Multi-modality NDE fusion using encoder–decoder networks for identify multiple neurological disorders from EEG signals

Abstract

Background:

The complexity and diversity of brain activity patterns make it difficult to accurately diagnose neurological disorders such epilepsy, Parkinson's disease, schizophrenia, stroke, and Alzheimer's disease. Integrated and effective analysis of multiple data sources is often beyond the scope of traditional diagnostic procedures. With the use of multi-modal data, recent developments in neural network approaches present encouraging opportunities for raising diagnostic accuracy.

Objectives:

A novel approach has been proposed toward the integration of different Nondestructive Evaluation data with EEG signals for improving the diagnosis of neurological disorders such as stroke, epilepsy, Parkinson's disease, and schizophrenia, by leveraging advanced neural network techniques in order to improve the identification and correlation of shared latent features across heterogeneous NDE datasets.

Methods:

We determined the 2D scalogram images using a specific encoder-decoder neural network after transforming the EEG signals using wavelet signal processing. Several NDE data types can be easily integrated for thorough analysis due to this network's ability to extract and correlate important aspects from each form of data. Aiming to uncover common patterns indicating of neurological disorders, the technique was evaluated on datasets containing EEG signals and corresponding NDE data.

Results:

Our method demonstrated a significant improvement in diagnostic accuracy and efficiency. The encoder-decoder network effectively identified shared latent features across the heterogeneous NDE datasets, leading to more precise and reliable diagnoses. The fusion of multi-modality NDE data with EEG signals provided a robust framework for the automatic identification of multiple neurological disorders.

Conclusions:

This innovative approach represents a substantial advancement in the field of neurological disorder diagnosis. By integrating diverse NDE data with EEG signals through advanced neural network techniques, we have developed a method that enhances the accuracy and efficiency of diagnosing multiple neurological conditions. This fusion of multi-modality data has the potential to revolutionize current diagnostic practices in neurology, paving the way for more precise and automated identification of neurological disorders.

Keywords

Neurological disorder identification deep learning in neuroimaging brain-Computer interface (BCI)encoder–Decoder networks multi-Modality fusion

1. Introduction

Neurological disorders comprise a range of diseases affecting the central and peripheral nervous systems, including neuropsychiatric, neurodevelopmental, and neurodegenerative diseases.¹ The neurological disorders category includes over 600 distinct diseases. These include well-known disorders such as epilepsy (EP), Parkinson's disease (PD), stroke (ST), Alzheimer's disease (AD) and other dementia, schizophrenia (SZ), cerebrovascular diseases, including stroke, brain tumors, migraine and headache-related diseases, and developmental disorders, including attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD).¹ In the world, neurological disorders rank second in terms of cause of death and are the primary cause of disability. In the last thirty years, there has been a significant rise in the mortality and disability rates linked to neurological diseases, particularly in low- and middle-income nations. Global population growth and aging are expected to cause further increases.² Anxiety and depression reduce productivity by the global economy by almost $1 trillion USD annually.³ Neurological disorders can affect patients’ and their families’ daily lives and even result in patient death. Early detection and medical intervention can improve clinical outcomes, even though it can be difficult to diagnose diseases in their early stages.Additionally, in low-income nations, there are less than two mental health professionals for every 100,000 people, whereas in high-income nations, there are more than 70.³

The serviceability and sustainability of neural structures, much like physical structures, are influenced by numerous factors necessitating a diverse array of diagnostic techniques for assessing neurological conditions. In the realm of neurology, clinicians employ various methods to detect and understand brain disorders, akin to how engineers use Nondestructive Evaluation (NDE) methods for structural inspections.⁴ As the application of neurological data grows, particularly EEG data, the need for advanced methods to discern meaningful relationships and reduce diagnostic uncertainty becomes increasingly crucial.

EEG data has been used in many studies to assess neurological disorders by extracting features through the use of mathematical transforms. Karhunen-Loeve transform,^5,6 Singular Value Decomposition (SVD),^7–9 Fourier transform,^10–12 Discrete cosine transform,^13,14 and wavelet transform^15–17 are some of the widely used techniques. Computational complexity and the effectiveness of information representation are important factors to take into account, as each approach has benefits and drawbacks. Wavelet transforms, in particular, have demonstrated a high degree of success in effectively extracting pertinent features from EEG data because of their capacity to capture time-frequency information.¹⁶ Examples of high computational complexity methods are singular value decomposition and the Karhunen-Loeve transform (O(n3)).¹⁸

Nondestructive Evaluation (NDE) systems can greatly improve the extraction of pertinent information in nonstationary environments by integrating EEG-based neural network techniques.¹⁹ Deep neural network models are increasingly being applied; they are especially well-suited for complex and dynamic data scenarios.²⁰ Convolutional Neural Networks (CNNs) are an essential deep learning algorithm for multi-dimensional data analysis.²¹ CNNs work particularly well with grid-like data formats, such as signal and image data, because of their adaptive filters, which can recognize intricate spatial relationships.²² In the context of NDE, this CNN functionality can be suitably modified for the analysis of EEG data. CNNs have demonstrated proficiency with NDE data formats,²³ and they can process and interpret EEG signals for neurological assessments with comparable efficiency.This means that Regarding the evaluation of EEG information for neurological disorders, it could be possible to find complex patterns linked to different brain conditions, which would improve the diagnostic procedure.

Nowadays, EEG data fusion method are frequently used in domains such as smart systems and sensor physical process, and their use in the context of EEG data analysis is growing in popularity. Increased signal-to-noise ratio, decreased false alarm rates, and improved disorder detection reliability can all result from integrating and correlating multiple EEG data sources.^24–27 Nonetheless, there is still a need to specifically identify correlations between various EEG data sources in order to diagnose neurological conditions.

Comprehensive studies that correlate data across different neurological evaluation techniques are conspicuously lacking, with each technique providing varying levels of diagnostic uncertainty and unique insights.²⁸ Due to variables like overlapping symptoms or variability in neurological conditions that affect diagnostic outcomes, this makes it difficult to evaluate such information over the course of a patient's medical journey. CNNs with EEG data can be used to classify cognitive states in an efficient manner, as shown in the research Decoding Cognitive Processes in Arithmetic Tasks. The model demonstrates strong performance metrics, such as 94% accuracy, and can be applied to neurofeedback and brain-computer interface (BCI) applications because to its real-time categorization capacity. The work emphasizes the need of scientific rigor by utilizing strategies to reduce overfitting, such as early stopping and learning rate decrease. It is advised that future studies investigate different neural network topologies and develop the model with larger datasets, thereby expanding the area of cognitive neuroscience.²⁹ A new technique that improves the mapping and assessment of cognitive phenomena using EEG data is introduced in Cognitive Phenomena assessment using Time Window-Based. Customizable analytic settings, including EEG channels and time frame lengths, provide versatility. It also performs exceptionally well in short-time brain mapping, capturing dynamic brain activity. The tool exhibits accuracy and potential for research and clinical applications after being validated using actual EEG data. Subsequent efforts will focus on enhancing its functionality and broadening its scope, emphasizing its noteworthy influence on cognitive neuroscience studies.³⁰ The presence of latent information offers a cohesive foundation for analyzing data, which could lead to enhanced precision and thoroughness in neurology diagnosis methods, as will be demonstrated in this study.

The following list includes this study's principal contributions:

1)
This paper presents a novel method for identifying multiple neurological disorders from EEG signals by combining multi-modality NDE fusion with encoder-decoder networks. Through the analysis of more complex EEG data, this novel approach improves the ability to identify and distinguish between distinct circumstances.
2)
This approach fills a major gap in the present research landscape by utilizing sophisticated neural network designs. Through the successful correlation of heterogeneous neurological data, it is possible to identify latent patterns shared by several EEG datasets, leading to a more profound comprehension of these disorders.
3)
The proposed methodology not only improves diagnostic precision but also offers a cohesive framework for comprehensive neurological evaluation. This approach has the potential to transform existing diagnostic practices, paving the way for more accurate and thorough identification of brain disorders.
The remainder of this paper is structured as follows: first, we describe our methodology, focusing on the integration of multi-modality NDE fusion and encoder-decoder networks in the analysis of EEG data for neurological disorders. Next, we present an experimental study that simulates the methodology of structural NDE assessments using EEG data. Finally, we discuss the implications of our findings and the potential for improving diagnostic practices in neurology.
2. Methodology

EEG data are preprocessed and converted using wavelet decomposition into 2D time-frequency scalograms before analysis. Then, using these EEG scalograms as training data, encoder-decoder networks are trained to map one set of EEG data to another, capturing shared features suggestive of neurological disorders. This method concentrates on the crucial details relevant to various neurological disorders by condensing the data into a latent representation. Figure 1 illustrates the overall procedure.

Figure 1.

EEG data analysis framework using encoder-decoder networks for neurological disorder identification. (a) Data preparation and pre-processing phase. (b) A schematic representation of our method.

Figure 1 shows a framework for processing EEG data in relation to diagnosing neurological conditions. The first step involves applying the Continuous Wavelet Transform (CWT) to convert raw EEG data into time-frequency representations, or scalograms. These scalograms are then fed into an autoencoder network, which is intended to extract the essential elements of the underlying EEG patterns from the complex data by condensing it into a smaller set of essential features. By discovering and using these latent features, the autoencoder seeks to reconstruct the input scalogram. This procedure could potentially speed up the identification and diagnosis of neurological disorders by distinguishing between EEG signals linked to different conditions.

These shared latent features are learned by deep neural networks. More specifically, the dimensionality of NDE data is reduced to a hidden layer cognitive content using a deep learning architecture called an autoencoder. The autoencoder reconstructs (decodes) the output cognitive content, which is defined as a complementary NDE measuring, from this reduced representation. Therefore, the two data types share basic NDE information in the form of the hidden layer weights.

2.1 EEG data preprocessing

2.1.1 Data alignment for correlation analysis of neurological disorders

The correlation analysis of EEG data in the identification of neurological disorders depends on accurate data alignment.³¹ As an instance, EEG data may be continuously recorded along a specific brain region or recorded at specific intervals during a study. This causes variations in the points at which data are collected between different EEG recording methods. Therefore, for consistent analysis, these datasets need to be spatially aligned. Assume that whereas one set of EEG data is recorded continuously at a given sampling rate, another set is obtained at predetermined intervals. Under such circumstances, the continuously recorded EEG data must be interpolated to match the grid points of the interval-based EEG data.

Let x₀,x₁,…,x_n be the collection points of the continuous EEG data, and y₀,y₁,…,y_n be the corresponding EEG measurements at these points. If we need to interpolate these data points onto a new set of grid points z₀,z₁,…,z_m, linear interpolation can be applied. The interpolated EEG value Y at a new grid point z can be calculated using the formula:

Y (z) = y_{i} + \frac{(y_{i + 1} - y_{i})}{(x_{i + 1} - x_{i})} \times (z - x_{i})

(1)

Where x_i and x_i₊₁are the original EEG data points nearest to z, and y_i and y_i₊₁ are the EEG measurements at these points.

Data from various EEG recording methods are synchronized and errors are reduced through this interpolation process.³² To align continuous EEG measurements with interval-based recordings, they would be linearly interpolated onto the predetermined grid points, as shown in the hypothetical figure. When trying to prepare the various sets of EEG data for later wavelet decomposition and analysis, After alignment, the EEG data are converted into a wavelet decomposition format that is appropriate for neural network analysis. Effective analysis and correlation of brain activity patterns associated with multiple neurological disorders depend on the described process of spatial alignment and interpolation of EEG data.

2.1.2 Wavelet transform for EEG signal analysis

The Wavelet Transform is particularly effective for EEG signal analysis due to its quality to decompose signals in both the temporal domain and spatial domain at the same time. This makes it ideal for representing the periodic, transient, and noisy characteristics often found in EEG data.³³ Wavelets can be thought of as basis functions that are obtained by manipulating a mother wavelet function, ψ_a(t). The following equation describes the wavelet transform function.³⁴:

W_{s} (b, a) = \int_{- \infty}^{\infty} \frac{s (t)}{\sqrt{| a |}} φ (\frac{t - b}{a}) d t

(2)

Where a,b∈R_i, a

\neq

0. Here,W_s is the mother function is scaley and translated, where the mother wavelet is represented by ψ(t), the translation parameter is b, and the dilation parameter is a.

Both continuous and discrete wavelet transforms are possible. Because it is effective at analyzing transient and noisy signals, which are typical in EEG data, the Continuous Wavelet Transform (CWT) is used in this work.³⁵ The advantage of the wavelet transform is its variable basis function length, which enables ideal adaptation to various frequency components. The flexibility of the wavelet transform to accommodate both disconnected changes and slow, persistent changes (using long basis functions) is essential for EEG analysis, which records both low-frequency and high-frequency brain waves.

The proposed Wavelet Transform-based EEG signal analysis framework is shown in Figure 2. It illustrates the various stages that the Continuous Wavelet Transform (CWT) goes through when processing EEG information. The system starts with the collection of EEG data, which is subsequently modified to capture aspects of both transient and persistent signals using the Complex Morlet wavelet. A scalogram, a time-frequency depiction of the EEG signal that emphasizes the energy distribution across various frequencies and times, is the end result. The encoder-decoder network, which is intended to evaluate and comprehend the wavelet-transformed data for additional processing and analysis, receives this scalogram after that.

Figure 2.

Overview of proposed framework.

This paper makes use of the Complex Morlet wavelet. The geometric properties of EEG waveform data are effectively represented by the Morlet function due to its impulse geometric shape.³⁴ A scalogram is a time-frequency internal representation of the energy distribution of a signal, denoted by the squared order of magnitude of the wavelet transform.³⁶ A scalogram in EEG analysis is represented as an image where the pixel luminance intensity at each time-frequency point is correlated with the energy density. The encoder-decoder network in our methodology then uses this type of wavelet-transformed EEG data as input.

2.2 Deep neural network for EEG analysis

2.2.1 EEG signal processing autoencoder networks

This part covers the encoder-decoder network that has been modified for processing EEG signals.For the purpose of processing both EEG data for neurological disorder identification and NDE data for structural health analysis, the autoencoder network—which consists of an encoder and a decoder—is employed in a dual capacity. The network is made up of two main parts: fully connected feedforward artificial neural networks (ANNs) called the encoder and the decoder. The encoder (a fully connected ANN) receives the input EEG scalogram first, and after that, it creates an encoded layer that is less dimensional than the original input. This encoded layer extracts the most essential characteristics from the EEG data.

Let the input data (EEG or NDE) be represented as a vector X∈Rn, where n is the no. of input neurons.

The encoded layer Z is obtained by applying a transformation to X. This can be represented mathematically as:

Z = f (W_{e} X + b_{e})

(3)

Where f is the activation function (such as sigmoid, ReLU), b_e is the encoder's bias vector, and W_e is the encoder's weight matrix.

The encoded layer Z represents a compressed version of the input signal in a lower-dimensional space (latent space). If the input dimension is n and the encoded layer dimension is m, then Z∈Rm with m < n.

The decoder reconstructs the end product exclusively using the encoded layers internal state, mirroring the framework of the encoder. Although symmetrical structures are common, they are not required. Reconstructing brain activity patterns for EEG and reconstructing images or signals that indicate structural integrity for NDE are two possible uses for this technique. The primary prerequisite is the same dimensionality for both the input and the output. We resize the input scalograms in our study to 64 × 64 × 1, and set the encoded layer to 8 × 8 × 1. This suggests that the shared data is compactly represented in an 8 × 8 2D array.

The output Y is reconstructed from the encoded layer. This is mathematically represented as:

Y = g (W_{d} Z + b_{d})

(4)

The weight matrix for the decoder is represented by W_d, the bias vector by b_d, and the activation function by g—which might or might not be the same as f—is represented by g.

To extract characteristics from the EEG data, such as edges, lines, and curves, autoencoder models learn a range of filters. Typically, these filters are 2D arrays of weights with values between 0 and 255. We utilize 3 × 3 size filters for our investigation. These filters could be seen in the context of EEG data in Figure 4(a). By using these filters on an EEG scalogram, spatial features can be converted into motifs referred to as activation maps or feature maps. Aspects of the EEG scalogram that are highlighted by these feature maps include regions of dominating frequency.

Figure 3.

EEG Data Linear Interpolation and Scalogram Transformation.

Figure 4.

Encoder–decoder network architecture.

As illustrated in Figure 4, the preliminary feature maps exhibit basic patterns and edges in the input image. Feature maps get increasingly abstract as we delve deeper into the network, encoding more intricate aspects of the EEG data. The most basic information retrieved from the input image is contained in the coded hidden layer (Figure 4), a crucial layer in the autoencoder that is reached by this process. This layer, which gradually combines high-level features to form higher-layer features, is essential for reconstructing the output image (Figure 3).

The autoencoder's goal is to reduce the variation between the input (X) and the output (Y). This is often done using a loss function, such as Mean Squared Error (MSE):

Loss = \frac{1}{n} \sum_{i = 1}^{n} (X_{i} - Y_{i})^{2}

(5)

Where X_i and Y_i are the h _i ^th elements of vectors X and Y, respectively.

The common patterns that are shared by the EEG images in the input and output are captured by autoencoders, which operate as nonlinear models. Reconstruction error is minimized and shared information is maximized in this architecture. The inverse process could be seen in Figure 4(b), which shows how the coded hidden layer in the EEG data domain efficiently represents a distinct latent mapping. We may observe that there is a strong latent mapping between the various EEG data representations when we compare the encoded layers in Figure 3.

2.2.2 Autoencoder architecture for EEG analysis

The two primary components of the autoencoder architecture are the left-hand and right-hand encoding and decoding paths (Figure 2). In order to gradually extract complex features from the input EEG scalogram, the encoding section adheres to the conventional architecture of Convolutional Neural Networks (CNNs). The process of applying convolutional blocks repeatedly results in this. For feature map extraction, every convolutional block has a 2D convolutional layer with 3 × 3 kernels. To reduce the spatial resolution (dimension), 2 × 2 max-pooling layers are employed. Along with a ReLU to add nonlinearity, there is a batch normalizing layers to enlarge the model and prevent overfitting.

To standardize the temporal resolution across various recordings, EEG data points are first subjected to linear interpolation, as shown in Figure 3. The time-frequency information of each interpolated EEG data point is then captured by applying a Continuous Wavelet Transform (CWT) to create a coefficient matrix. Following that, this matrix is converted into a 2D scalogram picture. The frequency and power of EEG signals over time are graphically represented by these scalograms, which are essential for recognizing the distinct neural patterns connected to different neurological disorders. The diagram would help to explain how unprocessed EEG data are transformed into a form that allows for additional analysis to check for the existence and traits of neurological disorders.

The different neural network architectures’ design parameters are compiled in Table 1. Network 1, Network 2, and Network 3 are the three standard networks with depths ranging from 6 to 10 layers. Network 1 Refined, Network 2 Refined, and Network 3 Refined are the networks with residual connections, which vary in depth and have three to five residual connections each. These networks may help with training deeper models. With eight layers and four residual connections, the U Network is a well-known tool for image segmentation tasks. Between the shallowest Network 1 and the deepest Network 3 Refined, the number of parameters increases from 2.7 thousand to over 55 thousand. The U Network has a significantly higher parameter count, almost 35 million, indicating its complexity and potential ability to learn complex patterns in data.

Table 1.
Architectures of the design parameters.

Network Name Network 1 Network 2 Network 3 Network 1 Refined Network 2 Refined Network 3 Refined U Network

Network Depth 6 8 10 6 8 10 8

Residual Connection Number - - - 3 4 5 4

Parametres (x 10³) 2.7 13.5 35.4 3.4 18.5 55.3 34,937

Network Name	Network 1	Network 2	Network 3	Network 1 Refined	Network 2 Refined	Network 3 Refined	U Network
Network Depth	6	8	10	6	8	10	8
Residual Connection Number	-	-	-	3	4	5	4
Parametres (x 10³)	2.7	13.5	35.4	3.4	18.5	55.3	34,937

The decoding path is crucial for generating the output EEG scalogram and progressively refining the feature resolution. By incorporating un-pooling layers that increase the spatial resolution of the feature maps, this decoding approach essentially mirrors the encoding path. To transform the final feature maps into the output scalogram, a concluding 1 × 1 convolutional layer is used.

The encoder serves to transform an input X∈R^iK into a hidden representation h∈R^iH, where:

h = f (W_{e} . X + b_{e})

(6)

Where, h represent the hidden representation. F is the activation function. We represent the weight associated with the encoding. Be respresent the bias terms.

The crucial elements of the EEG data are recorded during this process. The autoencoder's task is fulfilled by the decoding path, which uses the encoded representation to piece together the original EEG scalogram.

The X that has been rebuilt, or X′, is the outcome of the decoder function's mapping of the hidden representation h back to it:

\hat{X} = Φ_{2} (W_{2} h + b_{2})

(7)

Where a nonlinear activation function is denoted by Φ₂(.).The weight matrix W₂ is K × H.One bias vector is b₂.

An end-to-end mapping from one domain's input NDE data to another domain's equivalent NDE data can be applied to the network's entire structure. In this learning activity, the parameters (W₁, W₂, b₁, b₂) are estimated by minimizing the error may be occur VL between the refreshed data along with its Ground Truth:

L = \frac{1}{N} \sum_{i = 1}^{N} ‖ GT (x_{i}) - x_{i} ‖^{2}

(8)

Where, N is the number of data samples. Xi represents the input NDE data.GT(x_i) is the Ground Truth corresponding to ∥.∥² denotes the squared Euclidean distance.

2.2.3 Hyperparameter used for training

The model's performance and efficacy are greatly influenced by a number of critical hyperparameters that are included in the training process. The size of the steps taken during optimization is determined by the learning rate (0.001), which has an impact on how quickly or slowly the model converges to a minimum. More training epochs are needed to achieve the same level of precision in model weight adjustments as a smaller learning rate does. The number of samples processed prior to updating the model weights is specified by the minibatch size (32) which balances the stability of gradient estimates with computational efficiency. A weight decay value of 0.0001 is used as a regularization parameter to penalize big weights and encourage simpler models, preventing overfitting and enhancing generalization. To make sure the model gains sufficient knowledge from the data, it is trained for a total of 30 epochs, which is the number of full runs over the training dataset. Adam is the optimizer that's been used; it's well-known for its ability to handle sparse gradients efficiently and adaptably. Activation functions define the non-linearity introduced at each layer, influencing how effectively the model captures complicated patterns in the data, while dropout rate is employed to minimize overfitting by randomly removing neurons during training. Table 2 provides a summary of these hyperparameters and their values.

Table 2.
Hyperparameters used for training.

Hyperparameter Description Value

Learning Rate Size of the steps taken during optimization 0.001

Minibatch Size Number of samples processed before updating model weights 32

Weight Decay Factor Regularization parameter to prevent overfitting 0.0001

Number of Epochs Total passes through the entire training dataset 30

Optimizer Algorithm used for updating model weights Adam

Hyperparameter	Description	Value
Learning Rate	Size of the steps taken during optimization	0.001
Minibatch Size	Number of samples processed before updating model weights	32
Weight Decay Factor	Regularization parameter to prevent overfitting	0.0001
Number of Epochs	Total passes through the entire training dataset	30
Optimizer	Algorithm used for updating model weights	Adam

3. Results

3.1 Candidate learning architectures

In order to analyze EEG signals and identify multiple neurological disorders, we investigated different autoencoder architectures. Six distinct autoencoder networks, Network 1 through Network 3 Refined, which varied in depth and complexity, were used in our experiment. Basic autoencoders are represented by the Network 1, Network 2, and Network 3 architectures. They are matched in depth by their counterparts, Network 1 Refined, Network 2 Refined, and Network 3 Refined, but they are improved by residual connections that span three to five layers. These links are essential because they enable the network to integrate high-level abstraction with low-level feature recognition, which may enhance the detection of intricate patterns linked to neurological conditions. In addition, the U Network architecture, which is well-known for its competence in medical image segmentation, was modified and evaluated for performance in our field. A major component of our evaluation involved the network's performance, with an emphasis on its capacity to fuse multi-modal NDE data and recognize characteristic EEG patterns linked to neurological disorders.

Our research started out with more basic three-layer networks in order to get a general understanding of the issue. Remarkably, these less complicated models outperformed the more intricate, earlier U Network models in terms of performance, indicating the possibility of efficient model design without sacrificing diagnostic precision. Then, in order to assess the effects of each change, we methodically added layers to our models and integrated residual connections to make them more complex.

Figure 5.

The results of various evaluation metrics for the test dataset, including: a) PSNR, b) MSE, c) RMSE, d) PCC, and e) MAE.

The evaluation metrics for the several autoencoder architectures utilized in EEG signal analysis are shown in Figure 5. Mean Absolute Error (MAE), Pearson Correlation Coefficient (PCC), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR) are among the measurements. In terms of image quality, U-Net obtains the highest PSNR, while Net-2R does well with the lowest MSE, RMSE, and MAE, demonstrating its excellent accuracy and low prediction error. There is a good correlation between the PCC values of Network 2 Refined, Network 3, and U Network and the real data. These findings demonstrate a trade-off between performance and complexity, with Network 2 Refined producing accurate predictions with smaller mistakes and U-Net providing high-quality reconstructions.

The Table 3, which shows performance metrics for various neural networks evaluated on a dataset, demonstrates that U Network and Network 2 Refined perform better than the others in most cases. The best image quality is indicated by U Network's highest PSNR, while Network 2 Refined's lowest MSE, RMSE, and MAE suggest it makes the most accurate predictions with the least amount of error. There are significant correlations between the predictions made by U-Net and Net-2R and the real data, as evidenced by their high PCC values. Overall, U Network successfully strikes a balance between quality and accuracy, whereas Network 2 Refined excels in accuracy.

Table 3.

Values of evaluation metrics for the testing dataset.

Networks Name	Network 1	Network 2	Network 3	Network 1 Refined	Network 2 Refined	Network 3 Refined	U Network
PSNR	12.19	19.20	21.31	15.14	25.14	20.74	26.92
MSE	13,197.30	2778.40	1540.75	6550.75	827.15	1830.15	1050.80
RMSE	1111.90	50.42	38.65	79.18	26.60	41.70	32.40
PCC	0.53	0.88	0.91	0.81	0.92	0.91	0.86
MAE	176.04	75.90	56.00	116.14	33.40	64.05	42.57

Upon comparing the architectures, it was observed that U Network's convolutional blocks are significantly larger than those of Network 3 Refined, our most complex model, suggesting that U-Net has more complexity and hyperparameters. Understanding the trade-offs between model performance and complexity in the context of EEG signal analysis for neurological disorder identification was made possible thanks in large part to this comparison.

3.2 Experiment

3.2.1 Datasets

Five separate datasets from five different neurological disorders—Epilepsy (Ep), Parkinson's (Pd), Alzheimer's (Ad), Schizophrenia (Sz), and Stroke (St)—were employed in this study. Below, we go into further detail about these datasets.

1)
Epileptic Seizure Dataset:
The UCI Machine Learning Repository³⁷ is the source of the Epileptic Seizure dataset, which has been preprocessed and segmented into 11500 data entries, each consisting of 178 data points sampled during one second. The labels on the dataset relate to five categories: classes 2 through 5 capture different non-seizure states, such as eyes open or closed, EEG from healthy brain areas, and EEG from tumor-ridden sites. Type 1 reflects recordings of epileptic seizure activity. Researchers often use binary categorization to separate class 1 (seizures) from the other categories. This dataset is an essential source for research on epilepsy and EEG signal analysis. 2)
Parkinson's Disease Dataset:
The University of Iowa in Iowa City, Iowa, provides a publicly available dataset that is gathered for this study.³⁸ Along with 14 age and gender matched control participants, the sample consists of 14 PD patients (8 females and 6 males, mean ages 72.33 and 69.13 years, respectively). The Brain Vision system recorded EEG data at 500 Hz sample rate during resting-state, using 64 channels and Pz as the baseline mention channel. 3)
Alzheimer's Disease Dataset:
EEG recordings of 88 subjects resting with closed eyes are available to the public via openneuro.org.³⁹ The participants are divided into three groups: individuals with frontotemporal dementia (FTD group: 23 subjects), individuals with Alzheimer's disease (AD group: 36 subjects), and individuals in the healthy group (CN group: 29 subjects). Concerning cognitive state, the AD group scored an average of 17.75 on the Mini-Mental State Examination (MMSE), the FTD group scored 22.17, and the CN group scored a perfect 30. Utilizing 19 scalp electrodes and a 500 Hz sampling rate, the EEG recordings were obtained exploitation a Nihon Kohden EEG 21 hundred clinical device. 4)
Schizophrenia Disease Dataset:
This openly accessible dataset comprises 28 patients admitted to the Institute of Psychiatry and Neurology located in Warsaw, Poland.⁴⁰ It includes 14 healthy activity of a similar age group and 14 patients with a diagnosis of paranoid schizophrenia who are evenly distributed between genders and have an average of 28.3 years of age. EEG data was obtained using the international standard 10–20 EEG montage during a resting-state condition. Fifteen minutes of data were collected from nineteen channels. This dataset is useful for examining EEG patterns connected with schizophrenia and comparing them to healthy activity because the sampling frequency for EEG data acquisition was set at 250 Hz. 5)
Stroke Dataset:
This publicly accessible dataset (figshere.com)⁴¹ includes EEG recordings from fifty individuals who suffered from acute ischemic stroke. The participants’ ages range from thirty to seventy-seven years, and their gender distribution is eleven females and 39 males. The dataset includes EEG data from 22 individuals (all originally right-handed) with right hemiplegia and 28 individuals (all left hemiplegia) with EEG data obtained during the first 30 days following a stroke. Participants performed motor imagery exercises while watching corresponding videos, concentrating on left- or right-hand movements. This dataset is an essential resource for neurological research because it provides insightful information about motor imagery processes and post-stroke neural activity.

Table 4 summarizes the descriptive statistics for the datasets used in the study. The Epileptic Seizure Dataset features 11,500 samples with 178 features, showing 60% events and 40% censored data. The Parkinson's Disease Dataset includes 28 patients and 64 features, with 64.3% events. The Alzheimer's Disease Dataset has 88 subjects, 19 features, and 67.0% events. The Schizophrenia Dataset comprises 28 patients, 20 features, and 64.3% events. The Stroke Dataset consists of 50 patients, 64 features, with 60.0% events.

Table 4.
Descriptive statistics of neurological disorder datasets.

Dataset Name Dataset Type No. Patients No. Features No. Events No. Censoring

Epileptic Seizure Dataset³⁷ Single Risk 11,500 178 6900 (60%) 4600 (40%)

Parkinson's Disease Dataset³⁸ Single Risk 28 (14 PD patients, 14 controls) 64 18 (64.3%) 10 (35.7%)

Alzheimer's Disease Dataset³⁹ Single Risk 88 (23 FTD, 36 AD, 29 CN) 19 59 (67.0%) 29 (33.0%)

Schizophrenia Disease Dataset⁴⁰ Single Risk 28 (14 patients, 14 controls) 20 18 (64.3%) 10 (35.7%)

Stroke Dataset⁴¹ Single Risk 50 (22 right hemiplegia, 28 left hemiplegia) 64 30 (60.0%) 20 (40.0%)

All of these datasets are publicly available on the internet, and when the data was being collected, each subjects provided their informed permission for it to be published.
3.2.2 Multiple EEG data collection for neurological disorders

Dataset Name	Dataset Type	No. Patients	No. Features	No. Events	No. Censoring
Epileptic Seizure Dataset³⁷	Single Risk	11,500	178	6900 (60%)	4600 (40%)
Parkinson's Disease Dataset³⁸	Single Risk	28 (14 PD patients, 14 controls)	64	18 (64.3%)	10 (35.7%)
Alzheimer's Disease Dataset³⁹	Single Risk	88 (23 FTD, 36 AD, 29 CN)	19	59 (67.0%)	29 (33.0%)
Schizophrenia Disease Dataset⁴⁰	Single Risk	28 (14 patients, 14 controls)	20	18 (64.3%)	10 (35.7%)
Stroke Dataset⁴¹	Single Risk	50 (22 right hemiplegia, 28 left hemiplegia)	64	30 (60.0%)	20 (40.0%)

In this study, the effectiveness of multi-modality NDE fusion with encoder-decoder networks was examined through the collection of EEG data from patients suffering from a range of neurological conditions. The electrical activity of the brain can be taped non-invasively using EEG data, which can be used to gain important insights into diseases like epilepsy, Parkinson's, Alzheimer's, schizophrenia, and stroke. The modified procedure for gathering and analyzing data is as follows:

A collection of 6000 EEG samples was gathered to represent the different neurological disorders. To ensure representation from each condition across all sets, EEG data for each disorder was divided into training (70%), validation (10%), and testing (20%) datasets. The EEG devices used sensors to identify electrical impulses in the brain, just like IE equipment does. In order to capture the subtle activity linked to various brain states, the sampling frequency for EEG data was usually set at a high resolution. It was preferred to use time-frequency analysis, such as wavelet decomposition, to interpret the intricate EEG data. For example, compared to the aberrant signals from diseased states, normal brain activity would exhibit different dominant frequencies. To visualize these frequencies and differentiate between healthy and pathological brain regions, scalograms were created.

In addition to the EEG, other modalities that record the structure and function of the brain, such as functional MRI or CT scans, add layers of information to the picture. When combined, these modalities provide a complete picture of the state of the brain, with differences in the patterns seen representing defects or anomalies. To efficiently correlate the data, all modalities’ data were spatially aligned and interpolated. The correlation procedure ensured flexibility in managing heterogeneous EEG datasets by supporting data in scalar, waveform, or image formats.

This revised explanation puts the procedure for gathering and analyzing data in line with the particulars of identifying and researching a variety of neurological disorders using EEG and other neuroimaging methods. The methodology aims to improve diagnostic accuracy by synthesizing and interpreting complex datasets by utilizing the strengths of encoder-decoder networks.

3.2.3 EEG data augmentation for neurological disorder analysis

In the field of EEG analysis for neurological disorders, data augmentation refers to the process of intentionally growing the dataset by transforming the original EEG recordings in different ways. By preventing overfitting in neural networks, this method ensures that the models function properly when used with new, untested data. Overfitting occurs when a model learns the training set too thoroughly, incorporating noise and outliers. This can negatively impact the model's performance when applied to new data.

Data augmentation is particularly important in the field of neurology, where EEG signals are susceptible to a variety of interferences, such as electrical noise from medical equipment or ambient electromagnetic sources. By taking into account these possible disturbances that might not have been present in the original dataset, it helps the encoder-decoder network learn a more robust representation of the EEG signals.

Typical methods for augmenting EEG data consist of:

Noise Addition: To mimic possible measurement noise, zero-mean Gaussian white noise is added to the EEG signals.

Scaling: Modifying the EEG signals’ amplitude to reflect changes in signal strength.

Rotation and Translation: Changing the EEG signal sequence to mimic potential drifts and shifts that could happen

while the signal is being recorded.

Flipping is the process of representing possible polarity reversals by inverting the signal.

To enhance the EEG datasets in this study, we specifically used x-y translations, and noise insertion. Prior to augmentation, great care was taken to keep the training, validation, and testing datasets apart. By taking this precaution, you can be sure that information will not leak between the training and testing stages due to augmented data. The enhanced dataset, which was created by storing the augmented data alongside the original EEG recordings, was used to train more robust neural network models that could recognize minute patterns linked to a range of neurological conditions.

3.3 Evaluation metrics for EEG analysis

The performance evaluation of encoder-decoder networks is an important step in the neurological disorders from EEG signals. The networks are responsible for reconstructing representations of EEG signals, which are frequently obtained from a Continuous Wavelet Transform (CWT) as 2D scalograms. We use a number of image similarity metrics to compare the decoded outputs (predicted EEG scalograms) with the original (ground truth) EEG scalograms in order to evaluate the accuracy of these reconstructions. It is significant to remember that there are no set absolute performance thresholds; rather, these metrics are used for comparison.

Mean Absolute Error (MAE): To assess the accuracy of the EEG scalogram images, this metric measures the

pixel-by-pixel absolute difference between the ground truth image x and the predicted image px.

MAE = \frac{1}{n} \sum_{i = 1}^{n} | x (i) - px (i) |

(9)

Where n is the total number of pixels in the images.

Mean Squared Error (MSE): By squaring the squares of the pixel-wise errors, the Mean Squared Error (MSE) are

computed as the average squared difference between the actual and predicted values:

MSE = \frac{1}{n} \sum_{i = 1}^{n} (x (i) - px (i))^{2}

(10)

Root Mean Squared Error (RMSE): This statistic indicates the size of the error and is calculated by taking the

square root of the MSE.

RMSE = \sqrt{\frac{1}{n} \sum_{i = 1}^{n} {(x (i) - px (i))}^{2}}

(11)

PCC or Pearson Correlation Coefficient: The ground truth and predicted image linear correlation is measured by

the PCC, R, which is defined as follows:

R = \frac{Σ (x (i) - \bar{x}) (px (i) - p \bar{x})}{\sqrt{Σ {(x (i) - \bar{x})}^{2} Σ {(px (i) - p \bar{x})}^{2}}}

(12)

Where the mean values of the pixel intensities in the predicted and ground truth images are denoted by

p \bar{x}

and

\bar{x}

, respectively.

Peak Signal-to-Noise Ratio (PSNR): This metric assesses how well the anticipated EEG scalograms were

reconstructed. With increasing PSNR, the reconstructed image quality improves. It is provided by:

PSNR = 10 lo g_{10} (\frac{R^{2}}{MSE})

(13)

Where R is the maximum possible pixel value of the images. For instance, if the images are 8-bit, R would be 255.

3.4 Implementation of encoder-decoder networks for EEG analysis

We trained and deployed seven distinct encoder-decoder networks in our study, which were created especially for the EEG signals linked to neurological conditions. The implementation was completed with the help of TensorFlow 2.0, a potent open-source machine learning software library, and Keras, a high-level neural networks API.In order to maximize performance, the following hyper-parameters were selected through empirical testing and used in the training of these networks:

Learning Rate: Set to 0.001. This parameter determines the size of the steps taken during optimization. A smaller

learning rate ensures more precise adjustments to the model weights.

Minibatch Size: Fixed at 32. This specifies the number of training samples to be used in each iteration of the model

training process. A minibatch size of 32 strikes a balance between computational efficiency and the ability to

generalize from the training data.

Weight Decay Factor: Set to 0.0001. Weight decay, also known as L2 regularization, helps prevent overfitting by

penalizing large weights in the model.

Number of Epochs: Limited to 30. An epoch represents one complete pass of the training dataset through the

algorithm. Thirty epochs ensure that the networks undergo sufficient training to learn from the data without

overfitting.

We selected the Adam optimizer as the optimization algorithm. Given the complex and variable nature of EEG data in the analysis of neurological disorders, Adam (Adaptive Moment Estimation) is well known for its effectiveness in handling sparse gradients and adaptability to various problem scales.

4. Discussion

The outcomes of the encoder-decoder network implementations, which attempted to reconstruct EEG signal data for the purpose of identifying neurological disorders, are discussed in this section. For clinical end-users who depend on precise EEG signal interpretations, the findings are important.

Important findings from the research consist of:

Network Performance: Network with residual connections (Network 1 Refined to Network 3 Refined) generally

outperformed those without (Network 1 to Network 3), particularly in shallower networks. This was evident

across various evaluation metrics (MSE, RMSE, MAE, PSNR, PCC).

Impact of Network Depth: In non-residual networks, increasing depth led to better reconstructions and data

correlation. The deeper networks showed significant improvements in MSE, RMSE, MAE, PSNR, and PCC.

Optimal Network Complexity: For networks with residual connections, a mid-level depth (Network 2 Refined)

showed the best performance. Overly complex models (like Network 3 Refined) did not necessarily enhance

accuracy, indicating a balance between model depth and performance efficiency.

Pre-trained U Net Performance: The U-net model, known for handling nonlinear data associations, showed the

highest mean PSNR among all networks, indicating its robustness in EEG data reconstruction.

Qualitative Analysis: Visual inspection of reconstructed EEG data suggested that Network 2 Refined provided

more realistic results with fewer artifacts and blurriness. However, some networks rendered artifacts that could

lead to misinterpretations in clinical settings.

Residual Connections: Networks with residual connections displayed better reconstruction performance compared

to their counterparts without residuals. This improvement was attributed to the facilitated gradient flow in these

networks.

The outcomes show that encoder-decoder networks are a useful tool for processing EEG data and diagnosing neurological disorders. To ensure accurate and dependable EEG signal reconstruction for clinical use, they also stress the significance of selecting the appropriate network architecture and depth in order to strike a balance between complexity and performance.

Figure 6.

Training and Validation Loss Curves for Each Network Architecture Across 60 Epochs.

Six neural network models—Network 1, Network 2, Network 3, Network 1 Refined, Network 2 Refined, and Network 3 Refined—trained with learning rates of 0.001 and 0.0005 are shown in Figure 6 as plots comparing the training and validation Mean Absolute Error (MAE) loss curves. It is evident from each figure that the models are learning as the training and validation losses decrease over epochs. It is possible that overfitting occurs in models with a learning rate of 0.001, as evidenced by the greater volatility and wider difference between training and validation losses.Models with a learning rate of 0.0005 less frequently, on the other hand, have smoother and more stable loss curves, which point to improved generalization and decreased overfitting. While all models learn well, the general tendency is that more steady and dependable performance is attained by models with a lower learning rate.

Figure 7.

Mean absolute error (MAE) loss curves for six neural network models over 60 epochs.

The Mean Absolute Error (MAE) loss curves for six neural network models (Network 1, Network 2, Network 3, Network 1 Refined, Network 2 Refined, and Network 3 Refined) are shown in Figure 7 over 60 epochs, with a different color and line style for each model. For every model, the graphs show an overall decrease in MAE loss, indicating successful learning. Large variations are seen early in training, especially in models with a higher learning rate (0.001), which exhibit notable variability all through the training epochs. On the other hand, models with a lower learning rate (0.0005) show learning curves that are smoother and more stable, indicating improved generalization potential and consistency. Although the more stable routes of the lower learning rate models demonstrate their advantage in maintaining constant performance, by the end of the training period, all models attain similar MAE loss levels.

To summaries, networks lacking residual connections performed better when complexity increased when it came to EEG signal analysis for neurological disorder identification. Three subsets of the dataset comprise the training, validation, and test sets. The model is trained using the training set, which typically consists of 70–80% of the data, to update the weights. About 10–15% of the data are in the validation set, which is used to assess the model's performance and adjust hyperparameters during training to assist avoid overfitting. The test set, which assesses the model's ultimate performance on unobserved data, is designated for the remaining 10–15% of the data. The model's MAE loss is computed for the training and validation sets at each stage of the training process. This makes it possible to monitor the model's capacity for learning and generalization. This wasn't always the case, though, for networks with leftover connections. The less complex Network 2 Refined model outperformed the more complex Network 3 Refined model in particular. Given its high complexity and large number of parameters, this suggests that Network 3 Refined may be overfitting or not generalizing well enough, despite its sophistication. This suggests the need for a larger and more varied training dataset in order to adequately capture the wide range of EEG signal patterns linked to different neurological disorders, especially for more complex models.

4.1 Performance Metrics for Neurological Disorder Classification

We present the classification accuracy of our encoder-decoder networks for detecting multiple neurological disorders from EEG data. The model's performance was examined under five different conditions: epileptic seizures, Parkinson's disease, Alzheimer's disease, schizophrenia, and stroke. Table 5 summarizes the results, whereas Figure 8 depicts the model's accuracy for each disease.

Table 5.
Performance Metrics for Neurological Disorder Detection.

Condition Accuracy (%) Sensitivity (%) Specificity (%) Precision (%) F1 Score (%)

Epileptic Seizure 96.13 % 96.77% 97.56% 96.77% 96.77%

Parkinson Disease 95.68 % 93.33% 95.00% 93.33% 93.33%

Alzheimer Disease 97.93% 97.22% 97.67% 97.22% 97.22%

Schizophrenia 91.22% 90.91% 92.11% 86.96% 88.92%

Stroke 93.54% 92.31% 94.44% 92.31% 92.31%

Condition	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1 Score (%)
Epileptic Seizure	96.13 %	96.77%	97.56%	96.77%	96.77%
Parkinson Disease	95.68 %	93.33%	95.00%	93.33%	93.33%
Alzheimer Disease	97.93%	97.22%	97.67%	97.22%	97.22%
Schizophrenia	91.22%	90.91%	92.11%	86.96%	88.92%
Stroke	93.54%	92.31%	94.44%	92.31%	92.31%

Table 5 and Figure 8 demonstrate that our model performed well when categorizing EEG signals for Alzheimer's disease, with an accuracy of 97.93%. It also performed well for epileptic seizures 96.13% and Parkinson's disease 95.68%, demonstrating its capacity to detect specific EEG patterns for these disorders. While the accuracy for Schizophrenia was 91.22% and Stroke was 93.54%, these results just show good classification ability. Schizophrenia's decreased accuracy may highlight the complexity and variety of EEG patterns among patients.These performance measures highlight the potential of our technique in improving clinical diagnoses of neurological illnesses by providing a dependable tool for EEG data interpretation.

4.2 Comparative Analysis

In our study, we expanded our evaluation to include a comparative analysis with state-of-the-art methods, building on the experimental results from the previous section, where our proposed model exhibited superior performance in classifying neurological disorders compared to existing benchmarks. By implementing a robust evaluation framework, we ensured that our model's performance was consistently measured across the entire dataset. Our findings highlight the efficacy of Multi-Modality NDE fusion using Encoder–Decoder Networks in accurately identifying multiple neurological disorders from EEG signals, thereby advancing diagnostic methodologies in neurology.

Figure 9 presents a comparative analysis of the performance metrics for various machine learning approaches used in the classification of individual neurological diseases, as summarized in Table 6. Each method's effectiveness is evaluated based on key performance indicators: accuracy, sensitivity, specificity, precision, and F1-score. Notably, the Spatial-Temporal Autoencoder (STAE) with CNN-LSTM generative model demonstrates the highest accuracy of 96.30% in Alzheimer's disease classification, coupled with impressive sensitivity and precision metrics, indicating its robust performance⁴⁷. In contrast, the Multi-Head Graph Structure Learning approach for Parkinson's Disease shows a significantly lower accuracy of 69.40%, highlighting the challenges in effectively diagnosing this condition⁴⁴. The analysis further illustrates that approaches leveraging advanced deep learning techniques, such as autoencoders and hybrid models, generally outperform traditional methods across most diseases. This comprehensive comparison underscores the potential of machine learning in enhancing diagnostic accuracy and efficiency in neurology, paving the way for improved patient outcomes.

Table 6.
Comparison table of individual neurological disease classification.

Diseases Approaches Accuracy Sensitivity Specificity Precision F1-Score

Epileptic Seizure Autoencoder-Based Deep Neural Network with Long Short-Term Memory (LSTM) and a one-dimensional (1D) Convolutional Neural Network (CNN)⁴² 92.47% 88.89% 95.45% 94.12% 91.48%

AE-CDNN (Autoencoder-Convolutional Deep Neural Network)⁴³ 92.00% 90.17% 91.32% 85.44% 87.51%

Parkinson’s Disease Multi-Head Graph Structure Learning with Gradient Weighted Graph Attention⁴⁴ 69.40% 68.2% 71.6% 68.8% 67.50%

Common Spatial Pattern-based Approaches⁴⁵ 92.85% 92.22% 94.33% 93.42% 92.87%

Alzheimer Disease Dual-Branch Feature Fusion Network (DBN)⁴⁶ 80.23% 73.20% 75.45% 74.8% 74%

Spatial-Temporal Autoencoder (STAE) with CNN-LSTM Generative Model⁴⁷ 96.30% 97.73% 94.69% 94.8% 96.3%

Schizophrenia Dynamic Functional Connectivity (DFC) Analysis using Multifractal Analysis and Entropy Estimation⁴⁸ 89.29% 78.57% 100% 100% 87.8%

Convolutional 3D Autoencoder⁴⁹ 73.44% 83.87% 63.64% 69.7% 76.1%

Stroke Bidirectional Encoder Representations from Transformers (BERT) base model⁵⁰ 68.9% 77.9% 77.6% 77.9% 77.9%

Stacked Autoencoders for decoding asynchronous reaching movements⁵¹ 88% 80.17% 96.42% 95.2% 86.13%

Diseases	Approaches	Accuracy	Sensitivity	Specificity	Precision	F1-Score
Epileptic Seizure	Autoencoder-Based Deep Neural Network with Long Short-Term Memory (LSTM) and a one-dimensional (1D) Convolutional Neural Network (CNN)⁴²	92.47%	88.89%	95.45%	94.12%	91.48%
AE-CDNN (Autoencoder-Convolutional Deep Neural Network)⁴³	92.00%	90.17%	91.32%	85.44%	87.51%
Parkinson’s Disease	Multi-Head Graph Structure Learning with Gradient Weighted Graph Attention⁴⁴	69.40%	68.2%	71.6%	68.8%	67.50%
Common Spatial Pattern-based Approaches⁴⁵	92.85%	92.22%	94.33%	93.42%	92.87%
Alzheimer Disease	Dual-Branch Feature Fusion Network (DBN)⁴⁶	80.23%	73.20%	75.45%	74.8%	74%
Spatial-Temporal Autoencoder (STAE) with CNN-LSTM Generative Model⁴⁷	96.30%	97.73%	94.69%	94.8%	96.3%
Schizophrenia	Dynamic Functional Connectivity (DFC) Analysis using Multifractal Analysis and Entropy Estimation⁴⁸	89.29%	78.57%	100%	100%	87.8%
Convolutional 3D Autoencoder⁴⁹	73.44%	83.87%	63.64%	69.7%	76.1%
Stroke	Bidirectional Encoder Representations from Transformers (BERT) base model⁵⁰	68.9%	77.9%	77.6%	77.9%	77.9%
Stacked Autoencoders for decoding asynchronous reaching movements⁵¹	88%	80.17%	96.42%	95.2%	86.13%

Figure 8.

Accuracy for Different Neurological Disorders.

Figure 9.

Comparison of proposed methods results with individual existing methods of the datasets.

5. Sources of error and limitations in EEG data analysis

This study's use of encoder-decoder networks for the interpretation of EEG data came with a number of unique restrictions. An important obstacle was aligning EEG data from various scanners and modalities, which adds variability that might impact model performance. Due to variations in signal collection and noise characteristics, models trained on data from one type of EEG scanner may not generalize well to data from another. Furthermore, the model's capacity to generalize across a variety of circumstances may be limited by the study's dataset's incomplete representation of the broad spectrum of neurological disorders or brain patterns. Reducing these problems requires standardizing preparation techniques and employing a more diverse dataset. Moreover, model performance is strongly influenced by the specificity of brain patterns and the conditions present in the training data, emphasizing the necessity for extensive datasets that precisely reflect the range of disorders being studied. To increase the models’ resilience and clinical usefulness, these drawbacks must be addressed.

6. Conclusion

In conclusion, this work presents a novel method for detecting neurological disorders from EEG data by analyzing EEG signals using the Continuous Wavelet Transform and deep autoencoder neural networks. Even though the preliminary findings are encouraging, this study is only exploratory in nature and lays the groundwork for more comprehensive research. In order to improve the diagnosis of neurological disorders, future work will concentrate on validating this method in real-world scenarios and investigating how latent feature extraction through autoencoders can enhance data fusion.

Footnotes

ORCID iD

Shraddha Jain

Author contributions

All the authors have contributed equally.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

Datasets used in this study are publicly available and accessible.

References

World Health Organization. Neurological disorders: public health challenges. Geneva, Switzerland: World Health Organization, 2006.

Feigin

Vos

Nichols

, et al. The global burden of neurological disorders: translating evidence into policy. Lancet Neurol. 2020; 19: 255–265.

Quarshie

Davies

Otoo

. Setting mental health priorities in Ghana: a 15-year contextual analysis of the presidential state of the nation address. Challenges 2021; 12: 22.

National Research Council, Division on Engineering, Physical Sciences, Committee on Applied, Theoretical Statistics, Policy, Global Affairs, Committee on Science, Law, Committee on Identifying the Needs of the Forensic Sciences Community. Strengthening forensic science in the United States: a path forward. National Academies Press, 2009 Aug 29.

Rangayyan

Krishnan

. Biomedical signal analysis. John Wiley & Sons, 2024.

González-Vidal

Martinez-Ibarra

Skarmeta

. MultiBEATS: blocks of eigenvalues algorithm for multivariate time series dimensionality reduction. Inf Fusion. 2024; 104: 102159.

Yao

Feng

, et al. SVD-AE: an asymmetric autoencoder with SVD regularization for multivariate time series anomaly detection. Neural Netw. 2024; 170: 535–547.

Haider

Hanif

Rashid

, et al. EEG-based schizophrenia classification using penalized sequential dictionary learning in the context of mobile healthcare. Biomed Signal Process Control. 2024; 90: 105856.

Aljalal

Molinas

Aldosari

, et al. Mild cognitive impairment detection with optimally selected EEG channels based on variational mode decomposition and supervised machine learning. Biomed Signal Process Control. 2024; 87: 105462.

10.

Govil

Yao

Borao

. Multi-Dimensional Framework for EEG Signal Processing and Denoising Through Tensor-based Architecture. arXiv preprint arXiv:2401.05589. 2024 Jan 10.

11.

Sun

Peng

, et al. From eeg data to brain networks: Graph learning based brain disease diagnosis. IEEE Intelligent Systems, 2024 Jan 11.

12.

Lin

Yang

. Building a virtual psychological counselor by integrating EEG emotion detection with large-scale NLP models. In: Third international conference on biological engineering and medical science (ICBioMed2023), Vol. 12924. SPIE, 2024, pp.460–466.

13.

Kantipudi

Kumar

Aluvalu

, et al. An improved GBSO-TAENN-based EEG signal classification model for epileptic seizure detection. Sci Rep. 2024; 14: 843.

14.

Liu

, et al. Interpersonal Relationship Analysis with Dyadic EEG Signals via Learning Spatial-Temporal Patterns. arXiv preprint arXiv:2401.03250. 2024 Jan 6.

15.

Shah

Larijani

Gibson

, et al. Epileptic seizure classification based on random neural networks using discrete wavelet transform for electroencephalogram signal decomposition. Appl Sci. 2024; 14: 599.

16.

Dadhich

Patel

Tiwari

, et al. A flexible analytic wavelet transform and ensemble bagged tree model for electroencephalogram-based meditative mind-wandering detection. Healthcare Analytics 2024; 5: 100286.

17.

Devi

Sophia

. GA-CNN: analyzing student’s cognitive skills with EEG data using a hybrid deep learning approach. Biomed Signal Process Control. 2024; 90: 105888.

18.

Marco

Şentürk

Jeste

, et al. Functional mixed membership models. J Comput Graph Stat. 2024: 1–1.

19.

Sikora

Baniukiewicz

Chady

, et al. Artificial neural networks and fuzzy logic in nondestructive evaluation. In: Electromagnetic nondestructive evaluation (XVI). IOS Press, 2014, pp.137–151.

20.

Aldweesh

Derhab

Emam

. Deep learning approaches for anomaly-based intrusion detection systems: a survey, taxonomy, and open issues. Knowl Based Syst. 2020; 189: 105124.

21.

Croce

Zappasodi

Marzetti

, et al. Deep convolutional neural networks for feature-less automatic classification of independent components in multi-channel electrophysiological brain recordings. IEEE Trans Biomed Eng. 2018; 66: 2372–2380.

22.

Ditthapron

Banluesombatkul

Ketrat

, et al. Universal joint feature extraction for P300 EEG classification using multi-task autoencoder. IEEE access 2019; 7: 68415–68428.

23.

. CNN-Based Image analysis for EEG signal characterization. In: Advances in biomedical and bioinformatics engineering. IOS Press, 2023, pp.20–30.

24.

Khare

March

Barua

, et al. Application of data fusion for automated detection of children with developmental and mental disorders: a systematic review of the last decade. Inf Fusion. 2023: 101898.

25.

Guerrero

Parada

Espitia

. EEG signal analysis using classification techniques: Logistic regression, artificial neural networks, support vector machines, and convolutional neural networks. Heliyon 2021; 7(6).

26.

Yadav

Maini

. Electroencephalogram based brain-computer interface: applications, challenges, and opportunities. Multimed Tools Appl. 2023; 82: 47003–47047.

27.

Emara

El-Shafai

Algarni

, et al. A Hybrid Compressive Sensing and Classification Approach for Dynamic Storage Management of Vital Biomedical Signals. IEEE Access 2023.

28.

Munoz-Gama

Martin

Fernandez-Llatas

, et al. Process mining for healthcare: characteristics and challenges. J Biomed Inform. 2022; 127: 103994.

29.

Petrović

Spahić

Mandić

, et al. Decoding Cognitive Processes in Arithmetic Tasks: An EEG-Based Convolutional Neural Network Model.

30.

Petrović

Mandić

Borojević

, et al. Cognitive phenomena measurement with time window-based multispectral brain mapping. Technol Health Care. 2024: 1–0.

31.

Gil Ávila

Bott

Tiemann

, et al. DISCOVER-EEG: an open, fully automated EEG pipeline for biomarker discovery in clinical neuroscience. Sci Data. 2023; 10: 613.

32.

Lee

Tramontano

. Integrating Various neural features based on mechanism of intricate balance and ongoing activity: unified neural account underlying and correspondent to mental phenomena. World J Neurosci. 2021; 11: 161–210.

33.

Zhang

Zhou

Chen

, et al. The applied principles of EEG analysis methods in neuroscience and clinical neurology. Mil Med Res. 2023; 10: 67.

34.

Bajaj

. Wavelets for EEG analysis. Wavelet Theory 2020: 1–6.

35.

Wachowiak

Wachowiak-Smolíková

Johnson

, et al. Quantitative feature analysis of continuous analytic wavelet transforms of electrocardiography and electromyography. Philosophical transactions of the royal society A: mathematical. Phys Eng Sci. 2018; 376: 20170250.

36.

Top

. Classification of eeg signals using transfer learning on convolutional neural networks via spectrogram (Doctoral dissertation, Ankara Yıldırım Beyazıt Üniversitesi Fen Bilimleri Enstitüsü).

37.

Fokoue

. Epileptic seizure recognition. UCI Machine Learning Repository. (Accessed Nov 28, 2023).

38.

Anjum

Dasgupta

Mudumbai

, et al. Linear predictive coding distinguishes spectral EEG features of Parkinson's disease. Parkinsonism Relat Disord. 2020; 79: 79–85. (Accessed Nov 28, 2023).

39.

Miltiadous

Tzimourta

Afrantou

, et al. A dataset of EEG recordings from: Alzheimer's disease. Frontotemporal Dementia and Healthy Subjects. OpenNeuro 2023; 1: 88. (Accessed Nov 28, 2023).

40.

Olejarczyk

Jernajczyk

. Graph-based analysis of brain connectivity in schizophrenia. PloS one 2017; 12: e0188629. (Accessed Nov 28, 2023).

41.

https://figshare.com/articles/dataset/EEG_datasets_of_stroke_patients/21679035 (Accessed Nov 28, 2023).

42.

Mahapatra

Tariq

Anup

. Autoencoder-based deep neural architecture for epileptic seizures classification. In: Soft computing for problem solving: proceedings of the socpros 2022. Singapore: Springer Nature Singapore, 2023, pp. 109–119.

43.

Wen

Zhongnan

. Deep convolution neural network and autoencoders-based unsupervised feature learning of EEG signals. IEEE Access 2018; 6: 25399–25410.

44.

Neves

Yong

Yiming

. Parkinson's disease detection from resting state EEG using multi-head graph structure learning with gradient weighted graph attention explanations. arXiv preprint arXiv:2408.00906 (2024).

45.

Aljalal

Saeed

Khalil

Akram

Fahd

. Parkinson’s disease detection from resting-state EEG signals using common spatial pattern, entropy, and machine learning techniques. Diagnostics 2022; 12(5): 1033.

46.

Chen

Huabin

Dailei

Liping

Liang

. Multi-feature fusion learning for Alzheimer's disease prediction using EEG signals in resting state. Front Neurosci. 2023; 17: 1272834.

47.

Quanfa

Jing

Haitao

. Efficient identification of Alzheimer’s brain dynamics with Spatial-Temporal Autoencoder: A deep learning approach for diagnosing brain disorders. Biomed Signal Process Control. 2023; 86: 104917.

48.

Racz

Orestis

Peter

Andras

. Multifractal and entropy-based analysis of delta band neural activity reveals altered functional connectivity dynamics in schizophrenia. Front Syst Neurosci. 2020; 14: 49.

49.

Sun

Guoting

Xiaoqi

John

Qiyong

. Morphological fingerprinting: Identifying patients with first-episode schizophrenia using auto-encoded morphological patterns. Hum. Brain Mapp. 2023; 44(2): 779–789.

50.

Caiola

Avaneesh

Meijun

. EEG classification of traumatic brain injury and stroke from a nonspecific population using neural networks. PLOS Digital Health. 2023; 2(7): e0000282.

51.

Pei