A design of bat-based optimized deep learning model for EEG signal analysis

Depression is a mental illness that negatively affects a person’s thinking, action, and feeling. Thus the rate of depression is identified by analysing Electroencephalogram (EEG) signals. Because of noise, the problem of classifying depression rate has some issues, such as low accuracy and required high training time. In this research work, a novel Bat-based U-NET Signal Analysis (BUSA) architecture is developed to estimate the patient’s depression rate with an EEG dataset. This technique involves pre-processing, feature selection, feature extraction, and classification. After the data training, the pre-processing function was activated to neglect the noise in the brain signal. Hereafter, the noiseless Signal is used for the further process. Here, the bat algorithm mimics the behaviour of the bat’s frequency and loudness, increasing the accuracy of prediction and classification. This fitness function is upgraded in the U-NET classification phase. Moreover, the brain signal’s feature selection and depression rate were classified using the bat fitness that has helped to gain the desired output. Finally, the performance metrics of the proposed BUSA technique are compared with other existing methods regarding the accuracy, AUC, precision, recall, and power. The proposed BUSA framework attained a high accuracy rate of about 99.64%, a maximum precision level of approximately 99.98%, a high recall rate of approximately 99.95%, and a high AUC of approximately 99.2%. The developed framework has attained better results in classifying depression rates.


Introduction
Generally, EEG is an electrical activity used to record the brain on the scalp [32].It is used to measure the electrical Signal of the brain with the Brain-Computer Interface (BCI) [12].BCI is a technology that enables the communication between the brain and an external device, such as a computer or a robotic system, without requiring physical movement.BCI systems typically use non-invasive or invasive techniques to record and interpret electrical signals generated by the brain, also known as brain waves.The potential applications of BCI technology are numerous.They include helping people with disabilities communicate or control their environment, enabling them to control prosthetic limbs, and assisting in treating neurological disorders such as Parkinson's disease or epilepsy.BCI technology is still in the early stages of development, but it holds great promise for improving the quality of life for many individuals.Moreover, it will deliver mental activity-related information, and the EEG signals are unique, non-stationary, and weak because they contain much noise [25,37].So, EEG signals are the most crucial task to process and analyze EEG signals [35].It includes four processing jobs: feature selection, extraction, classification, and preprocessing [27].Initially, the data is collected, and then in the pre-processing module, the null/error data are removed from the data.Then the meaningful features were detected, and then the features were extracted.At last, the extracted features were classified.The basic process of signal analysis in EEG is detailed in Fig. 1.
Furthermore, it deals with the classification and feature extraction of EEG signals and transfers the gained EEG signal to the feature vector with the help of the classifier [40].For better EEG signal analysis, feature extraction to a certain EEG category is needed [21,43].Moreover, feature extraction extracts relevant and meaningful data or information for assisting good investigation [7,36].It also starts with raw signal data from EEG and creates the values of the feature vector [11,34].However, signal processing of EEG contains Autoregressive (AR), Fourier transform, entropy and discrete wavelet transform [14,26].It provides a wide range of EEG signals which is useful to the time-frequency transformation and contains various sub-bands to attain accurate information [19,29].Moreover, the AR technique includes forecasted variable interests that are helpful to linear combinations of past variable values [8,9].Additionally, the coefficient of the AR technique is useful to the feature vector of the BCI system.Still, it is more difficult to gather the transient features of EEG signals because of noisy Signals and error rate [13,38].Nonetheless, many techniques were developed for feature extraction of EEG signals, but it has the issue of information loss [17].The main problems of the EEG signal are noisy data, miss classification, and loss of information [4,33].There are a lot of techniques introduced to overcome these issues, such as the EMD-AR technique [41], EEG-AR model [30], convolution neural network (CNN) [10], and so on, but has the issues of noise in the Signal, loss of information, improper classification [3].So, this paper has developed a novel AR and deep learning (DL) algorithm for the modelling and analysis of EEG signals.It will solve the problem of information loss noise error.
Bats use echolocation to navigate and locate prey, and their auditory system is highly adapted for processing complex sound signals.This ability of bats to process complex acoustic signals has inspired researchers to develop optimization algorithms based on bat behavior, known as Bat Algorithm (BA).In recent years, there has been growing interest in using deep learning models for analyzing EEG signals.EEG signals represent the brain's electrical activity and can be used to diagnose various neurological disorders.However, designing an optimized deep learning model for EEG signal analysis is challenging due to the high dimensionality and variability of the data.One approach to addressing this challenge is to use the Bat Algorithm to optimize the hyperparameters of the deep learning model.The bat-optimized algorithm is a metaheuristic algorithm inspired by the behavior of bats, where they use echolocation to locate prey and avoid obstacles.The BA is particularly useful for optimizing complex, high-dimensional problems, such as those encountered in deep learning.By using the Bat Algorithm to optimize the hyperparameters of the deep learning model, it is possible to achieve better performance on EEG signal analysis tasks.It can lead to more accurate diagnoses of neurological disorders and better patient treatment outcomes.
The arrangement of this article is structured as follows.The related work based on EEG signals from 2018 to 2022 is detailed in Section 2, and the system model and problem statement are elaborated in Section 3. Also, the process of the planned methodology is described in Section 4. Finally, the achieved outcomes of the developed technique are mentioned in Section 5, and the conclusion is detailed in Section 6.

Related works
A few recent literature surveys based on EEG signal from 2018 to 2022 is detailed below, Zhang et al. [41] have proposed an EEG-based Empirical Mode Decomposition (EMD) and Auto-Regressive (AR) technique for classifying emotional data.Initially, the EMD technique is employed and decomposed to the EEG signal; it will calculate the features based on the function of intrinsic mode.Thus the developed technique attains 86.28% in emotion recognition, but noise in the Signal is high.
EEG is mostly used as a tool for assisting in the diagnosis of epilepsy.It has recorded the epilepsy diagnosis of the patient.Ouyang et al. [30] designed an EEG AR model to predict the error in the patient through epilepsy control.Thus the technique assists the diagnosis using epilepsy without the discharge of epileptiform.However, the classification rate of patient diagnosis is less because of the noisy signals.
Zhang et al. [42] have developed an AR technique and wavelength packet rottenness to classify the Signal based on the brain system.Moreover, the AR coefficient calculates the features depending on the assembling, and the sub-band has decomposed the EEG signals.Thus the computed feature extraction is fed into the vector machine to classify the EEG signal.But the process of classification takes more time to categorize the Signal.
Generally, EEG and Brain-Computer Interface (BCI) are used to increase the classification's performance.Dose et al. [10] have proposed a convolution neural system for generalized learning features.It will reduce the dimension, and a fully connected layer is helpful for classification.Thus the developed technique attains better results in classification accuracy at 80.38%.But it has various dimensions, so some misclassification has occurred.
Zhang et al. [42] have designed a convolutional recurrent attention technique for encoding the EEG signals high-level representation and generalized pattern with good performance.Moreover, the recurrent attention technique explores EEG signals of temporal dynamics, mainly focusing on discriminative temporal periods.But it has the problem of error noise in the Signal, so it is more difficult to attain better outcomes in EEG signal.
Ajagbe et al. [1] study the effectiveness of the DL techniques in bio-inspired by six performance parameters.Using CNN, the bio-inspired database integrated the beetles bee and morder hornet on the Matlab platform.The proposed model's outcomes are superior to other DL-based techniques concerning sensitivity, computational duration, accuracy, specificity, and precision.However, bio-inspired databases may need a huge amount of labelled data.
With the help of the advancement of the DL-dependent model, it is possible to develop a prompt and accurate technique for eye state classification issues.So, Larabi-Marie-Sainte et al. [23] proposed a novel compact bat-optimized algorithm (CBA) with a DLbased method for bio-medical EEG eye state categorization (CBADL-BEESC) method to classify the EEG eye state presence.Using the ALexNet method, the proposed CBADL-BEESC extract the feature extraction process.The extreme learning tuning auto encoder (ELM-AE) uses CBA to classify the EEG signals.Yet, this model is only applicable to the limited dataset.
In order to detect an optimized feature selection method which may classify the state of mental stress while improving the overall performance in the classification process.Hence, Hag et al. [15] introduced a hybrid feature detection model by maximum relevant and minimal redundancy with particle swarm optimized algorithm (PSO) and support vector machines (SVM) (mRmR-PSO-SVM) to choose the optimized feature subset.The presented method attained reasonable performance improvement compared with other methods.But, their computational complexity may occur.
Kundu and Ari, [22] proposed a brain-computer interface model dependent upon the P300 signal for character recognition.In order to identify the word/character without any movement of muscle from the EEG signal, a P300 is utilized.For P300 detection, sparse autoencoder (SAE) and stacked sparse autoencoder (SSAE) dependent feature extraction techniques are introduced.The SSAE method extracts a maximum range of information about the dataset compared with SAE.Two datasets are utilized to test the proposed method, and the results validate that the proposed model outperformed others.But, it required a high-dimensional dataset.
Polat and Nour, [31] proposed SVM for EEG signal classification for five cases comprising healthy, open eyes, closed eyes, an epileptic seizure, and from the tumour region has been conducted.The SVM classifiers contain various kernel functions comprising medium, linear, and cubic Gaussian.The proposed model attained a reasonable accuracy rate in classification performance compared with other recent convolution techniques.However, this system is limited generalizability.
Khare and Bajaj [20] proposed an optimal tunable Q wavelet transform (TQWT) using various optimized algorithms for the adaptive selection of decomposition metrics.The mean square error of decomposition is decreased using an optimized algorithm.To decompose the signals to sub-bands, optimized decomposition matrices are utilized.The proposed technique attained an accuracy rate of about 96.14% in the classification process.Anyhow, there is a lack of interpretability.
Azad et al. [5] proposed a linked open data search engine for searching semantic web documents depending on the user queries.Two semantic search methods, forward and backwards, are presented for triple searching.To enhance the search results, domain and triple-ranking methods were utilized.The results show the performance of the proposed model.Anyhow, this proposed technique attains a low accuracy level.
Azad et al. [6] proposed a query expansion method to tackle the low retrieval efficiency by integrating the three weighted techniques depending upon tf-idf, kNNdependent cosine similarity, and correlation level.The proposed novel technique improves the mean precision level by up to 25.89% and the geometric mean precision range by up to 30.83% in the FIRE dataset.Comparative assessments were carried out to show the superiority of the proposed model.Yet, it is only effective for this dataset.
To reduce the unclean noise filtering, large-size data, and lower signal-to-noise ratio (SNR), Liu, [24] integrated the weight-splitting methodology with the backpropagation (BP) neural model called the improved BP neural model algorithm.To overcome the extracting issue, this research utilized the non-linear mapping function of the basic BP neural model and computed the low-weight elements by integrating the PSO.Hence it enhanced the performance of the extraction of the entire BP system.The proposed algorithm reduced the SNR and unclean filtering in EEG signals.But, data acquisition is difficult.
In order to degrade the noise in EEG signals of depression, Kaur et al. [18] proposed novel hybrid methods of discrete wavelet transform (DWT) and wavelet packet transform (WPT) with variational mode decomposition (VMD) called DWT-VMD and WPT-VMD.In this system, the dataset has decomposed the signal to different elements then WPT-VMD is utilized to clear the noisy contents in the artifactual elements.This novel technique decreases the error rate and SNR and improves accuracy.Yet, there is complexity in selecting the metrics.
Akbari et al. [2] introduced a state of art technique depending upon centered correntropy (CC) and empirical wavelet transform (EWT) for categorizing the normal as well as depressed EEG signals.Using EWT, the EEG signals were decomposed, and as the discriminated feature CC of rhythms was trained and imported into kNN and SVM classifiers.These techniques attained better mean classification accuracy rate, specificity, and sensitivity.However, computational sources are complex.
The key step process of this research work is described as follows, • The normal and abnormal EEG signal standard dataset was initially taken from a net source and trained in the system.• Consequently, a novel Bat-based U-NET signal analysis model is designed in the MATLAB environment.• Then the present noise in the EEG signal is removed in the pre-processing frame.
• Then the pre-processed Signal enters into the classification layer of the optimized deep learning model.• While the testing process, the sensed patient EEG dataset is trained in the system; by matching the Patient EEG dataset with the normal EEG dataset, the Depression rate is predicted successfully in the MATLAB environment.• Hereafter, the evaluated metrics are validated with other models regarding the accuracy, AUC, power, precision, and recall.2. The EEG signal system is vital in analysing the human depression rate.However, the key demerits of the EEG signal system are less accuracy in depression classification.The automated type is impossible in the EEG signalling framework if the depression ranges are not classified.In addition, if the data is too complex, it reports very few depression classification measures.Also, the conventional model takes more time to complete the process.The gained performance of the developed BUSA framework is compared with other existing techniques such as CRA, CND and AARM.Moreover, the CND technique attained AUC is 98.3%, and the AARM technique gained 87.54% in AUC.Furthermore, the CRA method earns 81.86% in AUC, but the developed technique achieves high performance in AUC at 99.2%.

Proposed BUSA methodology for EEG signal classification
In the beginning phase, the EEG signals are taken from the standard datasets and trained in the system.Then a novel Bat-based U-NET signal analysis (BUSA) framework is designed in the MATLAB environment.In the initial phase, the errors are removed in the pre-processing layers.Hereafter, the error-free data enters the classification layer.Thus the architecture of the proposed technique is shown in Fig. 3.
The fitness of the bat is utilized to specify the depression rate from the trained EEG signal.Finally, the presented model is validated concerning accuracy, precision, AUC, recall, and power.Where, k i is denoted as the rate of input EEG signal and k is considered as the mean value of the input dataset.Moreover, η is represented as pre-processing, and mis regarded as noise present in the EEG signal.Thus the pre-processing helps to attain a perfect classification of the depression rate and will change the raw Signal into a normal neutral signal.

Feature extraction
Then the pre-processed dataset is sent to the feature extraction layer, which will extract the features of the EEG signal with the help of the Band-pass Butterworth filter.Thus, feature extraction happens based on common frequency bands such as theta, alpha, delta, and beta.Moreover, the frequency range of delta was 0.5-4 Hz, and the frequency range of theta was 4-8 Hz.Also, alpha has a frequency range of 8-13 Hz, and the frequency range of beta is 13-30 Hz.The feature extraction's main process is to extract the related features of the depression rate based on time-frequency series from the pre-processed dataset.
Moreover, bat fitness is used in the developed technique for extracting the features of the input EEG signal based on the function of time or frequency and signal energy.It will convert the discrete-time Signal into a discrete frequency sequence (K x ) using Eq. ( 2).
Where m is denoted as the number of samples and h(t) is designated as the Signal.Moreover, p is called a time-domain signal.The signal dimension provides more information about the nature of the system, and the fractional extent is used to estimate the experimental data dimension.Moreover, the correlation dimension redirects the degree of association among state-space points that measure the system's complexity.Thus the probability of issues in the same set is obtained using Eq.(3).
Let Q(t) is denoted as the correlation integral and θ is called as the probability of complexity in the system.Moreover, D is represented as the dimension of the estimated slope value f,k(i), and k(j) is described as the fractional dimension of the EEG signal.

Feature selection
The bat-based U-NET technique is applied to EEG signals to detect patterns indicative of depression.EEG is a non-invasive method of measuring brain activity by recording electrical signals from the scalp.To obtain the depression rate using bat-based U-NET with EEG signals, a large dataset of EEG recordings is needed, including both individuals with depression and healthy controls.The EEG signals are then processed using the bat-based U-NET technique, which can identify patterns of activity associated with depression.One of the major obstacles in using bat-based U-NET with EEG signals to detect depression is the lack of a definitive biomarker for depression.Depression is a complex disorder with a range of symptoms and no single EEG pattern or signal that can definitively diagnose depression.Generally, feature selection is the process of pattern recognition.It will select the features based on the subset.Typically, feature selection enhances classification accuracy performance and deals with high-dimensional data.Thus the difference in the generated new features is compared between depressed and normal patients.While the vector length of each chromosome corresponds to one of the features, it will select the feature, and the feature selection happens based on the k value of samples.At the same time, the k value is less than 0.05 means that the features are selected.Moreover, feature selection is based on relative power, centre frequency, correlation dimension, power spectral entropy, band power, and DFT.The process of the developed framework is detailed in Fig. 4.

Classification
Furthermore, selected features are sent to the fully connected layer and update the fitness function of a bat in this layer.Initially, it identified the rate of depression in the collected EEG signal by comparing normal EEG signals with the help of bat fitness.Next, calculate the distance between the selected and unknown samples using bat fitness obtained by Eq. ( 4), Based on some conditions, it will classify the normal Patient and depressed Patient, while the depression rate is Fr t a > 1 means depressed patient but Fr t a = 1 means normal patient.Thus the designed algorithm of the BUSA framework is detailed in algorithm 1, and the flowchart of the designed BUSA framework is shown in Fig. 5  accuracy, recall, precision, AUC and power.Furthermore, the introduced method classifies the depression rate of EEG signals.Finally, a successive score of the presented model is compared with other models to verify the proficiency score of the designed model.

Case study
Generally, EEG is one of the non-invasive techniques used to monitor brain activity with various classifications and detect brain disorders.Moreover, depression is a prevalent disease it will cause disability.So classifying the depression rate using EEG signal is the most critical task.Thus the developed BUSA framework is implemented in this case study.Initially, collecting the EEG signal dataset [39] of various patients was based on the channel location.the channel location includes F7, Fz, F6, C6, C7, Cz, H6, H8, T2, T5, T3, O2, O1, FP1, O6, FP2, etc., and the channel location is detailed in Fig. 6.
Then the collected EEG signal dataset is sent to the convolutional layer.The pre-processing contains a 1-42Hz Band-pass filter for removing the errors and noise present in the Signal.Additionally, the pre-processed dataset is sent to the feature extraction layer, and the layer extracts the features based on the channel location.Moreover, feature extraction includes correlation dimension, DFT, and band power.Thus the band power contains beta, alpha, theta, delta, and gamma.Also, the frequency range varies based on the channel location.The use of feature selection is choosing important information used for a classifier.Thus the feature selection maximizes the relevant features of the classifier's target.At the next level, the fitness function of a bat in the fully connected layer is updated, which identifies and classifies the depression rate of the patient to the normal patient.Moreover, it classifies the EEG signal based on normal and depressed patients.We have taken 4500 of the dataset which contains 2500 samples.Using the bat fitness function, we have set the threshold value as 7234.If the depression rate is above 7234, it is classified as a depressed patient.If the depression rate is lower than 7234, it is classified as a normal patient.Thus the classified depression rate is detailed in Table 1.
The depression rate is classified based on the channel location of each electrode; also, the channel's area varies while the EEG signal changes.Thus the developed BUSA framework attains better results in classifying depression rates.

Performance metrics
The planned BUSA model is implemented in LABVIEW or MATLAB tool, and the success rate of the designed scheme was analyzed using comparison assessment about the accuracy, AUC, power, recall, and precision.Thus the achieved performance is compared with other existing techniques related to analyzing EEG signals, such as the Classification of Emotion by AutoRegressive (AR) and Empirical Decomposition (CEAED) model [41], Analysis of EEG AR Modeling (AARM) [30], Classification of EEG signals using AR and Wavelet Packet Decomposition (AR-WPD) [3], end-to-end DL method to MI-EEG signal classification (E2E-EEG) [10], Convolutional Recurrent Attention (CRA) technique [42], Classification of Normal and Depressed EEG signals (CND) [2] and Prediction of Major Depressive Disorder (PMDD) [16].

Accuracy
In order to evaluate the performance of the proposed BUSA method, the performance parameter called accuracy is utilized.Accuracy is the ratio of the corrected detection to the total number of detection.The accuracy of the BUSA framework is identified based on the performance of classifying depression rates.Moreover, the accuracy of categorizing depression rates can be expressed using Eq. ( 5), Where, IPis denoted as a true classification of depression rate, INis represented as a true negative classification of depression rate.Moreover, APis expressed as a false positive classification of depression rate and ANis called a false-negative classification of depression rate.The comparison of accuracy with the exciting technique is detailed in Fig. 7.
The achieved accuracy rate of the proposed BUSA technique is compared with other replicas such as CEAED, CRA, CND, and so on.Thus the CEAED and AARM replicas attained 96.28% and 85.17% as accuracy, and the AR-WPD method gained 98.2% accuracy.Moreover, the E2E-EEG and PMDD methods earned 86.49% and 87% accuracy; also, the CND technique achieved 98.76%.Additionally, the developed BUSA technique achieves 99.64% accuracy.

Power
Initially, power is calculated by the quantity of the activity in a certain frequency of the band signal.The coherence among various electrodes is reflected in the degree of connection in the brain region.Thus the power is calculated based on the frequency range and appropriate average frequency for obtaining power values.
The CEAED and AARM techniques gained 128 and 256Hz in power.Also, the AR-WPD replica attained 250Hz in power.Moreover, E2E-EEG and CND techniques achieved 70Hz and 256Hz in power; also PMDD replica earned 500Hz.Thus the developed BUSA techniques reach 40Hz in power.Because of the less power rate, errors noise generated by the EEG signal is reduced, and the comparison of power is shown in Fig. 8.

Precision
Precision is a performance parameter utilized in ML to calculate the accuracy of a binary classification method.It estimates the ratio of true positive predictions among all positive predictions made by the process.The computation of precision (P) is operated to recognize the success of the proposed BUSA technique while classifying the depression rate.In addition, the measurement of the precision rate is obtained using Eq. ( 6) and comparison of precision has shown in Fig. 9.  Thus AARM replica attained 89.98% precision, and the E2E-EEGmethod gained 69.97% precision.Moreover, the CND method reached 98.65% Precision, and the PMDD technique achieved 91.3%.Additionally, the developed BUSA technique achieves 99.45% precision.

Recall
A recall is a metric used to calculate a classification method's performance.It measures the ratio of positive cases that the model correctly detects.Measurement of recall (R) is estimated to classify the depression rate of the developed BUSA technique.Additionally, recall is the term of true positive value to the addition of false-negative and true positive value.Moreover, the recall calculation of the BUSA method was obtained using Eq. ( 7), (7) Re call = IP IP + AN The achieved recall rate of the proposed BUSA technique is compared with other replicas such as AARM, E2E-EEG, CND, and PMDD.Thus the AARM replica attained 81.81% recall, and the E2E-EEG method gained 67.35%.Moreover, the CND method earned 98.47% recall, and the PMDD technique achieved 82.6%.Additionally, the developed BUSAtechnique reaches 99% in the recall, and the comparison of recall with the exciting technique is detailed in Fig. 10.

The area under curve (AUC)
It is the ratio of the measuring capability of a classifier to separate among classes that are also helpful for the summary of the ROC curve.It will provide better performance in the classification of depression rates using EEG signals.Moreover, AUC measures the classified rate depending on the gained absolute values.The comparison of the AUC is shown in Fig. 11.The gained performance of the developed BUSA framework is compared with other existing techniques such as CRA, CND and AARM.Moreover, the CND technique attained AUC is 98.3%, and the AARM technique gained 87.54% in AUC.Furthermore, the CRA method earns 81.86% in AUC, but the developed technique achieves high performance in AUC at 99.2%.

Discussion
This research paper takes the EEG signal dataset from the standard website.Generally, the dataset consists of several wanted and unwanted features and error/null values.These unwanted and error values may increase the computational time and reduce the classification process's accuracy.To increase the effectiveness in classification accuracy with low training time, the novel hybrid BUSA framework integrates the principles of bat optimized algorithm with the U-NET phase.Hybrid two effective DL-based models efficiently remove the error/invalid/unwanted data present in the dataset in the pre-processing layer.Hence, the dataset only contains the wanted features.Then, in the feature extraction phase, the meaningful features are extracted.The extracted features are classified in the classification phase with the help of a fitness function of the bat.Hence, the accuracy of the classification process is increased.Finally, the results are obtained.Reviewing all the metrics in the previous section, the proposed BUSA has gained high results in terms of accuracy, precision, recall, computational time, and AUC.It has revealed the stability of the proposed system in analyzing brain EEG signals.In addition, the reason for attaining the best accuracy is the up-gradation of the bat fitness in the U-NET layer.Usually, the fitness of the bat is an echo sound to find the location.Here, that fitness function is taken into account to track the Signal features to analyze the depression rate that has yielded the finest outcome.The outstanding gained metrics of BUDA framework comparisons are tabulated in Table 2.
Hence, the validation results have described that the proposed BUSA approach has improved the EEG signal classification system.Also, it is capable of performing in all EEG signal classification applications.Moreover, the limitation of this proposed work is design complexity.So, in future designing, the fuzzy set with this proposed model will be attained improved results.

Conclusion
The EEG signal system has contributed to analyzing individuals' brain functions in the medical field.Hence, the depression rate of the brain signal was estimated using any neural approaches or optimization framework.The main drawback behind the EEG signal system is a noisy signal.The noisy Signal can reduce classification accuracy by affording the low depression rate estimation, so in this work, a novel BUSA technique is proposed to classify the depression rate of patients.Here, the classification parameter of the U-NET was tuned by bat fitness to gain the finest results.By the mathematical modelling in MATLAB, the proposed technique achieved an accuracy of 99.64%, recall of 99.95%, AUC of 99.2%, and precision of 99.98%.Hence, the proposed BUSA model has improved the depression rate classification accuracy by up to 20% compared to other techniques.Thus, the robustness of the proposed method was proved.However, the presented architecture may experience a few problems because of the design complexity.It slows down the process in the middle phase of the model.So, there is a need to design another architecture in future work that could overcome these problems of the proposed approach without removing any significant layers and features.So, in future designing, the fuzzy set with this proposed model will be attained improved results.

Fig. 2
Fig.2System model and problem

Fig. 10
Fig. 10 Comparison of recall Several neural-based algorithms were used to estimate the brain signal; EEG has been taken to analyze the brain signal's health.Moreover, the sampling frequency regularized the EEG signal by converting the Signal to 256Hz.Furthermore, feature extraction was done to analyze the Signal based on the activity of the electrical brain.Fourier analysis is a mathematical technique used to represent a complex signal as a combination of simple sine and cosine waves of different frequencies.Wavelet analysis is a mathematical technique that allows the decomposition of a signal into wavelets, which are small waves that are scaled and shifted in time.It is useful in analysing non-stationary signals and will enable to focus on specific signal features at different scales.Using Fourier and wavelet transform, the important features are set.But it has a problem classifying EEG signals because of noise signals, less accuracy, and high execution time.The system model and problem definition are detailed in Fig.
Process of BUSA frameworkWhere, αand φ are denoted as constant, b 0 a is indicated as the depression rate of the normal patient, b t+1 a is represented as the depression rate of the affected patient and Fr t a is called as depression rate classifier. .

Table 1
Classified results

Table 2
Overall performance Performance assessment