phase is modelled
utilizing the equation given below,
()
()
,,
,1
,
,,
..,
.,
ag g ag H g
new F
ag
ag ag g
tnHAt BTBT
tntH else
t
+− <
=
+−
(27)
where, 𝐻 specifies the elected leader, 𝐴
represents an integer with value as 1 or 2, and 𝐵𝑇
is
its objective function value, 𝐻
is EBOA 𝑔
ℎ
dimension,
,1
,
new F
ag
t
refers the new created position for
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal
857
the
th
a
member,
,1
,
new F
ag
t
is its 𝑔
ℎ
dimension, and
,1new F
a
BT
indicates the objective function value.
This update process is modified by integrating the
HLBO in the EBOA, and the updated equation of
HEBO is expressed as,
𝑡
,
(
𝑥+1
)
=
1
2.𝑛.𝐴
𝐻
.𝑛
(
1 +𝑛.𝐴
)
−𝑛.𝑇𝐻
,
(
1 −𝑛.𝐴
)
(28)
In order to apply FC, subtracting 𝑡
,
(
𝑥
)
on both
sides,
By applying Fractional calculus [12],
() ()
()
()
()
()()
,, , ,
1
11..1.
2. .
ag ag g ag ag
tx tx Hn nAnTH nA tt
nA

+− = − + − − −

(29)
()
()
()
()
()()
, ,,
1
1.1..1.
2. .
ag g ag ag
Tt x Hn nA nTH nA t x
nA
α


+= + − − −


(30)
() () ()()() ()( )()
()
()
()
()()
,,, , ,
,,
11 1
1. 1 1 2 1 2 3
26 24
1
.1. . 1.
2. .
ag ag ag ag ag
gagag
tx tx tx tx tx
Hn nA nTH nA t x
nA
αα α ααα
+− − −− − − + − − −

=+−−−

(31)
()
()
()
()
()
()
()()
()()() ()( )()
,,,
,, ,
1
1.1..1.1
2. .
11 1
11 2 1 2 3
26 24
ag g ag ag
ag ag ag
tx Hn nAnTH nA tx
nA
tx tx tx
α
ααααα
+= + − − + −
+−+−−+−−−
(32)
where,𝐴specifies randomly selected number in
(
1,2
)
,
n
is the random variable
(
0,1
)
, 𝑇𝐻represents
hybrid leader, 𝑡
,
(
𝑥−1
)
is where the solution is
located at iteration (𝑥−1), 𝑡
,
(
𝑥−3
)
at 𝑥−3, and
𝑡
,
(
𝑥−2
)
is the position at iteration 𝑥−2 .
iv)Phase 2: Exploitation
The awareness of person in society in the voting
and election procedure has a high effect on their
decisions. Additionally, each person's activities and
thoughts and the leader's authority on a person's
awareness can enhance the people's awareness. From
the mathematical view point, an optimal solution can
be found depending on the local search near any
designed solution and is formulated as,
()
,2
,,
,
12 ..1 . ,
new F
ag ag
ag
x
tnGt
C
t

=+− −


(33)
where is refers to the recently generated position
for 𝑎
ℎ
member and its 𝑔
ℎ
dimension is specified as
,2
,
new F
ag
t
, 𝐶 refers to the maximal count of iterations
and 𝑥 refers to iteration contour.
v) Re-evaluating the fitness
The objective of the modified solution is
evaluated by equation (34) once updation is complete,
and the solution that succeeds in attaining the lowest
objective is deemed ideal.
vi)Termination
The above process is continued until the maximal
iteration is grasped.
3.6 Language Model
Here, GMM is used to determine the text transmitted
by EEG signals. Once the words are recognized by
the DRN, it is subjected to the GMM. The GMM (Afy
et al., n.d.) can map words from discrete space to
continuous space. It consists of a linear layer that
maps a vector in word space to a continuous
parameter space. The vocabulary size and vector size
are considered to be same. After that, these words are
applied to a multi-layer perceptron (MLP) and
combined based on the history. MLP selects an output
word for all input words or assigns a probability value
to every word in the vocabulary. Let us assume a
vocabulary 𝜒 with dimension 𝜀
1
and every word
𝑦,1 ≤𝑦≤𝜀
1
is represented by a vector with value ‘1’
at 𝑦
ℎ
position and remaining positions have a value
‘0’.The vector 𝜉
is mapped to a vector
r
e
with less
dimension 𝜀
2
with the support of a matrix 𝑊 as
below,
𝑒
=𝑊𝜉
(34)
where, matrix 𝑊 has a dimension 𝜀
1
×𝜀
2
.
Further, all the mapped words can be combined and
represented as,
𝑣
=𝑒
𝛽
1
…𝑒
(
𝛽
1
)
(35)
Here, 𝛽
specifies the
th
k
history and 𝐸 denotes
to the argument’s word identity. The GMM is built by
performing linear mapping of every history into a
small space given as follows,
𝑡
=𝑄𝑣
(36)
where, 𝑡 is the new feature space, which is textual
format contained in the brain thoughts. Thus, the EEG
signals are converted into text and 𝑄 specifies the
linear mapping.
INCOFT 2025 - International Conference on Futuristic Technology
858
4 RESULT AND DISCUSSION
In this section, the assessment of the outcomes
obtained by the DRN-FHEBO for conversion of brain
thoughts to text is portrayed. Also, performance
metrics and the dataset are discussed below.
4.1 Experimental setup
The designed FHEBO-DRN for conversion of brain
thoughts to text is executed by the MATLAB tool.
4.2 Dataset description
Here, conversion from brain thoughts-to-text is
carried out based on the data taken from the
handwriting BCI dataset (GitHub -
Fwillett/HandwritingBCI: Code from the Paper
“High-Performance Brain-to-Text Communication
via Handwriting,” n.d.). This dataset includes neural
activity in attempted handwriting, recorded with
43,501 characters in 1000 sentences over a period of
10.7 hours. The neuronal activity was recorded by
placing two microelectrode arrays with 96 electrodes
fixed to the hand region of the motor cortex. Also, it
includes the signature output of BCI in real-time.
4.3 Evaluation measures
The effectiveness of FHEBO-DRN is estimated by
utilizing assessment metrics such as Recall, F-
Measure, Text conversion accuracy, and Precision.
i) Precision: Precision estimates the fraction of
correctly classified samples or events among those
that are positively classified and is given as follows,
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =


(37)
wherein, 𝑇𝑃is the True Positive and 𝐹𝑃specifies
False Negative.
ii) F-Measure: This metric yields the Weighted
Harmonic Mean of recall and precision, and is given
by,
𝐹−𝑀𝑒𝑎𝑠𝑢𝑟𝑒=
2
(
2
)
(38)
where, 𝐹𝑁is the False Negative.
iii)Recall: Recall is a measure of how frequently
a machine learning method exactly identifies positive
instances from every true positive sample in the
dataset. It is expressed as,
𝑅𝑒𝑐 𝑎𝑙𝑙 =


(39)
iv)Text conversion accuracy: The accuracy of text
conversion is determined by determining the
proportion of the number of correctly converted
words to the whole number of words.
v)Error rate: The error rate is the measure of the
prediction error of the generated model concerning
the true model. Error rate is often used in the context
of classification models.
𝐸𝑟𝑟𝑜𝑟𝑟𝑎𝑡𝑒 =


(40)
4.4 Comparative methods
The designed FHEBO-DRN is examined based on the
techniques such as DCNN (Ullah & Halim, 2021),
RNN (Junwei et al., 2019), Hybrid DL (Willett et al.,
2021), and RF-classifier (Zhang et al., 2017) , HEBO-
Distributed Long Short-Term Memory (DLSTM).
Figure 3: Comparative analysis of FHEBO-DRN based on
learning set with a) precision, b) recall c) F-measure d) text
conversion accuracy and e) error rate
4.5 Comparative Analysis
The assessment of FHEBO-DRN for brain thoughts
to text conversion is executed with respect to learning
set. The assessment of FHEBO-DRN for brain
thoughts-to text conversion is done by learning set
and is shown in figure 3 Figure 3a) illustrates the
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal
859
analysis of FHEBO-DRN based on precision. The
precision recorded by FHEBO-DRN, with a learning
set of 90% is 92.765% and the precision computed by
DCNN is 82.765%, RNN is 84.878%, Hybrid DL is
86.766%, RF-classifier is 88.865% and HEBO-
DLSTM is 90.756%. This illustrates FHEBO-DRN
was successful in generating an enhanced
performance of 2.17% than the HEBO-DLSTM.
Figure 3b) describes the evaluation of FHEBO-DRN
based on recall. The recall obtained by FHEBO-DRN
is 95.765% with the learning set of 90% and the recall
figured by DCNN is 85.865%, RNN is 86.867%,
Hybrid DL is 88.867%, RF-classifier is 91.766% and
HEBO-DLSTM is 93.987%. The improved
performance of 9.29% is figured by FHEBO-DRN
than existing RNN. Figure 3c) explains the analysis
of FHEBO-DRN considering F-measure. The F-
measure figured by FHEBO-DRN, with a learning set
of 90% is 94.241% and the F-measure computed by
DCNN is 84.287%, RNN is 85.861%, Hybrid DL is
87.804%, RF-classifier is 90.292% and HEBO-
DLSTM is 92.343%. This demonstrates FHEBO-
DRN successfully recorded an enhanced performance
of 6.83% than Hybrid DL. In figure 3d), the valuation
of FHEBO-DRN considering text conversion
accuracy is illustrated. With learning set of 90%, the
text conversion accuracy recorded by FHEBO-DRN
is 91.765%. The text conversion accuracy valuated by
DCNN is 81.867%, RNN is 83.998%, Hybrid DL is
85.877%, and RF-classifier is 86.786%, HEBO-
DLSTM is 89.887%. The FHEBO-DRN achieved an
improved performance by 5.42% than RF-classifier.
Figure 3e) demonstrates the evaluation of FHEBO-
DRN on the basis of error rate. The error rate obtained
by FHEBO-DRN is 8.235%, with the learning set of
90% and the error rate figured by DCNN is 18.132%,
RNN is 16.001%, Hybrid DL is 14.123%, RF-
classifier is 13.213% and HEBO-DLSTM is
10.112%.
The comparative discussion of the FHEBO-DRN
is demonstrated in table 1.
The application of FHEBO for tuning the DRN
improved the convergence rate leading to accurate
word recognition. Further, the application of MAP for
signal segmentation and the high accuracy of DRN all
enabled effective attainment of conversion of brain
thoughts to text, resulting in excellent outputs.
Table 1: Comparative discussion
5 CONCLUSION
The conversion of brain thoughts to text depending
on BCI is seen as an emerging field. In this paper, a
technique for conversion of brain thoughts to text is
designed. Here, Input Brain EEG signals are taken
from the dataset, which contains EEG signals
corresponding to attempted imaginary thoughts of
questions/statements. Subsequently, the gaussian
filter is used for removing the noise in EEG signal and
then, signal segmentation and feature extraction are
processed. Later, word recognition is carried out
using DRN structurally optimized using the FHEBO.
Here, the proposed FHEBO is developed by the
combination of FC and HEBO. The HEBO is
developed by the combination HLBO, and EBOA.
Further, the Language modelling for speech
recognition is carried out utilizing the GMM.
Furthermore, the efficacy of the designed technique
is analysed and the FHEBO-DRN obtained a
maximal value of precision is 92.765%, F-measure of
94.241%, recall is 95.765% text conversion accuracy
of 91.765% and minimal error rate is 8.235%. In
future, advanced deep learning models will be
developed for enhancing the performance further. In
addition to this, EEG signals from other datasets can
be considered to validate the generalizability of the
approach.
Metrics HEB
O-
DLS
TM
Prop
osed
FHE
BO-
DRN
DC
NN
RN
N
Hyb
rid
DL
RF
classi
fier
Precisi
on
(
%
)
90.7
56
92.76
5
82.7
65
84.
878
86.7
66
88.86
5
Recall
(%)
93.9
87
95.76
5
85.8
65
86.
867
88.8
67
91.76
6
F-
measur
e
(
%
)
92.3
43
94.24
1
84.2
87
85.
861
87.8
04
90.29
2
Total
convers
ion
accurac
y(
%
)
89.8
87
91.76
5
81.8
67
83.
998
85.8
77
86.78
6
Error
rate
(
%
)
10.1
12
8.235 18.1
32
16.
001
14.1
23
13.21
3
INCOFT 2025 - International Conference on Futuristic Technology
860
REFERENCES
Afy, M., Siohan, O., & Sarikaya, R. (n.d.). GAUSSIAN
MIXTURE LANGUAGE MODELS FOR SPEECH
RECOGNITION.
Ahn, H.-J., & Lee, D.-H. (2021). Decoding 3D
Representation of Visual Imagery EEG using
Attention-based Dual-Stream Convolutional
NeuralNetwork. http://arxiv.org/abs/2112.07148
Bhaladhare, P. R., & Jinwala, D. C. (2014). A Clustering
Approach for the l -Diversity Model in Privacy
Preserving Data Mining Using Fractional Calculus-
Bacterial Foraging Optimization Algorithm . Advances
in Computer Engineering, 2014, 1–12.
https://doi.org/10.1155/2014/396529
Chen, Z., Chen, Y., Wu, L., Cheng, S., & Lin, P. (2019).
Deep residual network based fault detection and
diagnosis of photovoltaic arrays using current-voltage
curves and ambient conditions. Energy Conversion and
Management, 198.
https://doi.org/10.1016/j.enconman.2019.111793
Dehghani, M., & Trojovský, P. (2022). Hybrid leader based
optimization: a new stochastic optimization algorithm
for solving optimization applications. Scientific
Reports, 12(1). https://doi.org/10.1038/s41598-022-
09514-0
GitHub - fwillett/handwritingBCI: Code from the paper
“High-Performance Brain-to-Text Communication via
Handwriting.” (n.d.). Retrieved December 17, 2024,
from https://github.com/fwillett/handwritingBCI
Gupta, A., Sahu, H., Nanecha, N., Kumar, P., Pratim Roy,
P., & Chang, V. (n.d.). Enhancing text using emotion
detected from EEG signals.
Hassan, A. R., Siuly, S., & Zhang, Y. (2016). Epileptic
seizure detection in EEG signals using tunable-Q factor
wavelet transform and bootstrap aggregating.
Computer Methods and Programs in Biomedicine, 137,
247–259. https://doi.org/10.1016/j.cmpb.2016.09.008
Hong, K. S., Khan, M. J., & Hong, M. J. (2018). Feature
Extraction and Classification Methods for Hybrid
fNIRS-EEG Brain-Computer Interfaces. In Frontiers in
Human Neuroscience (Vol. 12). Frontiers Media S.A.
https://doi.org/10.3389/fnhum.2018.00246
Junwei, L., Ramkumar, S., Emayavaramban, G., Vinod, D.
F., Thilagaraj, M., Muneeswaran, V., Pallikonda
Rajasekaran, M., Venkataraman, V., & Hussein, A. F.
(2019). Brain Computer Interface for
Neurodegenerative Person Using
Electroencephalogram. IEEE Access, 7, 2439–2452.
https://doi.org/10.1109/ACCESS.2018.2886708
Kopparapu, S. K., & Satish, M. (2011). Identifying optimal
Gaussian filter for Gaussian noise removal.
Proceedings - 2011 3rd National Conference on
Computer Vision, Pattern Recognition, Image
Processing and Graphics, NCVPRIPG 2011, 126–129.
https://doi.org/10.1109/NCVPRIPG.2011.34
Kumar, P., Saini, R., Roy, P. P., Sahu, P. K., & Dogra, D.
P. (2018). Envisioned speech recognition using EEG
sensors. Personal and Ubiquitous Computing, 22(1),
185–199. https://doi.org/10.1007/s00779-017-1083-4
Mannepalli, K., Sastry, P. N., & Suman, M. (2017). A novel
Adaptive Fractional Deep Belief Networks for speaker
emotion recognition. Alexandria Engineering Journal,
56(4), 485–497.
https://doi.org/10.1016/j.aej.2016.09.002
Popescu, T. D. (2021). Signal segmentation using
maximum a posteriori probability estimator with
application in eeg data analysis. International Journal of
Circuits, Systems and Signal Processing, 15, 1336–
1345. https://doi.org/10.46300/9106.2021.15.144
Rajesh, S., Paul, V., Menon, V. G., Jacob, S., & Vinod, P.
(2020). Secure Brain-to-Brain Communication With
Edge Computing for Assisting Post-Stroke Paralyzed
Patients. IEEE Internet of Things Journal, 7(4), 2531–
2538. https://doi.org/10.1109/JIOT.2019.2951405
Trojovský, P., & Dehghani, M. (2022). A new optimization
algorithm based on mimicking the voting process for
leader selection. PeerJ Computer Science, 8, 1–40.
https://doi.org/10.7717/peerj-cs.976
Ullah, S., & Halim, Z. (2021). Imagined character
recognition through EEG signals using deep
convolutional neural network. Medical and Biological
Engineering and Computing, 59(5), 1167–1183.
https://doi.org/10.1007/s11517-021-02368-0
Willett, F. R., Avansino, D. T., Hochberg, L. R.,
Henderson, J. M., & Shenoy, K. V. (2021). High-
performance brain-to-text communication via
handwriting. Nature, 593(7858), 249–254.
https://doi.org/10.1038/s41586-021-03506-2
Zhang, X., Yao, L., Sheng, Q. Z., Kanhere, S. S., Gu, T., &
Zhang, D. (2017). Converting Your Thoughts to Texts:
Enabling Brain Typing via Deep Feature Learning of
EEG Signals. http://arxiv.org/abs/1709.08820
Fractional Hybrid Election Based Optimization with DRN for Brain Thoughts to Text Conversion Using EEG Signal
861