Design and Implementation of Hybrid Techniques and DA-based

Reconfigurable FIR Filter Design for Noise Removal in EEG Signals on

FPGA

C. SRINIVASA MURTHY, K. SRIDEVI

GITAM Institute of Technology, GITAM Deemed to be University, Visakhapatnam, Andhrapradesh,

INDIA

Abstract: - Virtual Reality (VR) technology assists physically challenged personnel in their daily routine activities.

The evolution of technology has enhanced the critical activities of people who use wheelchairs by extracting

features through electroencephalogram (EEG) and promoting options for their choice for decision-making on their

own. During extraction of EEG, signal artifacts may mislead the decision-making environment. Hence noise has to

be removed with help of an FIR filter for accuracy. In this context utilization of finite impulse response (FIR)

filters are so vital hence filters are incorporated with the hidden Markov model (HMM) and Gaussian mixture

model (GMM) and supervised machine learning architecture of multirate support vector machine (SVM). The

proposed EEG-based diagnosis system is a fully automated audio announcement system. The entire environment

has been developed by Verilog HDL and MATLAB. Validated on Artix-7 FPGA development board and

synthesized with Vivado Design Suite 2018.1. Obtained results exhibit an enhancement of 32% of signal-to-noise

ratio (SNR),7% of mean square error (MSE), and 69% of abnormality recognition.

Key-Words: - Virtual Reality (VR), electroencephalogram (EEG), hidden Markov model (HMM), Gaussian

mixture model (GMM), support vector machine (SVM).

Received: June 29, 2021. Revised: May 14, 2022. Accepted: June 12, 2022. Published: July 14, 2022.

1 Introduction

Electroencephalogram (EEG) signals are the primary

signals which can be used clinically to detect human

brain activities. EEG signals might be corrupted by a

motion of the lens during collecting data from

electrode results artifacts. This may occur because of

breathing and muscle contractions, as well as an

imbalanced contact between the skin and the surface

of the electrode [2-5]. Accidental noise signals,

known as motion artifacts, can hurt EEG reading and

make them unsuitable for analysis [6]. The

elimination of motion artifacts from EEG signals is

very necessary to determine the optimal EEG signal

for monitoring and analysis of brain activities.

Several alternative methods have been proposed to

remove artifacts from recorded EEG signals. The

most practical technique to eliminate motion artifacts

is to treat the signal with a digital filter. Filters can

eliminate extraneous signals without changing the

original signal if the predetermined objective is used

as a reference [8-11]. As a necessity, it is owned to

demand exact reference signals. The removal of EEG

motion artifacts is, however, inadequate. The

Discrete Wavelet transform (DWT) [12] is one of the

most widely utilized de-noising algorithms in

applications. In earlier research Hashim [14]

employed the wavelet threshold access to minimize

motion artifact noise in EEG recordings. EEG signal

baseline wander should be kept to a minimum. The

zero-phase high pass FIR Equi-ripple filtering

approach was suggested by Daqroud [8].

2 Related Work

In [7] passive wavelet transform has been employed

to extract the motion of artifacts from EEG signals

via non-contact capacitive coupling electrodes [13-

15]. The discrete wavelet process is required to pick

appropriate wavelet functions and thresholds, which

renders the wavelet a non-adaptive approach. To

address the wavelet transform's inadequacies in terms

of flexibility alone [16-19], Huang proposed an

empirical mode decomposition (EMD) procedure,

which is an adaptive technique for decomposing

signals into a limited number of intrinsic mode

functions (IMFs). Blanco-Velasco proposed an EEG

optimization strategy based on the EMD to reduce

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

324

Volume 17, 2022

high-frequency noise and baseline drift. Mode

mixing, on either side, is a common issue in the

EMD process that causes IMFs to be extracted

inadequately. Discrete wavelet transformations have

been advocated for quality removal in wavelet

vectors. For recognition of transmission line short-

circuit faults [20], the discrete wavelet transform has

been employed in fault analysis, EEG de-noising, and

disease recognition detection.

DWT is used to detect and eliminate movement

artifacts in EEG signals. RTP and RTCP methods

outperform the EMD method when a filter is

employed to reduce power-line intrusion and

accurate baseline drift. Hence we propose a method

to remove motion artifacts from practical EEG

recordings based on discrete wavelet transform and

wavelet thresholding in this article. At this point in

the process of validating the viability of the proposed

algorithm, real-time EEG has been captured and

reproduced signals. In terms of re-storage of the new

EEG theta delta alpha-beta (TDAB) complexes and

signal-to-noise ratio augmentation of EEG signals

associated with the discrete wavelet transform and

EMD motion artifact elimination have been

effectively enhanced and the proposed technique

executes much improved grades. The goal of the

current effort is to advance artifact exclusion practice

capable of identifying interference in EEG recordings

in a variety of situations and frames. The proposed

technique [18] includes a supplemental description

for improved recognition and organization of

abnormalities, as well as probable pulse-related

disruptions. For artifact removal, extracting features,

and detecting deviations the proposed strategy

consists of a three-stage method for assembling raw

EEG signals, transformed. The data has been

classified using classifiers, and the results have been

applied to FPGA for displaying the abnormal in the

display unit to take a decision. The method improves

the signal-to-noise ratio by gaining access to signals

that can protect the unique quality waveform without

corrupting the original raw data (SNR). The proposed

method expands the system which can recover all the

parameters available in an EEG raw signal. Several

denoising methods have been proposed in the

literature such as median filter, discrete cosine

transform, and discrete wavelet transform. In [21]

proposed multi-objective flower pollination

algorithm (MOFPA) with wavelet transform (WT)

for removal of artifacts and denoising of EEG signals

and different parameters analysis are done in terms of

SNR, PSNR, and MSE. In [22] is focused on Virtual

reality for improving the signal quality by removing

artifacts. Based on the results obtained, correction

prediction is 83% and 77% improvement in SNR. At

the onset, a filtering technique using a band-pass

finite impulse response (FIR) filter with a frequency

range of 0.5–40 Hz is used to remove various

artifacts and sounds mixed with raw EEG signals.

The DWT is then employed to evaluate the signals in

the time-frequency domain because EEGs are highly

non-linear and non-stationary signals in nature. For

feature extraction, DWT with four-level

decomposition is used with the db6 mother wavelet.

To effectively categorize the signal, a new feature set

formed of eleven non-linear statistical features

collected from each sub-band resulting from wavelet

decomposition is given to the input of an ANN.

Finally, to increase classification performance, a

unique approach called the sequential window

algorithm is used. This study achieved 99.44 percent

mean classification accuracy, 80.66 percent average

sensitivity, 4.12-second mean latency, and 0.2

percent average false positive rate (FPR). This

research successfully reduces latency time with

greater accuracy and a lower FPR [23]. Different

approaches for removing artifacts have been

proposed, however, artifact removal research remains

a work in progress. This paper focuses on the current

state of contaminated artifact removal. First, we'll go

through the features of EEG data and the many sorts

of artifacts. Following that, a basic overview of

current state-of-the-art procedures is offered, as well

as a detailed study of them. Finally, a comparison

analysis is presented to aid in the selection of

appropriate approaches for a given application [24].

Several novel methods for analyzing brain

bioelectrical signals were demonstrated and

compared. It also covers both traditional and modern

methods for removing noise contamination, such as

digital adaptive and non-adaptive filtering, signal

decomposition methods based on blind source

separation, and wavelet transformation [25]. It

distinguishes sleep stages and extracts novel

information from the sleep EEG to aid clinicians in

the diagnosis and treatment of sleep disorders. This

theory is based on exclusive EEG datasets from

Physionet, as well as the MIT-BIH, which is received

and reported by scientists for sleep range analysis

and prognosis. ML-based classifier utilizing

Ensemble Bagged Tree classifier achieved detection

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

325

Volume 17, 2022

accuracy of 95.9% on 18 records with 10197 epochs,

according to results [26].

Background: The EEG signals are a mix of small

amplitude in μV and different artifacts which are

contaminated due to the lens during the recording of

the signals, all these facts affect a change of signal

properties and these changes will create major issues

while analysing and diagnosing the abnormalities.

The EEG signals are also affected due to power line

noise, if signals frequency is above 45 Hz and it

makes it difficult to retrieve original information and

creates problems in effective diagnosis. In this

research work, the FIR filter plays a major role in

removing artifacts and creating noiseless signals for

proper further diagnosis. The EEG signals are

collected from https://physionet.org/news/post/397

and this database has a sampling rate of 256 with a

period of 5 seconds. The second process involved in

the FIR filter is an adjustment of DC shift and

removal of it by calculating the mean and setting it to

ground level. After the successful removal of

artifacts in EEG signals, different parameters like

SNR, MSE, and PNSR are analysed and compared

with existing works.

Motivation: In this work, the subjected EEG signals

are divided into different windows and each window

has 512 samples and applied to FIR filter design and

its performance analysis is carried out. For the

effective removal of artifacts, 64 tap FIR filter is

designed with the help fast RNS-based multiplier. To

minimize latency created by partial products, Parallel

Prefix Adder (PPA) is used along with the proposed

multiplier, the overall design is high throughput and

low latency. The digital FIR filter with 64th order is

subjected to epochs of each signal and analysis has

been done through different metrics to measure the

removal of noises and to retain signals quality. The

64-tap FIR filter uses Kaiser Window with low pass

filter response and based on metrics analysis, it is

found that the proposed design is better than existing

works.

Objectives: Design of efficient FIR filter for

effective removal of artifacts that are contaminated

through many sources such as 45Hx noise, eye blink,

EMG noise, and many more, but the noise which is

affected by power line of frequency of 45Hz is

dominated by noise and removal the of it, is a major

concern. Therefore, removal of such noises is a major

motivation of this work, and filter signals should

ready proper diagnosis and good quality with high

accuracy. Based on a literature survey done on

various techniques used for the removal of artifacts,

the following objectives are highlighted for this

research work.

1.Detailed study on different properties and

characteristics in EEG signals

2.Detailed analysis of FIR filter design for

optimization of power consumption, latency, and

improvement in throughput.

3.Design of FIR with both Kaiser Window and low

pass filter in Xilinx Design suite software using

Verilog HDL and generation of coefficients with help

FDA tool in MATLAB.

Detailed analysis of results obtained through

proposed FIR filter and comparison between

proposed and existing in terms of parameter metrics.

About EEG Databases: The two strategies for

eliminating raw EEG signals are simulated databases

and real-time databases. The standard database

contains a selection of EEG signal events of

comparable length, which is provided with the

dataset from the MIT-BIH Noise Stress Test

Database for duplication commitments. A distinct

signal includes time-shifting TDAB morphology in a

pair of regular and irregular EEG pulses and based on

the pair and irregular pulse, the abnormal is detected.

3 Proposed Automatic Abnormality

Detection System for Physically

Challenged Peoples

EEG files have been created by adding motion

distortions to EEG data outputs at various levels of

signal-to-noise ratio. In hospitals, EEG signal

acquisition systems are typically large and maintain

excellent precision and long-term monitoring. All

types of EEG recording equipment, from single-lead

to 12-lead, are available. In real-time monitoring of a

patient's condition through the use of many sensors’

certain wearable health-monitoring technologies

improve awareness on either side. The dataset

collecting stage of EEG includes the collection of

sensor forms, the position of sensors to point, the

number of sensors, and the hardware required for

data acquisition, archiving, and dissemination. In a

few real-time EEG detecting techniques, continuous

EEG sensor acquisition may be employed. It is a

difficult task to target EEG signals to multiple feature

assessments such as quality, reliability, and

punctuality. The erroneous disease identification and

a major impact on clinical outcomes will result from

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

326

Volume 17, 2022

the corrupted collected data. Fig.1 shows the

proposed overall architecture of abnormalities

detection and its classification which is based on

SVM and architecture consisting of an FIR filter,

Block memory generator and its controller, feature

extraction using HMM and GMM, and an SVM-

based classifier. The sample of EEG signal captured

by wireless Neurosky Mind wave Mobile 2 device is

stored in block memory which is off-chip BRAM.

The stored sample values are applied to the FIR filter

to eliminate noise samples that have high-frequency

signals and the filter uses fixed-point arithmetic

operations of size 16 bits including sign bit. The

block random access memory (BRAM) controller is

an IP core industry-standard component instantiation

in the top-level design which communicates between

overall system design and memory. The raw sample

of EEG signals is processed through an FIR filter and

then applied to HMM and GMM to extract abnormal

features present in the filtered signals. For each

extracted feature, the variances, energy, and

meaningful metrics are calculated and these metrics

are used for classifications. The multi-rate SVM

which uses the sigmoid function, has two phases such

as learning (training) and testing, in the first phase,

training the network against normal and abnormal

signals, and second, the phase is testing to detect

abnormal in the selected EEG signal from data sets

and measurement accuracy and based on obtained

metrics, the SVM classifies the dataset into normal

and abnormal. For the implementation of HMM,

GMM, and SVM functionality, double-precision

floating point-based multipliers and adders are used

in the proposed system. The main advantage of

floating-point operations compared to fixed-point

operation is that distortions are almost zero in

floating-point operations therefore detection of

abnormalities are accurate. In floating-point

operations format has 52 bits for mantissa, 11bits

exponent, and 1 bit for the sign. For validation of

floating operations results that are obtained in digital

software tools compared with MATLAB results since

MATLAB tool works on based floating-point and it

is found that both results are same.

4 Feature Extraction using HMM and

GMM Modules

In this paper, GMM with HMM is applied to filtered

EEG signals for extraction of features and

classifications of different categories. HMM,

mathematical theory models are provided to

understand the use of HMM as shown in Fig.2. For

abnormal features extraction in EEG signals through

HMM and GMM have been pipelined in this work

and each element of HMM has been made mandatory

for proper identification of disorders. Characteristics

of HMM are as follows. The number of states in

HMM is N and it has its physical significance and

acts a as hidden part of Baki’s model allowing

transitions to each state which has modeled the index

compato red the present state [10]. The number of

observations (M) is distinct from other observations

and these are the output of the modeled system and

all states are represented as S={S1, S2, S3,......... Sn}.

 The distribution of probability in each transition

state is X= {xij}which provides the probability of

states reaching to final decision from any other

states and it is given by.



For lower indexing then present state, the 

which has transitions to any other states.

 P is probability distribution of symbol which is

observed in j states i.e󰇛󰇜, where

󰇛󰇜



 Si is initial state distribution and it is given by

󰇟󰇠, therefore the complete features

extractions required two level decomposition

model parameters such as N , M to measure the

constants A,B, P and , the complete HMM

function is given by Q=(A,B,󰇜.

The features extraction using HMM and its selection

to get trained samples as shown in Fig.2.

Observations (M): The proposed system employs

numerous sequences to collect enough samples of

data to establish efficient model that will estimate

distinct samples from other samples. In GMM the

number of mixtures in each state (M) the

distributions in Gaussian for different weights will

form a new density probability which will be

generated by HMM. According to the fundamental

concept, a higher value of M would allow for

accurate signals modelling, because of very short

duration of other signals. A larger M value would

leave very few estimates for these signals. As a

result, accuracy in signal estimation occurs for M = 4

may produce better categorization results into normal

and abnormal shown in Fig.2.

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

327

Volume 17, 2022

The number of iterations and Convergence: The

two existing criteria in HMM are used for training to

convergence and the number of iterations. HMMs are

trained by raising the model's based on score for EEG

signals collection of observation samples.

Fig. 1: Proposed diagram of FIR filter for artifact

removal and abnormal detection using SVM

The accurate and tolerance level, which is set to

0.0001, refers to the percentage increase in this score

from the previous iteration, r model to be regarded

to have converged if the gain in score for the current

iteration is less than 0.0001. In general the HMMs

required 35-51 iterations to converge, however this

depends on a variety of criteria, including tolerance,

data, and good modelling parameter selection. The 80

iterations are a mandatory number having one of

these two conditions is met, the training comes for

final decision. Median filtering (MF) is used along

with FIR which is windows based model (w) can

replace the intermediate signals for optimization has

been applied for filtering the noise at very low

frequency signals to produce nonlinear smooth

signals. The prescribed series of n samples within a

window of length L as per equation (1), where M is

the median value.

W = {xn}, -L≤ i ≤L (1)

where M =median (xn)

In the VLSI environment effectiveness has to be

planned in terms of look up tables (LUTs) and slice

register (SR) for less consumption of hardware

resources. Hence a combination of DCT and FIR

filters has been employed for efficient filtrations at

both low and higher. The phases of EEG de-noising

based on MF can be summarized as follows: A

median filter with a width of 200 ms is used to

separate the TDAB complexes and P waves from the

noisy EEG input. A 600-ms MF is used to separate T

waves from the following signal. The second filter

action restricts the baseline of the EEG signal, which

has to be subtracted from the noisy EEG data to

obtain the adjusted baseline EEG signal. The basic

FIR filter function is given by

󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜









Where k is the length of the filter, in this design, the

k is 0 to 63 [21]. Equation (2) proposes FIR filter

operation with the direct structure which requires less

number of registers and its internal architecture is

shown in Fig.3. The Conventional FIR filter design

uses a binary number system for adders and

multipliers, which results in increased propagation

and net delays and limits the pace of operations. To

overcome these drawbacks, the suggested RNS-based

FIR filter presented in equation (1) employs a faster

modified Parallel Prefix Adder (PPA) that avoids

carry bit propagation The results of the existing PPA

EEG Signal Acquisition

System through Neuro-

Sky Mind ware Mobile

FIR filter for removal of

noise and artifacts

Pre-processed Signal at

10Hz, 20Hz, 30Hz and

40Hz frequencies

Features

extraction

using GMM

and HMM

algorithms

Power

estimation

using GMM

Clusters formation

Extracted

features for

base line

Trained

features for

final

validations

Test EEG signal

which is unknown

Final classified decision

and Information to

physically handicapped

about its decision

Machine

Learning based

SVM

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

328

Volume 17, 2022

and modified PPA are displayed in Table 1. Once the

power band performance from the EEG signals have

been extracted, computation of incident-linked

(de)synchronization (ERD/ERS) directory, a measure

of band power shift in EEG is carried out which has

first introduced by Pfurtscheller as well as Aranibar

[34] and can be determined as follows:





 

Where IBP is Interval band power is.

Fig. 2: Proposed abnormalities identification and

features extraction using HMM and SVM based

Classifications.

4.1 DCT Transform

The discrete cosine transform (DCT) is employed for

signal transmission as well as used for the

presentation of a digital image processing technique

that converts the time-domain function to a frequency

domain function. The Discrete Cosine Transform is

the true part of the Discrete Fourier Transform of a

signal w (k) represented in equation (3).

󰇛󰇜󰇛󰇜 󰇛󰇜󰇛󰇛󰇜󰇛󰇜





 ) (3)

Here the length of the noisy EEG signal is

represented by N as well as DCT quantities is

represented by w(k) which is computed as follows:

w(k) = 







Substitute k as well as n in the y (k) equation to

determine their inverse form. The DCT provides

excellent energy density in the case of highly

interlinked data. Once the energy compression

feature of DCT is obtained then this feature is

conversed and provided with information to various

predefined EEG signals operations. Further, we

found that the DCT spectrum is situated at a low

frequency, which represents the initial characteristics

of the DCT transform. The very first coefficients that

are generated using a one-dimensional DCT

transform are assumed to correlate with BW artifacts

in the EEG signal. In input EEG signals, correlated

data is extracted by DCT and calculated their energy

through transform coefficients.

5 Classification using Multi-Rate SVM

To categorize data into normal as well as abnormal

classifications, a precise classifier based on Neural

Networks using SVM classifiers can be employed.

Machine learning is a type of artificial intelligence

(AI) that enables computers to analyze the records of

an individual with trained algorithms, and also it can

be used to pre-process systems with various datasets,

and finally, it recognizes each record separately. The

technique of machine learning is to saturate the data

to identify probable patterns in the data as well as to

modify program actions. The technique of Machine

learning includes unsupervised as well as supervised

learning. Preparation of information in targeted

learning exists in the form (Xn,Yn)n = 1,2,.... N, with

Xn representing examples from an input space X as

well as Yn is generated by an undefined and

unpredictable process with input as Xn. The most

common technical issue faced in supervised learning

is classification as well as regression. On the other

hand, unsupervised Learning is provided with a set of

samples Xn, n = 1,2,.... N, however, the labels Yn are

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

329

Volume 17, 2022

unstable. Although Yn is not accessible, the

outcomes that seem to be available are referred to as

unsupervised learning. In this scenario, learning

problems include an increase in data clusters; the

center of clusters, as well as the clusters themselves

of the input data is often associated with pattern

classifications as well as structure evaluations. We

propose the Support Vector Machine (SVM) as a

machine learning algorithm that can provide better

results while separating datasets and have a more

stable presentation when dealing with data

uncertainty in our technique. The datasets that were

initially generated were stored in separate training

and testing sets that were independent of each other,

finally through VLSI programming, the SVM

algorithm was applied to the data pre-processing.

With the help of clustering, the excluded structures

are then linked to the training data set structure.

Cross-justification is often used to estimate how a

machine learning algorithm will operate when

confronted with unexpected input. The Machine

learning algorithms generate a hypothesis and

thereafter evaluate it using the same data. During K-

fold cross-authentication, the data is unevenly

divided into K divisions. To train the sets residual

data sets are used and one more data is used for

analysis and be evaluated, every division set should

be dignified K times. To test the viability of the

technique proposed, Three-fold cross-validation is

employed. Every row is a tradeoff to check the

accuracy using a classifier. Although the classifiers

are mongrelized, the classification accuracy gained

from them is described as the load.

6 Performance Analysis in Terms of

Accuracy, FER and FAR

Some of the parameters that can be employed to

measure the efficiency of the projected procedures

include theta, delta, alpha, and beta (TDAB) values

which have to be identified for accurate calculation

of SNR, accuracy, as well as error rate. Classifier

evaluation metrics areas followed. Three essential

analytical procedures are measured to identify the

classification performance of the proposed method.

We compute the accuracy of the algorithm

classifications as the fraction of all true

classifications of overall classifications as seen in

Fig.2.Initially, limit the algorithm's classiﬁcation for

accuracy calculation to a fraction of all true

classifications throughout all classifications.

Fig. 3: Internal Architecture of FIR filter with64 taps

[21]

Accuracy = 󰌣󰇛󰇜

󰌣󰇛󰇜

Where TP is true positive, TN is true negative, FP is

false positive and FN is false negative.

Furthermore, the classification of the algorithms for

false error rate (FER) has been computed as the ratio

of falsely classified as per the equation below:

FER= 󰌣󰇛󰇜

󰌣󰇛󰇜

Finally, the classification algorithm's sensitivity is

computed and is represented as follows:

s = 

󰇛󰇜

Here the hit rate and false acceptance rate is

represented by HR and FAR respectively and is

given as follows:

HR= 󰌣󰇛󰇜

󰌣󰇛󰇜

FAR= 󰌣󰇛󰇜

󰌣󰇛󰇜

The efficiencies and FAR are 1 and 0 respectively

only when the classifications are correct. Precision,

FAR, as well as s, are all computed separately.

7 Results and Discussion

There is a significant improvement in the detection of

abnormalities as compared to conventional

classifiers, as evidenced by performance indicators

like accuracy, SNR, PSNR, and Energy. This is due

to the fact that there are more samples available in

the given data set for training the HMMs, and the

processing time is more, in order to minimize the

elapsed time, GMM has been used to train the

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

330

Volume 17, 2022

network and stored in memory and then applied to

HMM. Furthermore, because each and every

sequence length is longer, hence DCT model

provides a superior fit for describing non-disease and

illness data. Hybrid types of HMMs, in which

multilayer perceptions or clustering algorithms are

utilized to obtain non-disease data, can be employed

to create better models. As the number of data

samples increases and the number of training samples

grows, we can see that performance improves in

obtained results shown in Table 4. The next stage is

real-time validation in EEG data samples that are

present in the database after designing the detector

and classifier HMMs. At this stage, no more data

processing or feature extraction is required. This is

profit accomplished using the HMMs from the

previous existing works. The validation method for

the performance of the HMM-based detection and

classification system and its accuracy is depicted in

Figure 7. For a few signals with definite uncertainty,

signal are considered in order to prove the

functionality of HMM system and this allows us to

compare abnormalities detection and their

performance to the diagnosis. Tables 3 and 4 exhibits

the utilization of resources by the proposed system

for effective detection of abnormalities and hardware

resources utilized are generated through the synthesis

process. Table 3 shows PSNR and SNR values for

different activities. The proposed HMM has been

used as the main model for feature extraction and

SVM classification for the detection of abnormality

based on trained features derived from the entire

dataset. In abnormal detection, much work has been

carried out with cascading HMMs such as HMM1,

HMM2 and HMM3 or using other classifiers like

MLP. The proposed hybrid method based on HMM,

GMM, and SVM provides better accuracy and

randomized parameters like PNSR and SNR. This

strategy has potentially provided well precise

outcomes. In addition, in the existing works the

sampling frequency employed varies by location of

abnormality due to more artifacts. Different EEG

monitors are accompanied by different power

supplies (amplitude in V), which can be adjusted

according to the expert's preferences. These

quantities may influence the extracted features and

have been standardized in the industry as well in

research. More trustworthy features can be extracted

with a better specification of the exact shape, length,

and other parameters of the EEG spike. In addition, a

uniform database for all EEGs would be extremely

useful in comparing and contrasting all of these

methods and features. Expert-provided signal

exemplars would also be very useful for the Non-

disease data. It is necessary to provide an example

dataset with signals or artifacts that are identified by

experts as being very similar to existing. Models

trained on such exemplars would very certainly be

producing superior results. All of the reported

datasets are associated with various anomalies, as

described below, in Table 1. Each dataset comprises

3600 samples, which are rehabilitated to binary using

MATLAB source code and stored in the FPGA's

block-RAM (BRAM). When compared to previous

works, all projected practices for precise irregularity

recognition can generate performance with 99%

exactness and there is a development of 5%in

accuracy and detection of abnormalities. Table 2

shows in positions of anomaly recognition and

recognition precision in percentage, the projected

effort, and the present effort are equated. The

projected automated abnormality recognition system

is originated to be proficient in noticing anomalies

from the consistent database, which are from the

MIT-BIH arrhythmia database the suggested and

validated by medical export and proposed results

shown in Table 1 and Table2, based on obtained

results the Atrial fibrillation with 99% accuracy,

Atrial flutter with 68.2% accuracy, Normal Sinus

rhythm with 93% accuracy, Sinus Tachycardia with

65 percent accuracy, and Ventricular Fibrillation with

100% precision, among other diseases are shown in

Table 2. The projected scheme has a precision of

85.24 percent on average as shown in Fig.4 and

SSIM is shown in Fig.3 (b) and PSNR is shown in

Fig.4(a).

Table 1. Number of records present in the database

for various abnormalities

Database

record

number

Name of the abnormalities

Database

count

2-14

Activity:1(More Reading)

14

2-17

Activity:2(Peak-Peak

variations)

17

2-11

Activity:3(Eyes

movement)

11

2-14

Activity:4(Left Leg

movement)

14

2-15

Activity:5(Right Leg

movement)

15

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

331

Volume 17, 2022

DCT has excellent energy compression and power

indulgence abilities compared to STFT, FF, and the

Fig. 8(b) shows the EEG database record no.15 of

atrial fibrillation and 12 of atrial flutter, and their

energy compaction results shown in Fig. 8(c) through

lettering Verilog HDL programming language and

manufactured by VIVADO design suite 2018.1. The

RTL design diagram is shown in Fig.9, the simulated

results are shown in Fig.9the simulated results of

artifacts deletion and types removed results are

shown in Fig.8 and 9the manufactured effects of the

projected health observing scheme are shown in

Table. 4. Since the device is quick and has been

tested against a standard database, the prototype

module established will undoubtedly aid in the

monitoring of patients’ health. The outcomes showed

that the system operates at 113.148MHz, power

consumption is of 0.082Watts, several LUTs utilized

is 485, area (no of slice register) is 284 and delay is

4.419ns. The proposed system is synthesized in

Xilinx software tool and final results are interface

with microcontroller (Arduino Uno board), from it,

the GSM is an interface to send a message about

soldier's health situations to medical professionals or

concerned persons so that soldiers life can be saved

instantly. The entire system is designed in Verilog

HDL and processed through FPAG design flow like

synthesis, mapping, translation, place & route, and

finally bit steam file is generated and the same file is

configured on FPGA. Based synthesis report, the

total delay of the proposed system is 2.5ns and power

consumption is 1.8mW.

Table 2. Comparative analysis of proposed research

work with existing work in terms of detection of

different abnormalities and accuracy

Databa

se

record

number

Name of

the

abnormali

ty

Detected

(YES/N

O)

Propose

d work

Accura

cy

Existin

g work

Accura

cy

1

2

3

4

5

6

7

8

9

10

11

Activity:1

Yes

98%

92.2%

12

Yes

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

Activity:2

Yes

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

81%

78.2%

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

Activity:3

Yes

No

Yes

89%

79%

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

332

Volume 17, 2022

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Activity:4

No

Yes

No

Yes

No

76%

69%

1

2

3

4

5

6

7

8

9

10

11

12

Activity:5

Yes

100%

89.7%

All pulse detecting algorithms functioned perfectly

for a large number of data present in the MIT-BIH

Arrhythmia Database. However, because of the

irregular heartbeat as well as interference effect, only

a few records in this database possessed dynamic

signals, especially 12, 18, 16, 15, 12 records. Earlier,

these records were utilized to examine the noise

tolerance of earlier investigations. As per the Physio

Net web-based resource, the signal from record 73

seems to be more challenging in MIT-BIH

Arrhythmia Database because of the signal's

characteristic pattern with uneven beats. Table 1

compares the algorithm’s output for another few

challenging records such as 12, 18, 16, 15, and 12.

The heartbeat indicator can fluctuate between normal

and abnormal beat signals as evidenced by records

12, 18, 16, 15, and 12. The signal from records 20

and 73 revealed a regular, varied heartbeat, and a

normal, atrial late, premature ventricular contraction

beat respectively. As seen in Fig.9, the EEG signal

for record 53 is a composite of regular as well as

controlled ventricular reduction beats. To remove the

artifact outcome, pre-processing is very important for

EEG signals. The proposed technique for removing

artifacts from raw EEG data uses a median filter,

HMM, GMM, as well as DCT with the height

constants of the DCT component set. SVM classifiers

have been used to distinguish correct datasets as

output. Furthermore, the output signal has been given

as input to RTP as well as RTCP before being saved

in loading procedures. The proposed algorithm is

more functional and provides enhanced performance.

In the database, there are 100 EEG signals taken

from the Physio net website, out of 100 EEG

databases, we have considered 10 EEG signals due to

the limitation of internal memory the of FPGA

hardware board.

Fig. 4: Average SNR of each activity for different abnormalities

0

20

40

60

80

100

Activit

y:1

Activit

y:2

Activit

y:3

Activit

y:4

Activit

y:5

SNR

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

333

Volume 17, 2022

Fig. 5: Average SSIM of each activity for different abnormalities

Fig. 6: Average PSNR of each activity for different abnormalities

Fig. 7: Average accuracy of all types of abnormalities and their corresponding activities

0,00E+00

2,00E+04

4,00E+04

6,00E+04

8,00E+04

1,00E+05

1,20E+05

1,40E+05

1,60E+05

1,80E+05

Activit

y:1

Activit

y:2

Activit

y:3

Activit

y:4

Activit

y:5

SSIM

0

20

40

60

80

100

120

140

160

Activit

y:1

Activit

y:2

Activit

y:3

Activit

y:4

Activit

y:5

PNSR

0%

20%

40%

60%

80%

100%

120%

Proposed work Accuracy Existing work Accuracy

Accuracy

Activity:1 Activity:2 Activity:3 Activity:4 Activity:5

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

334

Volume 17, 2022

(a) (b) (c)

Fig. 8: Database EEG signals, (a) packet loss w.r.t number of round to upload the packet into cloud for storage (c)

After removal of artifacts in MATLAB.

Fig. 9: Artifacts removed simulated results of EEG signals for automatic detection of abnormality

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

335

Volume 17, 2022

Database

record number

Name of the

abnormality

Detected

(YES/NO)

SNR

SSIM

PNSR

1

2

3

4

5

6

7

8

9

10

11

12

Activity:1

Yes

53.96

46.84

46.92

48.2324

50.7312

29.9189

53.9620

46.8485

46.9271

29.9198

50.7312

48.2324

1.0986e+05

4.6745e+04

7.8190e+03

5.4343e+04

7.2932e+04

1.6889e+04

1.0986e+05

4.6745e+04

7.8190e+03

1.6889e+04

7.2932e+04

5.4347e+04

145.0401

145.0402

131.9772

145.0401

145.0682

17.7023

145.0401

145.0402

131.9772

17.7023

145.0682

145.0401

1

2

3

4

5

6

7

8

9

10

11

12

13

15

16

17

18

Activity:2

Yes

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

42.8603

53.6969

88.5

50.3966

53.0835

67.45

42.8603

67.23

53.0835

68.2

50.3966

53.6969

65.14

53.0835

50.3966

67.23

53.6969

48.8604

1.4563e+03

1.0246e+05

1.5646e+05

7.0063e+04

4.2881e+04

84.5

1.4563e+03

1.452 e+03

3.3495e+03

3.4495e+03

7.1088e+03

1.0246e+05

1.3246e+05

4.2881e+04

7.0063e+04

3.0063e+04

1.0246e+05

1.4563e+04

126.3978

145.0476

145.4576

145.0401

127.9498

128.34

126.3978

126.7

127.9498

128.56

145.0401

145.0476

145.752

127.9498

145.0401

146.34

145.0476

138.9498

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

Activity:3

Yes

No

Yes

54.5508

56.7265

51.1981

56.7770

46.4857

51.5268

54.5508

56.7265

51.1981

56.7770

57.34

46.4857

51.5268

5.0267e+04

6.9477e+04

2.9540e+04

1.4857e+05

4.4852e+04

1.0205e+04

5.0267e+04

6.9477e+04

2.9540e+04

2.95402e+04

1.4857e+05

56.78

4.4852e+04

1.0205e+04

140.2728

117.8600

142.0947

145.0673

145.0401

127.6884

140.2728

117.8600

142.047

142.0947

145.0673

145.67

145.0401

127.6884

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

336

Volume 17, 2022

Table 3 Comparative analysis of proposed research work with existing work in terms of detection of different

abnormalities and accuracy

Table 4. Performance analysis and comparison of proposed research work with existing works in terms of SNR,

SSIM and PNSR.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Activity:4

No

Yes

No

Yes

No

53.0643

54.5508

53.5156

No

53.0643

54.5508

53.5156

No

3.3401e+04

1.3134e+04

3.1312e+04

No

3.3401e+04

1.3134e+04

1.3131e+04

3.1312e+04

3.1212e+04

No

121.2964

129.5245

142.1463

No

121.2964

129.5245

142.1463

No

1

2

3

4

5

6

7

8

9

10

11

12

Activity:5

Yes

53.2740

52.0412

53.6249

52.5063

52.3400

53.2740

52.1431

53.2740

52.0412

53.6249

52.5063

1.4207e+04

3.5743e+04

4.1498e+04

4.0358e+04

3.9676e+04

1.4207e+04

1.4357e+04

1.4207e+04

3.5743e+04

4.1498e+04

4.0358e+04

133.0471

120.9439

120.8381

128.1822

127.7840

133.0471

132.1271

133.0471

120.9439

120.8381

128.1822

Parameter

Existing work

Proposed work for Automated

abnormality detection system

EEG signal coefficients bit size

8

Flip-flop used

6902

Delay

7.31ns

Area (Slices)

11093(9%)

Power utilized

0.088W

Slice registers

2872

No of LUT’s

2871

Throughput

Size of the data/Delay=1.810Gbps

Latency

Product of Delay and size of

data=4.419ns*8=35.352ns

Area-delay product

4.145*35.352ns =52380.365ns

Power and Delay product

0.082*12.637mW=14.653404mW/ns

Frequency

MHz

113.148MHz

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

337

Volume 17, 2022

8 Conclusion

HMM and SVM produced results in terms of

sensitivity and accuracy, as shown in Tables 3 and

4. It can be observed from table.2 that the accuracy

has been always higher than the sensitivity due to

the availability of the number of training samples

for every class. The convergence graphs also

indicate that HMM requires a good number of

iterations to train. Eventually, increasing the number

of training samples can enhance performance. In

addition that multi-resolution analysis involved in

sampling the EEG at greater frequencies (> 1028Hz)

proved to increase the detection of abnormality and

might be used on EEG signals after FIR filter

filtering. Average accuracy and detection rates have

been considered since the models perform well in

the detection of abnormality for the processed data

sets, but it has been observed that when tested with

individual EEG signals several errors have been

produced in the form of false positive and false

negative. The total EEG signals have been divided

into windows of each 256 samples and processed

through HMM and GMM. The proposed system

exhibits a better outcome compared with other

strategies like thresholding EEGs and combining

ANNs with HMMs. Based on obtained results, a

comparison has been made between proposed and

previously published state-of-art works in the same

area, it is found that 27% improvement in speed of

process, 42% in identification of abnormalities in

EEG signals and 35% improvement in accuracy.

References:

[1]. S. Bose, A. De and I. Chakrabarti, "Area-

Delay-Power Efficient VLSI Architecture of

FIR Filter for Processing Seismic Signal," in

IEEE Transactions on Circuits and Systems II:

Express Briefs, vol. 68, no. 11, pp. 3451-

3455, Nov. 2021, doi:

10.1109/TCSII.2021.3081257.

[2]. E. G. Pale-Ramon, Y. S. Shmaliy, J. A.

Andrade-Lucio and L. J. Morales-Mendoza,

"Bias-Constrained H₂ Optimal Finite

Impulse Response Filtering for Object

Tracking Under Disturbances and Data

Errors," in IEEE Transactions on Control

Systems Technology, doi:

10.1109/TCST.2021.3118321.

[3]. X. Liu, M. Lewandowski and N. K. C. Nair,

"A Morlet Wavelet-Based Two-Point FIR

Filter Method for Phasor Estimation," in IEEE

Transactions on Instrumentation and

Measurement, vol. 70, pp. 1-10, 2021, Art no.

6503310, doi: 10.1109/TIM.2021.3075743.

[4]. X. X. Zheng, J. Yang, S. Y. Yang, W. Chen,

L. Y. Huang and X. Y. Zhang, "Synthesis of

Linear-Phase FIR Filters With a Complex

Exponential Impulse Response," in IEEE

Transactions on Signal Processing, vol. 69,

pp. 6101-6115, 2021, doi:

10.1109/TSP.2021.3115352.

[5]. Y. S. Shmaliy, Y. Xu, J. A. Andrade-Lucio

and O. Ibarra-Manzano, "Predictive Tracking

Under Persistent Disturbances and Data

Errors Using $H_2$ FIR Approach," in IEEE

Transactions on Industrial Electronics, vol.

69, no. 6, pp. 6121-6129, June 2022, doi:

10.1109/TIE.2021.3087403.

[6]. S. M. J. A. Tabatabaee, M. Rajabzadeh and

M. Towliat, "A Novel Low-Complexity

GFDM Relay Communication System: Relay

Selection and Filter-and-Forward," in IEEE

Transactions on Signal Processing, vol. 69,

pp. 5147-5158, 2021, doi:

10.1109/TSP.2021.3108679.

[7]. X. Liu, M. Lewandowski and N. K. C. Nair,

"Erratum to “A Morlet Wavelet-Based Two-

Point FIR Filter Method for Phasor

Estimation”," in IEEE Transactions on

Instrumentation and Measurement, vol. 71,

pp. 1-1, 2022, Art no. 9900201, doi:

10.1109/TIM.2021.3127767.

[8]. Y. -E. Lee, N. -S. Kwak and S. -W. Lee, "A

Real-Time Movement Artifact Removal

Method for Ambulatory Brain-Computer

Interfaces," in IEEE Transactions on Neural

Systems and Rehabilitation Engineering, vol.

28, no. 12, pp. 2660-2670, Dec. 2020, doi:

10.1109/TNSRE.2020.3040264.

[9]. N. Richer, R. J. Downey, W. D. Hairston, D.

P. Ferris and A. D. Nordin, "Motion and

Muscle Artifact Removal Validation Using an

Electrical Head Phantom, Robotic Motion

Platform, and Dual Layer Mobile EEG," in

IEEE Transactions on Neural Systems and

Rehabilitation Engineering, vol. 28, no. 8, pp.

1825-1835, Aug. 2020, doi:

10.1109/TNSRE.2020.3000971.

[10]. M. Dora and D. Holcman, "Adaptive Single-

Channel EEG Artifact Removal With

Applications to Clinical Monitoring," in IEEE

Transactions on Neural Systems and

Rehabilitation Engineering, vol. 30, pp. 286-

295, 2022, doi:

10.1109/TNSRE.2022.3147072.

[11]. S. Zahan, "Removing EOG artifacts from

EEG signal using noise-assisted multivariate

empirical mode decomposition," 2016 2nd

International Conference on Electrical,

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

338

Volume 17, 2022

Computer & Telecommunication Engineering

(ICECTE), 2016, pp. 1-5, doi:

10.1109/ICECTE.2016.7879634.

[12]. J. -S. Kang, S. Kavuri and M. Lee, "Adaptive

EEG noise filtering for coherence analysis,"

2014 International Winter Workshop on

Brain-Computer Interface (BCI), 2014, pp. 1-

4, doi: 10.1109/iww-BCI.2014.6782569.

[13]. C. Chou, T. Chen and W. Fang, "FPGA

implementation of EEG system-on-chip with

automatic artifacts removal based on BSS-

CCA method," 2016 IEEE Biomedical

Circuits and Systems Conference (BioCAS),

2016, pp. 224-227, doi:

10.1109/BioCAS.2016.7833772.

[14]. S. Siuly et al., "A New Framework for

Automatic Detection of Patients With Mild

Cognitive Impairment Using Resting-State

EEG Signals," in IEEE Transactions on

Neural Systems and Rehabilitation

Engineering, vol. 28, no. 9, pp. 1966-1976,

Sept. 2020, doi:

10.1109/TNSRE.2020.3013429.

[15]. Daqrouq, K.; Ajour, M.; Al-Qawasmi, A.R.;

Alkhateeb, A. “The discrete wavelet

transform based electrocardiographic

baseline wander reduction method for better

signal diagnosis”,. J. Med. Imag. Health

Inf.2018, 8, 1590–1597.

[16]. Lee, J.S.; Heo, J.; Lee, W.K.; Lim, Y.G.; Kim,

Y.H.; Park, K.S. “Flexible capacitive

electrodes for minimizing motion artifacts in

ambulatory electroencephalogram s”. Sensors

2014, 14, 14732–14743

[17]. Yin, C.; Zhou, H.; Li, J. Facile “one-step

hydrothermal synthesis of PEDOT:

PSS/MnO2 nanorod hybrids for high-rate

supercapacitor electrode materials”,. Ionics

2019, 25, 685–695

[18]. Michael R Gold et al. “The effect of TDAB

duration and morphology on cardiac

resynchronization therapy outcomes in mild

heart failure: results from the

resynchronization reverses Remodeling in

Systolic left ventricular dysfunction

(REVERSE) Study”, Circulation, page 112,

2012

[19]. Mei, Y.; Tan, G.Z.; Liu, Z.T.; Wu, H.

“Chaotic time series prediction based on

brain emotional learning model and self-

adaptive genetic algorithm”. Acta Phys. Sin.

2018

[20]. He, S.; Sun, K.; Wang, R. “Fractional fuzzy

entropy algorithm and the complexity analysis

for nonlinear time series”. Eur. Phys. J.-Spec.

Top. 2018, 227, 943–957.

[21]. Hashim, F.R.; Adnan, J.; Daud, N.G.N.;

Mokhtar, A.S.N.; Rashidi, A.F.; Rizman, Z.I.,

”Electroencephalogram noise cancellation

using wavelet transform.”, J. Fundam. Appl.

Sci. 2017, 9, 131–140

[22]. Feng, D.S.; Yang, D.X.; Wang, X. “Ground

penetrating radar numerical simulation with

interpolating wavelet scales method and

research on fourth-order Runge-Kutta

auxiliary differential equation perfectly

matched layer”, Acta Phys. Sin. 2016, 65, 23.

[23]. Karnewar, J.S.; Shandilya, D.V.K.;

Tambakhe, M.D. “A study on EEG signal

analysis and EEG databases”. Int. J. Res.

Advent Technol. 2019, 7, 188–195

[24]. Tychkov, A.; Alimuradov, A.; Churakov, P.

“The emperical mode decomposition for EEG

signal preprocessing”. In Proceedings of the

2019 3rd School on Dynamics of Complex

Networks and their Application in Intellectual

Robotics, DCNAIR”, Innopolis, Russia, 9–11

December 2019

[25]. K. Kalimeri and C. Saitis, “Exploring

multimodal biosignal features for stress

detection during indoor mobility,” in

Proceedings of the 18th ACM International

Conference on Multimodal Interaction, ICMI

2016, pp. 53–60, Japan, November 2016

[26]. Srinivasa Murthy et.al, "FPGA

Implementation of high speed-low energy

RNS based Reconfigurable-FIR Filter for

Cognitive Radio Applications", WSEAS

TRANSACTIONS on SYSTEMS and

CONTROLDOI:

10.37394/23203.2021.16.24, E-ISSN:

2224-2856, Volume 16, 2021.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US

WSEAS TRANSACTIONS on SYSTEMS and CONTROL

DOI: 10.37394/23203.2022.17.37

C. Srinivasa Murthy, K. Sridevi

E-ISSN: 2224-2856

339

Volume 17, 2022