Design and Enactment Evaluation of Adaptive Artifacts Removal from EEG Signal Records

One or more factors may be involved in the transmission of a signal. For a structure to be considered in acceptable condition or performance, it must be able to accurately record and provide information about a living structure, and then the signal is referred to as biological. When a CT scan is used, the X-ray absorption patterns can be used to create a spatial pattern, and a scalp electrode can be used to generate voltage. As a medical or biological source transmits, the biomedical signal [1] is considered to be a biomedical signal [2]. Waves in biomedical signals have been observed in a variety of physiological activities, tissues, and organs, ranging from gene and protein sequences to cardiac and brain rhythms [3].

Biological signals present a blend of signal and noise, necessitating careful observation. Noise, stemming from various sources such as equipment (sensors, amplifiers, filters), or unrelated asynchronous signals, can obscure valuable data [4]. Medical signal processing emerges as a vital technology aimed at extracting meaningful insights from biological signals. Employed by biologists and medical professionals alike, biomedical signal processing facilitates the exploration of novel biological phenomena and the identification of specific disorders. Central to this process is the accurate quantification of the signal model and its components, alongside the essential task of noise removal. Through meticulous signal processing, the objective is to delineate the underlying signal from the interfering noise, thereby enabling clearer interpretation and analysis of biological data. In Biological signal processing [5], the primary goal is to reduce background noise. Noise can be generated because of the interference between instruments and electrical lines. Subsystems that interact with the body in complex systems are also contained inside a biological subsystem. recordings; however, the process fails to entirely separate the ocular artefacts from the EEG signal, particularly when the amplitudes are similar. As a result, this study contributes to the development of new methods [6].

A new wavelet-based technique. W-ICA, which shows the wavelet scheme's performance capability, makes it easier to remove artefacts through the use of Haar wavelet with an ICA model. Experiments showed that the intended work is better than it was previously thought to be [7].

There are a variety of variables that limit the diagnostic procedure, including the inability to detect specific characteristics in signals and the limitations of human extraction skill. This manual analysis also suffers from human faults. These mistakes occur as a result of the decision-making process being weakened by weariness and subjectivity [8-10].

First, sensors are used to identify the physical characteristics of the biomedical system. An ECG signal, for example, measures the electrical activity of cardiac muscles and is used to determine the heart's functioning qualities. Filtering and pre-processing are the next steps. After the sensor has recorded the biological signal, this must be done. There may be unwanted noise in the biological signal, thus pre-processing and filtering are necessary. For example, the electrical impulses produced by the respiratory system are the primary source of ECG signal interference and noise [11].

Electroencephalography (EEG) is a method used in neurophysiology for measuring brain electrical activity through the use of electrodes attached to the scalp. These aberrations are a result of the mingling of EEG data with other physiological signals. Electrical cardiograms, Electro-oculogram, and electromyograms are all examples of such artefacts.

The medical field faces significant challenges in eliminating artefacts from EEG data, prompting the development of innovative solutions. This study introduces a learning algorithm based on FLM optimization for NN-enhanced adaptive filtering, aimed at mitigating artefacts effectively.

Initially, the adaptive filter is employed on the EEG signal to determine optimal weights using two renowned optimization methods: the Firefly (FF) algorithm and the Levenberg-Marquardt (LM) algorithm. These methods converge at the neural network (NN), which plays a pivotal role in achieving adaptive filtering by computing appropriate weights. Consequently, the proposed filtering approach enhances artefact removal capabilities.

Subsequently, the model's performance is evaluated using metrics like Signal-to-Noise Ratio (SNR), Mean Square Error (MSE), and Root Mean Square Error (RMSE) following adjustments to hidden neurons. These evaluations provide insight into the efficacy of the developed model in artefact reduction and signal enhancement.

By integrating FLM optimization with NN-enhanced adaptive filtering, this study offers a promising framework for addressing artefacts in EEG data. The utilization of established optimization algorithms coupled with neural network architecture showcases a robust methodology for improving EEG data quality and clinical interpretation.

It is generally accepted that adaptive noise cancellation can help get rid of artefacts. The fundamental schematic depicting the realisation of adaptive noise cancellation is shown in Figure 1. More so, we get our first data from the input source S(t). After that, we get the subsequent input from the source denoted by A(t). Here, the noise at issue pinpoints the place where EMG, EOG, and ECG artefacts first emerge. The interference signal I(t) is generated when the noisy source signal of the artefacts signal is "acquiesced" by unanticipated non-linear dynamics. Then, the clear signal adds to the interference signal to create the primary input signal, represented as follows:

$p\left( t \right)=s\left( t \right)+I\left( t \right)$    (1)

image004.png

Figure 1. The proposed algorithm

The signal at the input, denoted as S(t), and the signal at the output, denoted as I(t), are the two terms in Eq. (1). Subsequently, adaptive filtering is applied to the signal generated by the noise source in order to obtain the filtered output. The interference signal, which is the result of nonlinear dynamics, is remarkably similar to the output of the producing filter. Retrieving clean EEG data is the primary focus of noise cancellation. In addition, the filtered response is reduced relative to the original input (specified by Eq.) in order to retrieve the clean EEG data (2).

${{S}_{o}}\left( t \right)=S\left( t \right)+I\left( t \right)-F\left( t \right)$                 (2)

When analysing the model in terms of time series prediction, the NARX model works exceptionally well. Typically, a time series’ reaction is predicated on a signal whose values are sequentially correlated with both the input signal and the preceding data. The method is called NARX if and only if the input to the system can be uniquely identified as a computable quantity. NARX uses a three-vector-layer structure, with the hidden layer serving as the input and the output layer serving as the final layer.

Here, we have an input vector, a delayed output vector, and another input vector all making up the input layer of our data structure (delayed). After NN is applied, the resultant vector is L(n+1).

$L\left( n+1 \right)=f(L\left( n \right),\ldots ,L\left( n-{{D}_{L}} \right);V\left( n \right),\ldots ,~V\left( n-{{D}_{V}} \right))$           (3)

It was the input vector L(n) that was being referred to in Eq. (3). The delayed values of the regressed output vector are then represented by the notation L(n-1), L(n-2),..., L(n-D_L). V(n), V(n-1), V(n-D_V) is then the representation of the delayed input vector's sequence. The NARX-NN procedure begins with the assignment of weights between the hidden units and the input vector and between the hidden units and the regressed output vector.

In order to improve optimization performance, this model employs a hybrid method based on a combination of the FF algorithm and the LM algorithm. While attempting to implement the NARX-NN method, the FF approach is initially considered. As such, the blinking habits of the FF are necessary for the FF algorithm. In the FF model, the feasible solution vector is calculated from the coordinates of each FF. Brightness changes the position of FF. Any FF that can't pin down the source of the brightness just picks a number and moves it. Here, the brightness value is determined by the solution to an optimization problem. FF's unpredictable motion increases processing cost and slows down solution convergence. As a result, the firefly method produces non-feasible solutions and has a slower convergence rate.

This study employs the LM algorithm alongside the FF algorithm to resolve the problems associated with the convergence of the optimization concept. This is how the object's location is determined, which is necessary for reconciling theoretical and experimental results. In addition, the LM procedure collaborates with the Gauss-Newton model to function on the steepest descent clustering. The learning process is efficient and quick in the model as well. As a result, the hybrid learning method proved to be more effective in rapidly training the model.

Figure 2 depicts the NARX NN model, employing a learning technique to integrate three distinct weights: exogenous input, regressed result, and exogenous/regressed output. Initially, the LM and learning algorithm receive a randomly generated input vector. Once the optimization principle is finalized, both LM and FF learning models produce weight vectors. This approach ensures comprehensive incorporation of input factors, enhancing the model's adaptability and performance. The vectors are as follows:

${{W}^{LM}}={{W}^{L1}}+{{W}^{L2}}+{{W}^{L3}}+\ldots +{{W}^{Lk}}$              (4)

${{W}^{ff}}={{W}^{f1}}+{{W}^{f2}}+{{W}^{f3}}+\ldots +{{W}^{fk}}$                  (5)

image009.png

Figure 2. The proposed Hybrid NARX-NN scheme

The following steps are used for atomised EEG artifact detection and removal:

1) The population size is determined (along dimension). Additionally, both the beauty and intensity factors are initialised with made-up values. Finally, the weight vector is used to seed the random FF.

2) The entails developing the fitness function.

3) Compare different FFs based on their brightness levels. Every FF's location is updated based on the luminosity of its nearest neighbouring FF.

4) Determine the intensity, attractiveness, and fitness values.

5) The global solution or the best solution at the moment is determined by ranking fireflies.

6) Keep this pattern going till the maximum number of iterations is reached.

7) Formulate the value of error among the current value and the expected value.

8) The LM model is then used to initialise the weight vector in the NARX NN based on the magnitude of the hidden weights.

9) Calculate the sum squared error.

10) The LM algorithm equation is used to update the input weights.

11) If there is a minimal trail-weighed function rather performance index to update the weight according to the equation below.

12) Measuring the Learning Rate at a time when the Performance Index is lower than the Trail Function.

13) The LM formula is used to adjust the previously arbitrary weight.

14) Generate the gradient matrix from the Jacobian matrix.

15) The outcome of the LM method is represented by using the weights from the input vector and the updated weights.

16) The LM algorithm, an error value is calculated by subtracting the error from the ground truth value and applying the new weight.

17) The results from the FF model and the LM model are used to generate the final weight vector. The examination of the error value of the LM and FF model is important for moving forward with the optimization process. The LM algorithm's output is only relevant to the weight vector if its error value is smaller than the FF algorithm's error value (ELM<EFF).

Concurrently, the weight vector is selected from the FF algorithm's output if its error value is smaller than the LM algorithm's error value (ELM>EFF). When this happens, the current error value is less than the previous error value (E_t+1 E_t), and the damping factor value is at its lowest. When the current error is larger than the previous error (E_t, E_t+1), the damping factor is increased. Last but not least, the FLM learning algorithm's error value comparison yields the weight vector.

A. The database applied:

The authentic signal used in this study came from the Physionet database. Additionally, we conducted an analysis of the proposed method's efficiency. Various artefacts, including Electromyographic (EMG), Electrooculogram (EOG), and Electrocardiographic (ECG), were introduced into a genuine signal to show the presented method. Real-time applications were used to test the created model, which was then implemented in MATLAB. The CHB-MIT scalp EEG database is used as the input database. In this case, each of the first five signals downloaded lasted one minute.