A Greedy Optimized Intelligent Framework for Early Detection of Alzheimer's Disease Using EEG Signal

Recent researchers have been drawn to the analysis of electroencephalogram (EEG) signals in order to confirm the disease and severity range by viewing the EEG signal which has complicated the dataset. The conventional models such as machine learning, classifiers, and other mathematical models achieved the lowest classification score. The current study proposes to implement a novel deep feature with the best solution for EEG signal analysis and severity specification. A greedy sandpiper-based recurrent neural system (SbRNS) model for predicting Alzheimer's disease (AD) severity has been proposed. The filtered data are used as input for the feature analysis and the severity range is divided into three classes: low, medium, and high. The designed approach was then implemented in the matrix laboratory (MATLAB) system, and the effectiveness score was calculated using key metrics such as precision, recall, specificity, accuracy, and misclassification score. The validation results show that the proposed scheme achieved the best classification outcome.


Introduction
AD is a neurological malfunction identifed by debilitation of analytic functions and amnesia [1]. Tis is mainly due to the unusual increase of protein around brain cells [2]. AD is often considered as a usual aging process and afects older people more than younger ones [3]. In humans, AD signs are increased confusion and loss of learning ability and memory [4]. Generally, AD is categorized into three stages based on its symptoms and efects [5]. In the beginning stage (mild AD), the sign of AD is most commonly amnesia [6], which does not change a person's daily life. Te next phase is considered moderate AD [7], which is recognized by increased confusion and difculty in learning [8]. In this phase, patient's dependence on the other people increases [9]. Te third stage is severe AD, which is characterized by the entire debilitation of individual [10]. Te early identifcation of AD is important in the case of mild AD and moderate AD [11,12]. Detecting this disease is quite difcult because signs like amnesia are often considered as usual aging signs [13]. Moreover, the imaging and signal analysis system has been introduced in the digital feld to analyze the AD severity range with the help of artifcial intelligence (AI) [14,15]. Hence, several neural models and classifers were implemented to fnd the AD severity rate in the early stages without the requirement of high resource usage [16].
Generally, disease detection is done by large testing and eliminating the other disease cases [17]. Neurological testing, physiological and psychological examination, and blood test help to detect AD [18]. Te basic Alzheimer's disease detection with AI is represented in Figure 1. Perform brain scans, such as computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET), to support an Alzheimer's disease diagnosis or rule out other possible causes for symptoms. Psychological testing such as mini-mental state testing and image techniques based on AI were developed to detect AD [19]. Still, they faced several demerits such as inaccurate disease classifcation, high error rate, and so on [20]. Tus, the researchers paid attention to AD detection techniques based on the analysis of EEG signals [21]. Here, the brain's reading is recorded using electrodes directly by 10-20 electrode systems [22]. AD diagnosis based on EEG signals provides early detection and accurate disease classifcation [23].
Recently, many researchers tried to detect AD in the early stage by identifying changes in the synchrony of EEG signals [24]. However, the diagnosis is not accurate in the case of single synchrony [25]. In addition, for the EEG signal-based AD severity analysis, diferent and several brain samples are taken from one person [26]. Tis has helped to fnd the brain neuron movements based on their actions and emotions. Te EEG-based AD detection technique is used as a tool to diagnose AD early for a huge number of people [27].
To overcome the demerits faced by these techniques, a highly efcient and accurate AD detection was presented in this paper using EEG signals [28,29]. In addition, a novel deep network has been executed by incorporating the optimal solution for detecting disease range. Also, in the designed model, the noise removal function is executed in the hidden phases that have aforded the fnest noise feature tracking and neglecting. Te present noises were eliminated properly in the fltering stage, so the algorithm complexity was minimized, and the classifcation score was maximized. Finally, a comparison assessment was conducted to check the improvement score in analyzing the AD severity range.
Te current research is organized as follows. Section 2 describes the existing works in detail. Te problem related to analyzing the AD EEG signal is defned in Section 3. Te proposed scheme and the solution of the problem are mentioned in Section 4. Te proposed model's outcome is discussed in Section 5, and research arguments are concluded in Section 6.

Related Works
Some of the recent works related to AD detection are described below.
Saf and Saf [27] developed a technique by using Hjorth parameters to diagnose AD in its early stage. Tis model uses diferent fltering techniques such as empirical mode decomposition and discrete wavelet transform to flter the errors in the brain signals. Also, the K-nearest neighbor algorithm was applied to classify the diseases. In past decades, diferent neuropsychological testing based on magnetic resonance imaging (MRI) was developed to diagnose AD. Tus, Murugan et al. [28] developed a MRI-based convolutional neural network model to classify and detect AD accurately. Tis model is highly efcient in detection and classifcation of diseases. But, they did not provide accurate AD detection in case of a large population.
Te interface between the features of EEG signal and AD is not clarifed in existing detection models. Terefore, Li et al. [29] presented a novel technique to establish the relationship between AD detection and EEG signal which presents latent parameters in 3D phase and also provides high classifcation efciency. However, the computational cost is high in this model. Deepa and Chokkalingam [30] presented an optimized VGG-16 framework to classify AD using arithmetic optimization which uses MRI-based images for disease classifcation. Here, the AD was classifed as mild, moderate, and severe. Moreover, the computational complexity and cost are reduced in this technique. However, the experimental outcome of this technique for a large population does not provide efcient classifcation.
Recently, EEG signals have been widely used in medical felds to diagnose the disease in its early phase. Hence, Cejnek et al. [31] designed a model to detect mild AD based on EEG records. Tis technique uses an adaptive flter to preprocess the recorded EEG dataset. In predictive function, the linear NN technique with gradient adaption is used. Tis technique is validated by testing the data of AD patients. However, it does not attain high sensitivity and accuracy.

AD System Model and Problem Description
Usually, the disease forecasting process using signaling data is more complicated because of the feature variations in the minute level. Hence, the normal classifcation and signal analysis model has required more resources to fnd the feature variation and to identify the disease signal. In many  cases, the algorithm complexity has been recorded because of signal analysis delay. Tese issues have motivated this work toward presenting an optimized deep neural system for EEG signal analysis and disease signal feature classifcation. Te disease that was considered for validating the developed model is AD [32].
Analyzing the brain signal is the foremost topic in the medical industry to detect and treat diferent kinds of diseases. Especially, analyzing the EEG signal and identifying the disease features and their severity range are highly complicated because of the noisy unbalanced data. Te problems analyzing in EEG signal are described in Figure 2.

Computational Intelligence and Neuroscience
A novel sandpiper-based recurrent neural system (SbRNS) has been introduced for predicting Alzheimer's disease in an earlier stage using EEG signal. Primarily, the EEG signals were collected and trained to the system and then the preprocessing function was activated to flter the present noise in the trained datasets. Consequently, the feature analysis was performed, and the disease features were extracted and then the severity range was measured. Finally, the key parameters were estimated and compared with the other schemes.

Proposed Methodology
Te proposed architecture is illustrated in Figure 3. In the feature analysis steps, the required features were traced and mined based on the specifcation of 0 th class and 1 st class.
Here, the signal pixels were analyzed and traced for the 0 th class features and then the extraction was done. Consequently, the disease specifcation and severity range estimation were performed. Finally, the comparison analysis helped to measure the performance enhancement by the developed scheme.

Proposed SbRNS Design.
Te planned model has fve diferent phases that includes data importing, error fltering, classifcation module, optimal phase, and output phase. Te data importing process is executed in the training phase of the initial layer. Noise elimination function is performed in the second phase hidden layer. Feature analysis and the severity level identifcation are processed in the third classifcation phase, and then the classifcation phase parameters are tuned by the optimal layer. In the fourth layer, the optimal phase sandpiper ftness is utilized. Finally, the severity forecasting outcome is recorded in the ffth layer, output phase.
Te proposed scheme functions on the basis of the sandpiper approach [33] and recurrent neural scheme [34]. Tese layers are elaborated in Figure 4. Here, the reason for incorporating the sandpiper function in the recurrent neural classifcation phases is to earn the fnest severity forecasting score.

Preprocessing Model.
Te function preprocessing is designed to eliminate the training faws from the trained sets. Tis process tends to gain the fnest accurate severity analysis score and less computational complexity. Moreover, the data importing process has been processed by (1). Here, the AD brain EEG signal dataset is determined as b e , dataset training function is determined as F(b e ), and the n number of EEG signals is described as 1, 2, 3, 4, . . . , n: Te data training function is performed in the input phase of the novel SbRNS. After training the data, the function preprocessing is activated in the hidden phase to eliminate the noise features.
where J represents the preprocessing variable, q represents the normal features in the database, and the noise features are described as a. Te noise features that are present in the trained database are removed by the preprocessing process that is represented as b e (q, a) − b e (a).

Feature Analysis.
Te output of the preprocessing layer is earned as the error-free data. Ten, those data are considered as the input of the feature analysis process. Forecasting the disease or severity range in the entire data is difcult and has maximized the computational complexity. Considering this problem, the process feature analysis has been executed to track and extract the present features in the trained sets.
where the AD signal features are described as Z and F * determines the feature analysis variable. Te process of feature extraction is indicated in (3). Here, Z represents the 0 th class. In this, the 0 th class features are traced and mined. Also, the unwanted features y are neglected because y falls under the 1 st class. In (3), max denotes the maximum possible iterations. Tis feature analysis process is executed by the ftness of sandpiper.

Classifcation Module.
After the feature analysis function, the severity range of the AD has been predicted in the form of low, high, and medium. Te analyzed features Z are categorized under the 0 th class and then AD range is specifed as high severity; if Z falls under the 1 st class, then it is medium severity. If Z is not falling under classes 0 and 1, then it is specifed as low severity.
where D * represents the classifcation parameter. Hence, based on the "if" condition, the severity of the AD is specifed as high and medium. Te classifcation process is valued in (4). Te defned formulation of the novel SbRNS is represented in Algorithm 1. By processing each step of SbRNS algorithm, the EEG signal was analyzed, and the severity score was forecasted. In addition, the fow of the steps in executing the present scheme is elaborated in Figure 5.

Results and Discussion
Te discussed model is executed in the MATLAB programming platform and processed in Windows 10 platform. Initially, the EEG signal of AD is gathered from the standard site and taken as the input of the MATLAB system. Te parameter specifcations are indicated in Table 1.
In addition, the EEG signal processing framework functions in the ratio of 75 : 25, i.e., training 75% and testing 25%. Te class that has been considered in the severity specifcation is low, medium, and high.
Te ability of the model in specifying the severity range is described through the validation, and the graphical representation is given in Figure 6. Te validation graph has reached 100% accuracy range that described the exact prediction, and the loss is very low.
To measure the working range of the proposed design, a few diferent test samples are considered that are normal Computational Intelligence and Neuroscience signal, low AD signal, medium AD signal, and high AD signal. Here, the normal signal is also considered for this experiment to check the working rate of the presented design. Te results of the forecasted normal brain signal are shown in Figure 7.
Diferent EEG signal samples were tested. Te outcomes of medium AD, high AD, and low AD are indicated in Figures 8-10.

Performance Validation.
Recall validation is an evaluation of the sensitivity range in classifying the severity categories. Also, it has provided the stability range of the executed model in classifying the disease range in the presence of false prediction and actual prediction. Tus, if a system has recorded a good recall measures, it is better in severity specifcation. Te recall is was formulated as recall � true positive false negative + true positive .
To measure the positive values in the severity prediction, the precision metrics were valued. Hence, this has aforded the actual forecasting outcome in severity range specifcation. Te precision parameter is defned as Precision � true positive false positive + true positive .
Te exactness of severity specifcation is determined as accuracy. Hence, the gained accuracy has verifed the robustness of the designed approach in specifying the Alzheimer severity range. Te accuracy is defned as accuracy � exact forecast total forecast .
Te mean efciency of the severity classifcation process has been found in the F-measure validation. Te summation and product of the recall and precision have described the Fscore, which is indicated as Te metrics specifcation was estimated to assess the negative scores in predicting the AD severity range as follows: specificity � true negative false positive + true negative .
Here, the specifcity score was determined by only taking into account negative classifcations, i.e., a true negative and  Computational Intelligence and Neuroscience a false positive. Hence, the overall validation of the developed novel SbRNS was described in dual phases that are with and without incorporating the sandpiper ftness solution in the recurrent model, which is graphically defned in Figure 11. Before applying the sandpiper model in the recurrent system, the recorded specifcity score was 98%; after adding    Computational Intelligence and Neuroscience the sandpiper function in the classifcation phase of the recurrent system, 99.5% of recall was recorded.

Comparison Assessment.
Te conventional schemes that have been taken used to validate the designed model are the extreme intelligent model (EIM) and K-nearest model (KNM) [35]. Besides, the presented approach is measured in dual phases that are before and after applying the optimization model.
Te EIM has attained a precision score as of 99.69% and specifcity of 99.6%; the approach KNM has earned a specifcity rate as of 95.7% and precision of 96.7%. Moreover, before incorporating the sandpiper ftness, the presented scheme SbRNS has recorded the specifcity score as of 99.3% and precision of 99.6%. After applying the ftness solution of sandpiper function model, 99.8% precision and 99.9% specifcity were reported.
Te F-measure and accuracy rate gained by the scheme EIM are 98.9% and 98.7%, respectively; the approach KNM has gained an accuracy score of 98.1% and F-score of 98.3%. In addition, before performing the sandpiper function, the recorded F-score is 99% and accuracy is 99%. Furthermore, after executing the sandpiper model, the recorded F-score is    Computational Intelligence and Neuroscience 99.5%, and the exactness score is 99.8%. Te statistics before and after the optimization are shown in Figure 12. By comparing the recent scheme, the present approach has earned the best outcome in classifying the disease severity range. Tis has verifed the efectiveness of the developed model in EEG signal analysis. In addition, after incorporating the sandpiper, the highest performance was recorded. It is clearly seen that the need of the optimal model in severity specifcation is highly recommended. Te assessment of accuracy and F-measure is revealed in Figure 13. Also, the overall comparison assessment is described in Table 2.

5.
3. Discussion. Te validated parameters have described the profcient score of the designed model. Te working performance of the proposed model is described by estimating all classifcation metrics that are precision, specifcity, Fscore, accuracy, recall, and error validation that are detailed in Table 3.
In this case, the metrics error was examined in order to examine the falling score of the designed approach. As a result, the earned misclassifcation score is only 0.1% percent, which is a very low and negligible state. Tis described the efcacy of the devised scheme.

Conclusion
A novel SbRNS has been implemented in the MATLAB framework for the classifcation of AD severity ranges using EEG signals, and the severity analysis procedure was carried out. Te current SbRNS has a maximum severity forecast range of 99.8% when compared to other conventional approaches, and the classifcation rate was increased by 1%. Te recorded sensitivity rate is 99.9%. In comparison to previous approaches, the specifcity score has increased by 0.6% in comparison to earlier methods. Te novel SbRNS was able to attain an extremely low error rate of 0.1%. Furthermore, when compared to the conventional recurrent neural model before applying the sandpiper function, the present SbRNS improved performance by 1% to 2%. As a result, the designed method is the a necessary scheme for EEG signal analysis and severity range prediction; however, it is not applicable to other application datasets. Creating a tuned hyper-parameter model will provide the best classifcation and feature analysis results in the future.

Data Availability
Te data used to support the fndings of this study are available online.

Conflicts of Interest
Te author declares that there are no conficts of interest regarding the publication of this paper.