Motor imagery EEG signal classification using novel deep learning algorithm

Abstract

Electroencephalography (EEG) signal classification plays a critical role in various biomedical and cognitive research applications, including neurological disorder detection and cognitive state monitoring. However, these technologies face challenges and exhibit reduced performances due to signal noise, inter-subject variability, and real-time processing demands. Thus, to overcome these limitations a novel model is presented in this research work for motor imagery (MI) EEG signal classification. To begin, the preprocessing stage of the proposed approach includes an innovative hybrid approach that combines empirical mode decomposition (EMD) for extracting intrinsic signal modes. In addition to that, continuous wavelet transform (CWT) is used for multi-resolution analysis. For spatial feature enhancement the proposed approach utilizes source power coherence (SPoC) integrated with common spatial patterns (CSP) for robust feature extraction. For final feature classification, an adaptive deep belief network (ADBN) is proposed. To attain enhanced performance the parameters of the classifier network are optimized using the Far and near optimization (FNO) algorithm. This combined approach provides superior classification accuracy and adaptability to diverse conditions in EEG signal analysis. The evaluations of the proposed approach were conducted using benchmark BCI competition IV Dataset 2a and Physionet dataset. On the BCI dataset, the proposed approach achieves 95.7% accuracy, 96.2% recall, 95.9% precision, and 97.5% specificity. In addition, it delivers 94.1% accuracy, 94.0% recall, 93.6% precision, and 95.0% specificity on the PhysioNet dataset. With better results, the proposed model attained superior performance compared to existing methods such as CNN, LSTM, and BiLSTM algorithms.

Introduction

Electroencephalography (EEG) is widely used for neurological analysis, cognitive state monitoring, and disease diagnosis. Efficient classification of EEG signals is essential for detecting mental states, disorders such as epilepsy, and sleep stage classification. However, the complexity of EEG signals increases due to inherent noise and overlapping brain activities. This makes the classification into a challenging task¹. The foremost challenge is the EEG data low signal-to-noise ratio (SNR). Since EEG signals are weak and frequently affected by factors such as muscle movements, eye blinks, and other environmental noise. Also, the brain activity of inter-subject has variations which further complicates classification models across individuals. Another critical issue is its real-time processing which requires quick decoding of motor imagery (MI) tasks. For quick processing, the system requires more computational resources and advanced algorithms². Additionally, continuous usage of BCI systems leads to user fatigue and reduced signal quality which further affects the system reliability. Also, it requires a typical setup which is complex, high cost and requires specialized expertise. Due to this, the accessibility and usability of these systems in clinical and non-clinical environments remain limited.

Various techniques have been explored to address these challenges in MI-based EEG signal classification. Preprocessing methods such as filtering for noise removal and feature extraction are utilized to improve the signal quality^3,4. Techniques like spatial pattern and Independent Component Analysis (ICA) are widely employed for spatial filtering and dimensionality reduction. Also, methods like Short-Time Fourier Transform and Wavelet Transform are used for specific feature extraction⁵. In terms of classification, traditional learning algorithms such as k-nearest neighbors, support vector machines, and linear discriminant analysis exhibited promising results in decoding MI tasks^6,7. However, these methods often struggle with generalization across users and sessions which limit the applicability in real time analysis.

The adoption of deep learning approaches in recent years for MI-based BCI systems exhibits better performances⁸. Variants of Convolutional Neural Networks and Recurrent Neural Networks are used due to their ability in modelling complex, non-linear relationships in data⁹. Hybrid models that integrate these methods with advanced preprocessing techniques have demonstrated improved classification accuracy¹⁰. Though these techniques have numerous merits they have limitations such as sensitivity to noise and computational inefficiency. Thus, the primary motivation of this research work is to develop a novel model which should overcome these challenges in existing models. Specifically, the objective is to attain enhanced accuracy and robustness in MI EEG signal classification using an advanced preprocessing and classification technique.

An innovative approach for MI EEG signal classification is presented in this research work combining hybrid preprocessing and classification techniques. The preprocessing stage presents a hybrid approach that integrates empirical mode decomposition (END) and continuous wavelet transform (CWT) for effective noise isolation and enhanced feature processing. To refine spatial features source power coherence and common spatial patterns are incorporated. For final classification, an Adaptive Deep Belief Network (ADBN) is proposed which is further optimized using the Far and Near Optimization (FNO) algorithm. Due to this the proposed model attains high accuracy in MI based EEG signal classification. The novelty of this work is present in its comprehensive approach of managing the noise, variability, and real-time constraints in EEG signal analysis. The contributions of this research work are presented as follows.

• A hybrid preprocessing model is presented incorporating EMD and CWT for improved signal representation and feature enhancement. Also included a feature refinement procedure considering source power coherence (SPoC) integrated with common spatial patterns (CSP) for spatial feature enhancement and robust feature extraction.

• Proposed an optimized classification model using Adaptive Deep Belief Network (ADBN) which is optimized with the Far and Near Optimization (FNO) algorithm to attain superior classification accuracy.

• Presented the proposed model experimental analysis in detail using benchmark BCI Competition IV Dataset 2a and Physionet dataset. The performance evaluation included metrics like precision, recall, f1-score, specificity, and accuracy.

• Presented a detailed comparative analysis with existing learning algorithms like CNN, LSTM, BiLSTM, GRB and RF algorithms to demonstrate the superior performance of proposed model.

The remaining discussion in the article is presented as follows. “Related works” section provides a detailed discussion of existing research works; “Proposed work” section provides the proposed model and “Results and discussion” section provides the experimental results and discussion. The conclusion is presented in “Conclusion” section.

Related works

The literature review on MI EEG signal classification presents an overview of existing research works. A wearable BCI system using a wireless headband targeting MI detection for real-time neurorehabilitation is presented in¹¹. The presented model employs a closed-loop design with visual feedback and integrates neural network model EEGNet for decoding MI-EEG signals. The wearable system requires fewer EEG channels and maintains robust classification across a larger subject pool than similar existing systems. Experimental analysis demonstrates the model’s superior performance comparable to commercial systems. However, the obtained accuracy is in the range of 75–85% which limits the model performance for further processing. The MSCARNet model presented in¹² combines multi-scale convolutional attention and Riemannian geometry embedding to attain enhanced classification accuracy for motor imagery EEG classification. The presented model utilizes sliding windows to partition raw EEG signals which enables data expansion for processing through temporal convolutional layers with multiple kernel sizes. These features are then mapped into Riemannian space to utilize its geometry for dimensionality reduction and improved signal interpretation. Also, the model integrates a lightweight attention to highlight critical spatio-temporal features. Experimental evaluations using BCI dataset exhibit the better performance of the presented model in subject-dependent and subject-independent classifications over traditional models. However, the presented model faces challenges with subject-independent scenarios, attributed to individual variability and increased computational demands from the Riemannian mapping.

The EEG classification model presented in¹³ utilizes wavelet packet decomposition for feature extraction. Then feature computation is performed using approximate entropy feature selection is performed using a modified binary grey wolf optimization. The resulting high-dimensional feature vector retains the most informative features and it is classified using KNN. The experimental analysis demonstrates superior performance compared to existing methods like SVM and CSP-based approaches in terms of precision, sensitivity, and MCC. However, limitations include the potential redundancy in selected features and the computational burden of WPD at higher levels. EEG-based MI classification model presented in¹⁴ aimed to improve classification accuracy by utilizing multi-time–frequency features. The presented model integrates Riemannian geometry for feature selection. This process operates on spatial covariance matrices with sparse optimization to attain enhanced feature selection and reduced redundancies. Additionally, the Dempster-Shafer theory fuses probabilistic outputs from multiple time windows and avoids uncertainties in classification decisions. Experimental validation on the BCI Dataset demonstrates the significant improvements in classification accuracy over traditional methods. Though the model has advantages of using multiple overlapping time–frequency bands it is computationally intensive due to dimensionality of Riemannian features.

The MI classification model reported in¹⁵ utilizes a combination of wavelet scattering transforms with a fusion of complementary classifiers. The presented approach utilizes wavelet time scattering to extract temporal features. Further fuzzy recurrence plots (FRPs) are used to convert EEG signals into grayscale texture images which enables wavelet image scattering for spatial feature extraction. Two SVM classifiers are separately trained on the wavelet time and image scattering features. Finally, using a fuzzy rule-based system their outputs are fused to overcome the limitations of individual classifiers. Experimental analysis exhibits an average accuracy compared to other learning algorithms. However, the obtained low accuracy and the high computational complexity highlights the model’s poor performance over other approaches. The MI classification based on EEG signals presented in¹⁶ utilizes ensemble empirical mode decomposition to overcome the noise and non-stationary challenges in EEG data. The presented model decomposes EEG signals into intrinsic mode functions and extracts the time domain features. Further the most discriminative features are identified using Kruskal–Wallis and chi-square tests. These features are subsequently classified using a variety of ML algorithms, including NN, SVM, and KNN. Experimental evaluations exhibit the superior performance of neural network classifiers. However, the analysis did not focus on multi-class scenarios which are the minor limitation of the presented model.

The Feedforward Backpropagation Neural Network (FFBPNN) model presented in¹⁷ for MI classification in EEG-based BCIs. The presented model initially normalizes raw EEG signals using a min–max scaling technique. The proposed neural network architecture includes two hidden layers with a tan-sigmoid transfer function. Levenberg–Marquardt algorithm to optimize the network. Due to this, the presented model effectively captures nonlinear patterns in EEG data. Experimental analysis using BCI dataset highlights better accuracy over existing algorithms like Filter Bank CSP and Stacked Linear Discriminant Analysis. However, the presented model performs only binary classification tasks and dependency on static network architecture reduces its adaptability to real-time applications. The fused group LASSO model is presented in¹⁸ for EEG classification. The presented model incorporates group sparsity and spatial smoothness by combining the strengths of group LASSO with a total variation norm. This enables the presented model to enforce spatial smoothness constraints while selecting relevant channels and features. Additionally, the model is optimized using a primal–dual algorithm which further improves the classification performances. The results demonstrate superior performance compared to other sparse optimization and spatial filtering methods. The limitations of the presented model are dependency on predefined regularization parameters which can be challenging to generalize across diverse datasets.

The neural network model MSCTANN presented in¹⁹ for classifying MI EEG signals in multi-class scenarios addressing challenges like individual variability and feature redundancy. The presented model integrates a multi-scale module for extracting diverse EEG features at varying granularities. The residual module further fuses feature and avoids network degradation. Additionally, a Channel-Temporal Attention Module is incorporated for highlighting important features across temporal and spatial dimensions. Experimentations using BCI dataset validates the model superiority with better results over state-of-the-art models like FBCSP and EEG-TCNet. However, the limitations of the presented model include the computational overhead associated with the complex architecture and the need for further exploration of optimal multi-scale kernel combinations. An innovative zero-shot learning approach is presented in²⁰ for classifying known and unseen MI tasks in a BCI system using EEG signals. The presented model combines the common spatial patterns and a neural network for feature extraction and classification of EEG features. The outlier detection procedure differentiates seen and unseen classes and the classification performed using a SoftMax classifies the known tasks and distance-based measures for novel tasks. Experimental analysis demonstrates the system’s ability in classifying unseen MI tasks with an accuracy of 91.81% compared to the upper bound method, emphasizing the potential to expand the MI task set without additional training data. However, the presented model has limitations in terms of scalability due to the detection of only one unknown category.

A modified deep neural network-based system presented in²¹ classifies the discriminative features extracted for four-class MI tasks using Four Class Filter Bank Common Spatial Pattern algorithm. Experimental evaluation on the BCI dataset highlights the improvements of the presented model in terms of metrics like precision, recall, specificity, and F1-score. The results highlight the system’s robustness in handling inter-subject variability compared to traditional ML algorithms. The EEG-based MI classification framework presented in²² utilizes graph representations of brain connectivity. The presented model involves transforming EEG signals into graph structures through functional connectivity analysis and extracts important structures using a specialized matrix. The extracted information is then fed into the CNN model which includes graph embeddings to detect critical structures for classification. The experimental analysis exhibits impressive results in terms of accuracy over traditional methods. However, the study identifies limitations, such as focusing solely on connectivity strengths without incorporating node-specific information, which could further enhance classification accuracy.

The multi scale CNN model presented in²³ for MI EEG classification includes spatial–temporal convolution (STC) blocks with varying temporal kernel sizes for multi-scale feature extraction. To enhance the discrimination of features supervised contrastive learning with cross-entropy loss are used in the presented model. The experimentation evaluations highlight the substantial improvements over existing learning models. However, the computational overhead, high convergence time and optimal parameter tuning for multi-scale kernels limits the model adaptability. Similar CNN based MI classification in EEG signals is presented in²⁴ for the BCI system. The presented CNN model integrates temporal and spatial convolutional layers with optimization techniques like batch normalization and dropout regularization to balance feature extraction and computational efficient classification. Experimental analysis on the BCI dataset validates the model’s robustness and achieves better classification accuracy over methods like AlexNet and ResNet50.

The deep learning methodology presented in²⁵ combines traditional filter bank spatial filtering with a CNN model. This enhances the discriminative power of EEG signals by retaining temporal information. A stage-wise training strategy optimizes the model by utilizing triplet loss for feature extraction and cross-entropy loss for classification. Experimental analysis demonstrates that presented FBSF-TSCNN outperforms existing methods in terms of Cohen’s kappa and accuracy. However, the proposed model’s complexity is higher than traditional methods. Also, it depends on separate training for feature extraction and classification layers limit its scalability. The hierarchical transformer-based classification algorithm presented in²⁶ for MI EEG signals designed for BCIs includes a low and a high-level transformer model for extracting specific features. The short-term intervals within a long MI trial are processed by a low-level transformer to extract interval-specific features. The high-level transformer employs a self-attention mechanism to highlight the features from relevant intervals. This enhances the model focus on motor imagery segments and the hierarchical structure allows to capture short and long-term temporal dependencies effectively. Experimental results of proposed architecture in subject-dependent and subject-independent scenarios outperforms conventional methods. However, the computational cost of the transformer architecture particularly for high-dimensional EEG data and challenges in optimizing hyperparameters for different datasets are the limitations of the presented model. Table 1 provides a summary of reviewed literature.

Table 1. Literature review summary.

Ref	Algorithm/Methodology	Limitations
12	MSCARNet with multi-scale convolution and Riemannian embeddings	Computational intensity in subject-independent tasks and reliance on parameter tuning
13	Ensemble transformers for multi-class EEG classification	High resource demands and limited exploration of lightweight adaptations for real-time usage
14	Riemannian Sparse Optimization with Dempster-Shafer Fusion	Intensive computational pipeline and limited focus on inter-session variability
15	Wavelet scattering and fuzzy fusion classifiers	Computational burden from fuzzy recurrence plot transformations and manually set parameters
16	MSCNet with supervised contrastive learning and multi-scale STC	Requires significant training time and struggles with optimal parameter tuning
17	Feedforward Backpropagation Neural Network (FFBPNN)	Restricted to binary classification and lacks adaptability for real-time applications
18	Fused group LASSO with total variation norm	Generalization challenges due to static regularization parameters and lack of frequency-specific optimizations
19	MSCTANN with multi-scale and attention-based learning	Increased computational overhead due to complex architecture and parameter tuning for kernels
20	Zero-shot learning with CSP and neural network projection	Limited scalability with a single unseen category and reliance on traditional CSP
21	FCIF with FC-FBCSP and modified DNN classifier	High computational complexity and dependency on hyperparameter tuning
22	Graph-based feature extraction with Ego-CNN	Ignores node-specific information and has limited scalability for small datasets
23	Wavelet packet decomposition with MBGWO and KNN	Computational complexity of WPD and redundancy in selected features
24	CNN with optimized lightweight architecture	Challenges in handling inter-subject variability and limited interpretability of deep learning models
25	FBSF-TSCNN combining filter bank spatial filtering with CNN	High computational complexity and separate training for feature extraction and classification
26	Hierarchical transformer architecture (HLT, LLT)	Computationally expensive for high-dimensional data and requires extensive hyperparameter tuning

Research gap

The detailed literature analysis reveals several critical research gaps in MI-based BCI systems. While various methods have explored feature extraction through approaches like CSP, wavelet transforms and graph-based representations it has limitations such as low robustness to inter-subject variability. Also, the inability to fully utilize the spatio-temporal dependencies, and computational inefficiency are the observed limitations of feature extraction models. Learning algorithms exhibits better performance but struggle with interpretability and scalability for real-time applications. Whereas transformer-based methods require extensive computational resources which limit their practical deployment. Artifact removal techniques utilized so far are effective but it introduces latency or relies on rigid preprocessing pipelines which limit its adaptability to dynamic scenarios. Further most of the studies highlight binary classification, neglecting the complexities of large-scale real-world BCI tasks. These gaps highlight the need for a novel approach that integrates efficient feature extraction, adaptive learning mechanisms, and lightweight architectures. Attaining enhanced robustness, scalability, and generalizability while maintaining low computational overhead should be the outcomes of the developed model.

Proposed work

The proposed work for MI-based EEG signal classification is designed to perform robust classification performances. The process begins with raw EEG signals captured as multi-channel time-series data. The input signals inherently possess challenges such as noise, variability, and non-stationarity. To overcome these challenges the proposed work incorporates a hybrid preprocessing stage, advanced feature extraction, and an optimized classification model enhanced by the Far and Near Optimization (FNO) algorithm. The preprocessing performed using Empirical Mode Decomposition (EMD) adaptively decomposes the signal into intrinsic mode functions (IMFs) and isolates oscillatory components by filtering out noise. This step is performed to manage non-linear and non-stationary signals. Each IMF is then transformed using Continuous Wavelet Transform (CWT) where it performs a multi-resolution analysis considering both time and frequency information. The reconstruction of the signal at this stage ensures the critical feature preservation while minimizing noise which is helpful to improve the classification accuracy. Further to enhance the spatial structure of the data, Source Power Coherence (SPoC) is utilized which refines features by aligning them with task-specific signal power. Additionally, Common Spatial Patterns (CSP) is applied to extract discriminative spatial features which maximize variance differences between classes. In the final classification, the extracted features are fed into an adaptive deep belief network (ADBN) optimized with the FNO algorithm. The FNO dual strategy of global exploration and local refinement ensures efficient convergence and avoids local minima outperforming conventional SGD and Adam optimizers. By combining these advanced techniques, the proposed model exhibits superior classification accuracy and demonstrates its robustness to variability and noise. Figure 1 presents the complete process flow of the proposed model.

[See PDF for image]

Fig. 1

Process flow of proposed model.

Preprocessing

The preprocessing in the proposed model focuses on decomposing, transforming, and refining the raw EEG signals to enhance the quality of input so that relevant features can be extracted for MI classification. For that first signal decomposition is performed using empirical mode decomposition. EMD adaptively decomposes the signal into a set of intrinsic mode functions (IMFs) and a residual component. This decomposition isolates the oscillatory patterns in the signal which reduces the complexity in subsequent feature extraction and noise reduction. The EEG signal is iteratively decomposed into a series of IMFs, denoted as , where represents the order of the IMF. The process begins with identifying the local maxima and minima of . A cubic spline interpolation is performed to connect these points yielding the upper envelope and the lower envelope . The mean envelope is then computed as

1

where indicates the upper envelope interpolated through local maxima, indicates the lower envelope interpolated through local minima, indicates the mean envelope which represents the slow-changing trends in the signal. A candidate IMF is obtained by subtracting from the signal

2

To ensure that meets the criteria of an IMF and iterative shift is performed. Specifically, the envelopes of are recalculated and a new mean envelope is obtained. This results in an updated candidate which is formulated as

3

where represents the iteration index. The sifting process terminates when satisfies the IMF conditions such as the number of extrema, zero crossings differ by one and the mean envelope is negligibly small, ideally approaching zero. Once an IMF is extracted, it is subtracted from the signal to obtain the residual. This process is formulated as

4

The residual becomes the new input signal for extracting the next IMF, and the process continues iteratively. The decomposition process results in a finite number of IMFs in which each represents distinct oscillatory modes. The final residual that encapsulates the non-oscillatory trend and the original signal including all is expressed as

5

where indicates the total number of IMFs extracted from the signal, indicates the intrinsic mode function, indicates the residual component containing noise.

The second step in the preprocessing stage is multi-resolution transformation using wavelet transformation. This step applies the wavelet transformation to analyze the signal at multiple resolutions which is crucial for extracting both time-domain and frequency-domain features. Given an input signal where is the index of the intrinsic mode function obtained from the previous step, the wavelet transforms computes the correlation between the signal and scaled versions of a wavelet function . This process is mathematically formulated as

6

where indicates the wavelet coefficient for the IMF at scale and translation , indicates the wavelet function which is used as a basis function for analysis, indicates a parameter which defines the wavelet time shift, indicates the scale parameter which controls the wavelet’s frequency resolution, indicates the complex conjugate of the wavelet function. The continuous wavelet transforms the map of the signal into a two-dimensional domain where represents time localization and provides frequency resolution. Further the wavelet function is scaled and translated to adapt its shape for capturing localized features of the signal. The scaled and translated wavelet function is mathematically formulated as

7

where indicates the normalization factor ensuring energy preservation across scales, indicates the scaled and shifted wavelet function. The scale determines whether the wavelet captures high-frequency or low-frequency components. Further the signal is reconstructed by inverting the wavelet transform which integrates all wavelet coefficients across scales and translations. Mathematically it is formulated as

8

where indicates the constant specific to the wavelet function which is defined as

9

where indicates the Fourier transform of . This reconstruction ensures that the original IMF is accurately represented in terms of its wavelet coefficients. The wavelet transform provides a flexible procedure for analyzing non-stationary signals like EEG. By breaking down the signal into time–frequency components this step captures localized features which are essential for differentiating the MI tasks.

Further in the preprocessing, signal reconstruction is performed to combine the relevant wavelet coefficients obtained in the previous step to reconstruct a clean and interpretable version of the intrinsic mode functions (IMFs). This step ensures that noise and irrelevant features are minimized while retaining the main signal information. The wavelet transform represents a signal as a combination of its wavelet coefficients across scales and translations. To reconstruct the IMF from its wavelet coefficients , the inverse wavelet transform is applied as per Eq. (8). However continuous integration over all scales and translations is computationally complex thus discrete approximation is employed where and take specific discrete values. Mathematically the approximation is formulated as

10

where indicates the summation over discrete scales, indicates the summation over discrete translations. This discrete approximation significantly reduces computational complexity and preserves most of the essential signal characteristics. Once each is reconstructed they are summed along with the residual to form the final reconstructed signal which is mathematically formulated as

11

where indicates the fully reconstructed version of the signal, indicates the total number of IMFs extracted during the EMD step, indicates the residual component representing the low-frequency or noise. This ensures that all relevant oscillatory modes are recombined to approximate the original signal and avoid unwanted artifacts. In order to enhance the reconstruction quality thresholding is applied to the wavelet coefficients before reconstruction. The threshold is applied to suppress noise which mathematically expressed as

12

where indicates the thresholded wavelet coefficients, indicates the threshold value. Using thresholded coefficients reconstruction further improves the signal-to-noise ratio. In order to refine spatial features Source Power Coherence (SPoC) is employed in the proposed. SPoC refines spatial features by identifying spatial filters that optimize the correlation between EEG signals and the power of a target variable. This step is essential for enhancing the relevance of spatial features in EEG data, making them more informative for classification tasks. The input to SPoC is the reconstructed EEG data matrix in which each row indicates a channel and each column represents the time sample. Let where indicates EEG channels, indicates time samples. The objective is to find spatial filters that maximize the correlation between the source power and a target variable which encodes the MI task. SPoC begins by calculating the covariance matrix of the EEG data matrix which is formulated as

13

where indicates the normalized covariance matrix of , the unnormalized covariance matrix is indicated as , and indicates the trace operator which is obtained by summing the diagonal elements of a matrix. The normalization in this step ensures that the covariance matrix is scale-invariant which making it robust to variations in signal amplitude. Further the target variable power over a given time window is computed as

14

where indicates the power of the target variable, indicates the value of at time . The power indicates the variance of which serves as the criterion for SPoC optimization. In order to refine the spatial features SPoC identifies spatial filters that maximize the correlation between the source power and the target power . This optimization problem can be expressed as

15

where indicates the spatial filter to be optimized, indicates the correlation between two variables. The optimization problem is solved through generalized eigenvalue decomposition which is formulated as

16

where indicates the eigenvalue which represents the contribution of the corresponding eigenvector to the total variance. indicates the diagonal matrix constructed from the covariance of the target variable . The eigenvector corresponding to the largest eigenvalue is selected as the spatial filter . Once the optimal spatial filter is identified, the spatially filtered signal is computed as follows

17

where indicates the spatially filtered signal for time , indicates the transposed spatial filter. This signal represents the refined features, highlighting the most relevant spatial components for MI classification.

Feature extraction through common spatial patterns (CSP)

In the proposed work feature extraction is performed using Common Spatial Patterns (CSP). CSP identifies spatial filters which maximize the EEG signal variance for one class while minimizing it for another. This process enhances class separability which making it an effective tool for Motor Imagery (MI) classification. Let and represent EEG data matrices for two classes, where indicates EEG channels, indicates the number of time samples for classes 1 and 2. The covariance matrices for the two classes are mathematically expressed as

18

19

where indicates the covariance matrices for classes 1 and 2 respectively, indicates the trace operator which is obtained by summing the diagonal elements of a matrix. The composite covariance matrix is obtained by the sum of the two normalized covariance matrices which are mathematically formulated as

20

This composite matrix captures the overall variance structure of the data. To decorrelate the data whitening transformation is applied. The whitening matrix is computed using the eigenvalue decomposition of which is formulated as

21

22

where indicates the eigenvector matrix of , indicates the diagonal matrix of eigenvalues, indicates the inverse square root of eigenvalues. The whitening transformation is then applied to the covariance matrices which are formulated as

23

24

This step ensures that the covariance matrices are uncorrelated. Further the CSP algorithm optimizes spatial filters by solving the eigenvalue problem which is formulated as follows

25

where indicates the eigenvector, indicates the eigenvalue. Further in the spatial filtering and feature extraction process, spatial filtered signals are computed and for each spatial filtered signal column of its variance is computed as follows

26

where is the number of time samples in . These variances form the feature vector used for classification.

Classification using optimized adaptive deep belief network (ADBN)

The optimized ADBN model used in the proposed work for classification processes the refined spatial features extracted from EEG signals to predict the associated class label for Motor Imagery (MI) task. The first step of the classification model involves organizing the refined spatial features extracted from EEG signals into a structured format suitable for processing by the ADBN. The spatially filtered EEG data is represented as a feature matrix capturing discriminative information relevant to the classification task. Each sample is characterized by a feature vector and the entire data is expressed as

27

where indicates the feature matrix containing samples with features each, indicates the feature vector corresponding to the sample. indicates the dimensionality of each feature vector representing the number of extracted features, indicates the total number of samples. To ensure numerical stability and consistency across samples the features are normalized. Let represent the normalized feature matrix and the normalization process is mathematically expressed as

28

where indicates the normalized value for the sample and feature, indicates the mean of the feature across all samples which is mathematically expressed as

29

indicates the standard deviation of the feature across all samples which is mathematically expressed as

30

The normalization ensures that all features have a zero mean and unit variance which further reduces the impact of varying magnitudes or scales among features. The input representation step organizes the extracted features into a structure compatible with neural network processing. After representing the features in a structured format, the next step involves feeding the feature matrix into a network for classification. This process involves initializing the input layer, propagating the features through successive layers, and applying transformations at each stage. The neural network input is the normalized feature matrix prepared in the previous step. Let where indicates the input layer representation of the feature matrix, indicates the number of features, indicates the number of samples. Each column of corresponds to a feature vector of a single sample. The first hidden layer applies a linear transformation to the input features using a weight matrix and a bias vector . This transformation is mathematically expressed as

31

where indicates the output of the first hidden layer, with hidden units, indicates the weight matrix mapping input features to hidden units, indicates the first hidden layer bias vector, indicates the activation function ReLU which is formulated as

32

The activation function allows the network to learn complex, non-linear relationships in the data. For each subsequent hidden layer the transformation is generalized which is mathematically expressed as

33

where indicates the output of the hidden layer, indicates the layer weight matrix. indicates the layer bias vector. indicates the number of hidden units in the layer. The activation function is applied element-wise to ensure non-linearity at each layer.

The output layer in the classification model is responsible for generating class probabilities based on the learned features from the previous layers. The output layer applies a linear transformation to the final hidden layer’s output where is the number of hidden layers in the network. This transformation is mathematically expressed as

34

where indicates the class score matrix for classes and samples, indicates the final hidden layer output in which is the number of units in the last hidden layer. indicates the weight matrix mapping the hidden layer output to the class scores, indicates the output layer bias vector. This linear transformation computes the raw scores which indicate the network’s confidence in each class before normalization. The final layer transforms the last hidden layer output into class probabilities using the softmax function. The transformation is mathematically expressed as

35

where indicates the probability matrix for classes and samples, indicates the output layer weight matrix, indicates the output layer bias vector. Each column of represents the predicted probabilities for all classes for a given sample. To convert the class scores into probabilities, the SoftMax function is applied to each column of . The SoftMax transformation is given by

36

where indicates the predicted probability of sample belonging to class . indicates the class score for sample and class . The SoftMax function ensures that each sample probability sum is equal to one which is mathematically formulated as

37

The normalization allows the model to provide the outputs as probabilities and makes the predictions more understandable. The predicted class for each sample is obtained by selecting the class with the highest probability which is mathematically formulated as follows.

38

where indicates the predicted class label for sample , indicates the probability of sample belonging to class . The loss function is an important factor in the classification model as it quantifies the difference between the predicted outputs and the true labels. Also, it provides feedback for optimizing the model’s parameters. Categorical cross-entropy loss is used in the proposed work for multi-class classification tasks. The categorical cross-entropy loss function measures the dissimilarity between the true class labels and predicted probability distribution . Mathematically the loss function is formulated as

39

where indicates the total loss value, indicates the total number of classes, indicates the total number of samples, indicates the binary indicator. The loss function aggregates the negative log probabilities of the predicted outputs for the true classes, encouraging the model to assign high probabilities to the correct classes. For each sample the log-likelihood for its true class is computed as

40

where indicates the contribution to the loss from sample , indicates the natural logarithm of the predicted probability for the true class . Since is a binary indicator, this effectively reduces to where is the true class for sample . This step penalizes the model more heavily for low predicted probabilities assigned to the correct class. The total loss is obtained by averaging the loss contributions over all samples which are formulated as

41

Substituting gives

42

This normalization ensures that the loss function scales proportionally to the data size, making it invariant to . The gradient of the loss function with respect to the predicted probabilities is computed. Mathematically it is formulated as

43

where indicates the gradient of the loss function with respect to , indicates the ratio of the true label to the predicted probability. These gradients are propagated backward through the network using the chain rule for updating the weights and biases. To fine-tune the parameters of the network, an optimization model is adopted in the proposed work.

Far and near optimization (FNO)

The Far and Near Optimization (FNO) algorithm incorporated in the proposed work fine-tunes the parameters of the network through its local and global search strategies. The objective of FNO is to minimize the model loss function by balancing global exploration and local exploitation. This hybrid strategy provides strong convergence to an optimal solution and overcomes the challenges such as local minima which exist in traditional optimization algorithms. To find the optimal set of parameters the objective function is formulated including weights and biases are mathematically expressed as

44

where which indicates the set of all trainable parameters across layers of the neural network. The optimization iteratively combines far and near search strategies to explore and refine the parameter space. The far search step provides a global exploration which effectively avoids local minima. Mathematically the far search candidate is formulated as

45

where indicates the candidate parameters after iteration, indicates the parameters at iteration , the exploration step size for far search is indicated as and indicates the random direction vector sampled from a uniform distribution. The far search improves the optimization and allows the model to explore diverse regions of the parameter space. The near search step focuses on precise parameter tuning using gradient-based updates. The candidate parameter computation for near search is mathematically formulated as

46

where indicates the candidate parameters after iteration, indicates the near search step size, indicates the gradient of the loss function. The near search ensures that parameter updates reduce the loss function locally by utilizing the gradient information for effective refinement. At each iteration, the algorithm evaluates the loss function for far and near search candidates which are formulated as

47

48

where indicates the far search loss and indicates the near search loss. The update rule for the parameters is determined adaptively which is formulated as

49

This adaptive procedure selects the parameters which provide the minimized loss function and ensures efficient progress toward the optimal solution. The optimization process continues until the termination condition such as maximum number of iterations or convergence condition which is formulated as follows

50

The far search prevents the model from stagnating in suboptimal solutions, while the near search fine-tunes the parameters for precise convergence. This dual strategy combined with adaptive evaluation aligns the optimization process with the model’s objective of minimizing classification errors and leads to improved accuracy in Motor Imagery classification tasks. The summarized pseudocode for the proposed model is presented as follows.

Results and discussion

The proposed model performance is validated through experimentation by utilizing benchmark BCI Competition IV Dataset 2a. The dataset has multi-class MI tasks with pre-segmented EEG signals. The experimental implementation of the proposed model was conducted on a workstation equipped with an Intel Core i7 processor clocked at 3.60 GHz, 32 GB of RAM, and an NVIDIA GeForce RTX 3080 GPU with 10 GB dedicated memory to accelerate deep learning computations. The experiments were conducted on a Windows 10 64-bit operating system environment. For the development and execution of the model, Python 3.8 was used as the primary programming language. Key libraries and frameworks employed included TensorFlow 2.9 for building and training the deep learning models, Scikit-learn for classical machine learning utilities and performance evaluation, and NumPy and Pandas for numerical and data preprocessing operations. Visualization tasks such as plotting learning curves and comparative graphs were managed using Matplotlib and Seaborn libraries. This hardware and software configuration ensured efficient training of the proposed deep learning architecture, stable resource handling for large-scale EEG data, and reproducible results for validation and comparison. The proposed model developed and implemented using Python with TensorFlow and Scikit-learn libraries as a multi-layer architecture to integrate spatial, temporal, and frequency-domain features effectively. Performance metrics such as accuracy, precision, recall, F1-score, and kappa, are computed for a thorough comparative analysis and compared with conventional ML and DL models like CNN, LSTM, BiLSTM, and Random Forest (RF) algorithms. Hybrid models like CNN with SVM, CNN with RF and DNN with SVM are considered for comparative analysis. Table 2 presents the simulation hyperparameters used in the proposed model and conventional models experimentation.

Table 2. Simulation HYPERPARAMETER.

S. No	Model	Hyperparameter	Value
1	Proposed model	Learning Rate	0.001
2		Optimizer	Adam
3		Batch Size	64
4		Epochs	100
5		Dropout Rate	0.3
6		Activation Function	ReLU (hidden layers), Softmax (output layer)
7		Loss Function	Categorical Cross-Entropy
8	CNN	Learning Rate	0.001
9		Optimizer	Adam
10		Batch Size	64
11		Epochs	100
12		Dropout Rate	0.2
13		Convolutional Layers	3
14		Kernel Size	(3,3)
15		Pooling	MaxPooling (2 × 2)
16		Activation Function	ReLU (hidden layers), Softmax (output layer)
17	LSTM & BiLSTM	Learning Rate	0.001
18		Optimizer	RMSProp
19		Batch Size	64
20		Epochs	100
21		Dropout Rate	0.3
22		Number of LSTM Units	128
23		Activation Function	Tanh (hidden layers), Softmax (output layer)
24	Random Forest (RF)	Number of Trees	100
25		Criterion	Gini
26		Min Samples Split	2
27		Min Samples Leaf	1
28	CNN-SVM	Learning Rate	0.001
29		Optimizer	Adam
30		Batch Size	64
31		Epochs	100
32		Dropout Rate	0.2
33		Convolutional Layers	3
34		Kernel Size	(3,3)
35		Pooling	MaxPooling (2 × 2)
36		SVM Kernel	RBF
37		Regularization (C)	1.0
38	CNN-RF	Trees	100
39		Learning Rate	0.001
40		Optimizer	Adam
41		Batch Size	64
42		Epochs	100
43		Dropout Rate	0.2
44		Convolutional Layers	3
45		Kernel Size	(3,3)
46	DNN-SVM	DNN Layers	3 Fully Connected
47		Activation Function	ReLU (hidden), Softmax (intermediate)
48		Dropout Rate	0.25
49		Optimizer	Adam
50		Final Classifier	SVM (RBF Kernel)
51		Regularization (C)	1.0

Performance analysis of proposed model using BCI IV dataset

The dataset used for the proposed model experimentation is the BCI Competition IV Dataset 2a which is a standard benchmark for motor imagery (MI)-based EEG signal classification²⁷. This dataset includes EEG recordings from 9 subjects, each performing 4 motor imagery tasks such as left hand, right hand, both feet, and tongue movements. The data was recorded using 22 EEG channels placed according to the international 10–20 electrode placement system along with 3 electrooculography (EOG) channels to monitor and account for eye movement artifacts. Each subject took two sessions, one for training with labelled trials and another for testing with unlabelled trials. The EEG signals were sampled at a rate of 250 Hz and bandpass-filtered between 0.5 and 100 Hz. To remove powerline interference a notch filter was applied at 50 Hz. Each session consists of 288 trials with 72 trials per task which results in a total of 576 trials per subject across both sessions. This complete process provided 2592 trials for training and 2592 trials for testing across all subjects. The dataset supports both subject-dependent and subject-independent evaluation which allows to perform complete assessment of model generalizability. By providing clean, labelled data along with raw signals, the dataset provides flexibility for preprocessing and feature extraction techniques making it an ideal choice for validating advanced MI-BCI systems.

Figures 2 and 3 illustrate the training and validation trends for accuracy and loss. The training accuracy consistently improved and reached maximum at 96.2%, while the validation accuracy closely followed and reached 95.7% by the final epoch. The results indicate the model strong generalization and minimal overfitting features. Correspondingly, the training and validation loss steadily declined, converging to 0.03 and 0.04, respectively. The parallel behavior of both curves confirms stable learning, efficient optimization, and reliable performance of the model in motor imagery EEG signal classification tasks. These outcomes affirm the model’s robustness and applicability for real-time BCI environments.

[See PDF for image]

Fig. 2

Proposed model accuracy analysis for BCI IV dataset.

[See PDF for image]

Fig. 3

Proposed model loss analysis for BCI IV dataset.

Further the proposed model performance is evaluated through precision, recall, f1-score, accuracy, cohen kappa, AUC-ROC, and specificity metrics. Table 3 presents the training and testing results obtained by the proposed model for all the metrics.

Table 3. Performance analysis of proposed model.

S. No	Performance metrics	Training	Testing
1	Precision	0.964	0.959
2	Recall	0.962	0.957
3	F1-Score	0.963	0.958
4	Accuracy	0.962	0.957
5	Cohen Kappa	0.943	0.936
6	AUC-ROC	0.978	0.973
7	Specificity	0.981	0.976

As presented in Table 3, the proposed model consistently performs well on both training and testing datasets. It achieved 96.2% accuracy in training and 95.7% in testing, reflecting strong generalization with minimal overfitting. Precision and recall remained high across both phases, indicating the model’s reliability in detecting relevant motor imagery events with low false positive and false negative rates. The F1-scores of 96.3% (training) and 95.8% (testing) confirm balanced performance. Notably, the Cohen’s kappa values above 93% demonstrate strong alignment between predicted and actual labels. High AUC-ROC and specificity scores further validate the model’s discriminative power and ability to correctly identify non-target instances. Overall, these metrics underscore the effectiveness of the proposed system.

Figure 4 compares the precision of the proposed model with conventional ML and DL approaches. At the 100th epoch, the proposed method achieves the highest precision of 95.9%, outperforming BiLSTM (94.0%), LSTM (92.5%), RF (91.5%), and CNN (91.0%). This improvement is attributed to the model’s advanced feature extraction and optimization mechanisms. While LSTM and BiLSTM capture temporal patterns, they lack robust spatial analysis. RF relies on manual features and CNN struggles with temporal dynamics, leading to reduced performance. Hybrid models such as CNN-SVM (91.6%), CNN-RF (91.8%), and DNN-SVM (92.0%) show modest improvements over their respective base architectures, confirming the benefit of combining feature extraction with traditional classifiers. The consistent upward trend across epochs highlights the reliability and effectiveness of the proposed model in minimizing false positives.

[See PDF for image]

Fig. 4

Precision comparative analysis for BCI IV dataset.

As shown in Fig. 5, the proposed model achieves the highest recall of 96.2% by the 100th epoch, outperforming BiLSTM (94.5%), LSTM (93.0%), RF (92.0%), and CNN (91.5%). This demonstrates the model’s strong capability to detect motor imagery events with fewer missed classifications. While BiLSTM offers competitive results due to its bidirectional temporal learning, it lacks the spatial refinement incorporated in the proposed method. Models like CNN and RF show relatively lower recall, reflecting their limitations in processing complex spatial–temporal patterns in EEG data. Hybrid models such as CNN-SVM (92.0%), CNN-RF (92.2%), and DNN-SVM (92.5%) demonstrate improved performance compared to their base counterparts but still fall behind the proposed approach. Overall, the superior recall confirms the model’s effectiveness in identifying true motor imagery classes with high sensitivity.

[See PDF for image]

Fig. 5

Recall comparative analysis for BCI IV dataset.

Figure 6 shows the specificity comparison across models in distinguishing non-target motor imagery classes. The proposed model records the highest specificity of 97.5% at the 100th epoch, surpassing BiLSTM (96.0%), LSTM (95.0%), RF (94.5%), and CNN (94.0%). This high specificity reflects the model’s strong ability to filter out irrelevant signals and reduce false positives. Hybrid architectures like DNN-SVM (94.7%), CNN-RF (94.6%), and CNN-SVM (94.4%) show improvement over their individual counterparts due to enhanced feature representation. Nevertheless, they remain slightly behind BiLSTM (96.0%) and the proposed method, which leverages advanced preprocessing and spatial filtering. CNN and RF models, at 94.0% and 94.5%, respectively, exhibit limitations in discriminating non-relevant classes The consistent improvement over epochs further supports the proposed model’s reliability for real-time EEG classification tasks.

[See PDF for image]

Fig. 6

Specificity comparative analysis for BCI IV dataset.

As illustrated in Fig. 7, the proposed model reaches an F1-score of 96.0%, outperforming BiLSTM (94.5%), LSTM (93.0%), RF (92.0%), and CNN (91.0%). This indicates a strong balance between precision and recall, reflecting the model’s effectiveness in handling both false positives and false negatives. Hybrid methods such as DNN-SVM (92.2%), CNN-RF (92.0%), and CNN-SVM (91.8%) outperform their traditional methods, but still lags behind BiLSTM (94.5%) and the proposed model. This margin underscores the proposed method’s enhanced ability to capture spatiotemporal features and reduce both false positives and false negatives. The steady rise in F1-score throughout training further demonstrates the model’s stable learning behavior and efficient optimization. These results affirm the adaptability and reliability of the proposed approach in motor imagery EEG classification.

[See PDF for image]

Fig. 7

F1-score comparative analysis for BCI IV dataset.

The accuracy comparative analysis given in Fig. 8 for the proposed model and existing methods highlights the better performance of the proposed model. The proposed model achieves the highest accuracy of 95.7% at the 100th epoch which is significantly better than the existing methods. The accuracy attained by the existing BiLSTM is 94.0% which is 1.7% less than the proposed mode. LSTM attained an accuracy of 92.5% which is 3.2% below the proposed model. The accuracy attained by the RF is 91.5%, and CNN is 91.0% which is approximately 4% less than the proposed model. Among hybrid methods, DNN-SVM (92.4%), CNN-RF (92.3%), and CNN-SVM (92.1%) show modest gains over their base versions, validating the benefit of combining deep and traditional learning. However, none of them match the consistency and final accuracy of the proposed approach. The superior performance of the proposed model is due to its hybrid feature processing and the optimized classifier network.

[See PDF for image]

Fig. 8

Accuracy comparative analysis for BCI IV dataset.

Table 4 summarizes the comparative performance of all models, clearly demonstrating the superiority of the proposed method. It records the highest precision (95.9%) and recall (96.2%), reflecting its ability to accurately identify true positives while minimizing missed detections. The F1-score of 96.0% confirms its well-balanced classification capability. With a specificity of 97.5%, the model effectively distinguishes non-target instances, outperforming all other methods. The overall accuracy of 95.7% further highlights its robustness and reliability. Compared to other deep learning and hybrid models, the consistent lead across all metrics underscores the effectiveness of the proposed approach. These improvements are attributed to the model’s hybrid preprocessing, effective spatial–temporal feature integration, and optimization using the Far and Near Optimization algorithm, making it well-suited for real-world applications.

Table 4. Overall performance analysis.

Model	Precision	Recall	F1-Score	Specificity	Accuracy
CNN	0.910	0.915	0.912	0.940	0.910
LSTM	0.925	0.930	0.928	0.948	0.925
BiLSTM	0.940	0.945	0.943	0.960	0.940
RF	0.915	0.920	0.918	0.945	0.920
CNN-SVM	0.916	0.920	0.918	0.944	0.921
CNN-RF	0.918	0.922	0.920	0.946	0.923
DNN-SVM	0.920	0.925	0.922	0.947	0.924
Proposed	0.959	0.962	0.960	0.975	0.957

Performance analysis of proposed model using Physionet dataset

The second dataset used for the proposed model experimentation is the Physionet Dataset which is also a standard benchmark for motor imagery (MI)-based EEG signal classification²⁸. This publicly available dataset comprises EEG recordings collected from 109 subjects performing various motor tasks, including both actual and imagined movements such as opening and closing of fists or feet. For MI-specific evaluation, trials involving motor imagination of limb movements were selectively used to maintain consistency with typical MI-BCI applications. Recordings were captured using a 64-channel EEG system configured according to the international 10–10 placement standard. The signals were digitized at a sampling rate of 160 Hz, with pre-applied bandpass filtering between 0.5 and 40 Hz. To mitigate contamination from ocular and muscle artifacts, additional preprocessing was performed, including Independent Component Analysis (ICA) for artifact rejection. The dataset’s structured format, which includes event markers and synchronized labels, facilitates the implementation of flexible preprocessing pipelines and feature extraction methods. These characteristics make the PhysioNet dataset an excellent alternative for verifying the scalability and adaptability of advanced EEG classification models in MI-based BCI systems (Figs. 9, 10).

[See PDF for image]

Fig. 9

Proposed model accuracy analysis for Physionet dataset.

[See PDF for image]

Fig. 10

Proposed model loss analysis for Physionet dataset.

The training and testing accuracy and loss plots for the PhysioNet dataset clearly demonstrate the learning stability and generalization ability of the proposed model. From the accuracy curve, it is evident that the model progressively improves across epochs, reaching a training accuracy of approximately 94.3% and a testing accuracy close to 94.1% by the 100th epoch. The narrow margin between training and testing accuracy reflects minimal overfitting, indicating that the model effectively captures underlying patterns without becoming biased to the training data. Additionally, the loss curves exhibit a consistent downward trend, with both training and testing losses approaching near-zero values towards the final epochs. The gradual and stable decrease in loss across both curves suggests efficient convergence and optimization. Notably, the testing loss remains closely aligned with the training loss, confirming strong generalization and resilience against overfitting.

The quantitative evaluation of the proposed model on the PhysioNet dataset, as summarized in Table 5, affirms its strong and consistent performance across both training and testing phases. The model achieves a training precision of 93.8% and a closely aligned testing precision of 93.6%, indicating reliable classification with minimal false positives. Similarly, the recall scores of 94.3% during training and 94.0% during testing demonstrate the model’s effectiveness in identifying true positives. The F1-scores, which represent the harmonic mean of precision and recall, stand at 94.0% for training and 93.8% for testing, highlighting a balanced trade-off. An accuracy of 94.2% in training and 94.1% in testing further reinforces the model’s overall classification capability. Cohen’s Kappa values, 0.918 for training and 0.916 for testing, reflect a strong agreement between predicted and actual labels. Additionally, the model achieves high AUC-ROC values of 96.3% and 96.2%, demonstrating excellent discriminatory power. The specificity scores of 95.1% (training) and 95.0% (testing) confirm the model’s competence in correctly identifying non-target instances. These results collectively indicate the robustness, stability, and generalization ability of the proposed model for motor imagery classification tasks using EEG signals.

Table 5. Performance evaluation of the proposed model on PhysioNet dataset.

S. No	Performance Metric	Training	Testing
1	Precision	0.938	0.936
2	Recall	0.943	0.940
3	F1-Score	0.940	0.938
4	Accuracy	0.942	0.941
5	Cohen’s Kappa	0.918	0.916
6	AUC-ROC	0.963	0.962
7	Specificity	0.951	0.950

Figure 11 presents a comparative evaluation of precision values across 100 training epochs for the proposed method and various existing models on the PhysioNet dataset. By the final epoch, the proposed model achieves the highest precision of 93.6%, outperforming all other techniques. Specifically, BiLSTM attains a precision of 93.0%, LSTM records 91.8%, CNN-RF reaches 91.5%, CNN-SVM achieves 91.3%, DNN-SVM records 91.4%, RF reaches 91.2%, CNN stands at 90.0%, and CNN-LSTM yields 91.0%. These results emphasize the superior discriminative ability of the proposed model, particularly in reducing false positives during classification. Traditional models like CNN and RF show comparatively lower precision due to their inability to model complex spatiotemporal dependencies.

[See PDF for image]

Fig. 11

Comparative analysis of precision metric for Physionet dataset.

The recall comparative analysis of the proposed model and existing approaches presented in Fig. 12 using the PhysioNet dataset exhibits the proposed model highest recall of 94.0% which clearly demonstrating its superior ability to correctly identify true positive instances in motor imagery classification. In comparison, BiLSTM achieves a recall of 93.5%, LSTM reaches 92.0%, while CNN-RF, CNN-SVM, and DNN-SVM attain 91.6%, 91.5%, and 91.8%, respectively. Traditional classifiers such as RF and CNN yield 91.5% and 90.5%, indicating limited sensitivity in identifying relevant EEG patterns. Although BiLSTM and LSTM exhibit commendable performance due to their sequential learning structure, they are constrained by their limited spatial understanding. Hybrid models show modest improvements over their respective baselines by integrating learned features with advanced classifiers; however, they still fall short of the recall achieved by the proposed model.

[See PDF for image]

Fig. 12

Comparative analysis of recall metric for Physionet dataset.

The specificity analysis given in Fig. 13 for existing and proposed models highlights the proposed model highest specificity of 95.0% which clearly demonstrates its superior capability in correctly identifying negative instances and reduce false positives. In comparison, BiLSTM achieves 94.5%, LSTM records 94.0%, CNN-RF and CNN-SVM achieve 93.6% and 93.5%, while DNN-SVM reaches 93.7%. Traditional classifiers like RF and CNN show comparatively lower scores at 93.5% and 93.0%, respectively. Although BiLSTM maintains strong performance through bidirectional temporal learning, and LSTM captures sequential dependencies effectively, both lack comprehensive spatial feature handling. The hybrid models show moderate improvements over their base versions, reflecting the benefits of combining neural representations with classical classifiers. Nevertheless, they remain below the specificity attained by the proposed method.

[See PDF for image]

Fig. 13

Comparative analysis of specificity metric for Physionet dataset.

The comparative performance analysis given in Fig. 14 showcases the proposed model against several conventional and hybrid approaches for F1-score metric. At the 100th epoch, the proposed model achieves the highest F1-score of 93.8%, indicating a well-balanced trade-off between precision and recall. Among the compared models, BiLSTM records an F1-score of 93.3%, followed by LSTM at 91.8%. The hybrid models—CNN-RF, CNN-SVM, and DNN-SVM—obtain final F1-scores of 91.4%, 91.3%, and 91.6%, respectively. RF and CNN achieve slightly lower values of 91.2% and 90.2%, reflecting limitations in capturing the spatiotemporal characteristics of EEG signals. While BiLSTM benefits from bidirectional temporal learning and LSTM captures sequential dependencies effectively, both models show marginal gains compared to the proposed approach. Hybrid models show improvements over traditional models but still it is lesser than the proposed model’s performances.

[See PDF for image]

Fig. 14

Comparative analysis of F1-score metric for Physionet dataset.

The accuracy comparative analysis given in Fig. 15 depicts the proposed and existing models performances. The proposed model achieves the best outcome with an accuracy of 94.1%, showcasing its strong classification ability. BiLSTM follows with 93.0%, leveraging temporal bidirectionality. LSTM and CNN-RF reach 91.8% and 91.5%, respectively, while CNN-SVM and DNN-SVM yield 91.3% and 91.6%. CNN and RF lag slightly behind at 90.5% and 91.2%. Although hybrid models enhance the performance of their individual components, they remain below the proposed method. The consistently rising accuracy curve of the proposed model reflects its robust learning dynamics and adaptability to unseen EEG inputs. Meanwhile, traditional and sequence-only models show limited gains due to restricted spatial or contextual feature extraction.

[See PDF for image]

Fig. 15

Comparative analysis of accuracy metric for Physionet dataset.

Table 6 highlights the performance comparison among various models on the PhysioNet dataset. The proposed method outperforms all others, achieving top scores in precision (93.6%), recall (94.0%), F1-score (93.8%), specificity (95.0%), and accuracy (94.1%), confirming its strong generalization and classification capability. BiLSTM ranks second, with a notable accuracy of 93.0%, attributed to its strength in learning bidirectional temporal patterns, though it lacks in spatial representation. LSTM also delivers competitive results (91.8% accuracy), slightly behind due to less contextual awareness. Hybrid models like CNN-RF (91.5%) and DNN-SVM (91.6%) exhibit better precision and accuracy than their base versions, indicating that model fusion contributes to improved learning. CNN and RF, while foundational, perform comparatively lower, mainly due to limited ability to model both spatial and sequential EEG features.

Table 6. Overall performance analysis for PhysioNet Dataset.

Model	Precision	Recall	F1-Measure	Specificity	Accuracy
CNN	0.900	0.905	0.902	0.930	0.905
LSTM	0.915	0.920	0.918	0.940	0.918
BiLSTM	0.930	0.935	0.933	0.945	0.930
RF	0.910	0.915	0.912	0.935	0.912
CNN-SVM	0.910	0.915	0.913	0.935	0.913
CNN-RF	0.912	0.917	0.914	0.936	0.915
DNN-SVM	0.913	0.918	0.916	0.937	0.916
Proposed	0.936	0.940	0.938	0.950	0.941

The computational complexity of the proposed model was analyzed in terms of Floating-Point Operations Per Second (FLOPs) to provide a better understanding of its processing demands. The hybrid preprocessing stage, which involves Empirical Mode Decomposition (EMD) and Continuous Wavelet Transform (CWT), indeed introduces additional computational overhead compared to models that directly use raw EEG signals with deep learning architectures. The EMD step has a complexity of approximately O(N²) due to iterative sifting, while the CWT, depending on the scale and frequency resolution, contributes a complexity close to O(N log N) per signal segment. However, these steps are executed only once per signal, and their cost is offset by the significant improvement in feature quality and noise suppression. In contrast, end-to-end deep learning models trained on raw EEG data may seem faster in execution time but often suffer from reduced interpretability and struggle with noisy, non-stationary signals, leading to suboptimal generalization. The proposed model’s advantage lies in its balanced trade-off between accuracy and robustness. While existing models may be marginally faster, the proposed method significantly outperforms them in terms of precision, recall, and specificity, as shown in the experimental results. Moreover, the integration of Source Power Coherence (SPoC) and Common Spatial Patterns (CSP) enhances class separability, while the Adaptive Deep Belief Network (ADBN), optimized with the Far and Near Optimization (FNO) algorithm, ensures reliable convergence and avoids local minima. Thus, the proposed approach provides a compelling advantage in scenarios where classification accuracy, generalization across subjects, and robustness to noise are critical, such as in clinical and assistive BCI applications, where marginal improvements in accuracy can have significant real-world impact.

Conclusion

A novel optimized deep learning model is presented in this research work for MI EEG Signal Classification. By combining hybrid preprocessing techniques and optimized classification strategies the proposed model overcomes limitations in existing techniques. The hybrid preprocessing model utilizes Empirical Mode Decomposition (EMD) and Continuous Wavelet Transform (CWT) for isolating critical signal modes and enhancing time–frequency representation. Spatial feature refinement with Source Power Coherence (SPoC) and Common Spatial Patterns (CSP) in the proposed model provides better extraction of discriminative features. The Adaptive Deep Belief Network (ADBN) which is optimized using Far and Near Optimization (FNO) algorithm in the proposed model attained better classification accuracy in MI classification. Comprehensive evaluation was performed on two benchmark datasets such as BCI Competition IV 2a and the PhysioNet EEG Motor Imagery dataset. On the BCI-IV dataset, the proposed model achieved a classification accuracy of 95.7%, outperforming existing techniques such as CNN (91.0%) and BiLSTM (94.0%). It also recorded a precision of 95.9%, recall of 96.2%, and specificity of 97.5%, confirming its superior detection capability. On the PhysioNet dataset, the model maintained high performance, reaching an accuracy of 94.1%, precision of 93.6%, recall of 94.0%, and specificity of 95.0%, demonstrating its strong generalization across varied EEG sources. These results validate the proposed model efficiency in feature extraction and optimization strategies. Though the proposed model is better than existing models, it exhibits computational intensity due to its multi-step preprocessing and FNO-based optimization. Future research can address these limitations by integrating lightweight architectures which enhances its applicability to dynamic BCI environments.

Author contributions

Sathish Mathiyazhagan—Research proposal—construction of the work flow and model—Final Drafting—Survey of Existing works—Improvisation of the proposed model; Dr. Geetha Devasena M. S—Initial Drafting of the paper—Collection of datasets and choice of their suitability—Formulation of pseudocode.

Data availability

The datasets generated and/or analysed during the current study are available in the repository https://bbci.de/competition/iv/download/

Competing interests

The authors declare no competing interests.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Aggarwal, S; Chugh, N. Signal processing techniques for motor imagery brain computer interface: A review. Array; 2019; 1–2, pp. 1-12. [DOI: https://dx.doi.org/10.1016/j.array.2019.100003]

2. Awuah, WA; Ahluwalia, A; Darko, K; Sanker, V; Tan, JK; Pearl, TO; Atallah, O. Bridging minds and machines: The recent advances of brain-computer interfaces in neurological and neurosurgical applications. World Neurosurg.; 2024; 189, pp. 138-153. [DOI: https://dx.doi.org/10.1016/j.wneu.2024.05.104] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38789029]

3. Geng, X; Zhang, Xi; Yue, M; Weixin, Hu; Wang, L; Zhang, X; Ping, Yu; Long, D; Yan, H. A motor imagery EEG signal optimized processing algorithm. Alex. Eng. J.; 2024; 101, pp. 38-51. [DOI: https://dx.doi.org/10.1016/j.aej.2024.05.077]

4. George, O; Smith, R; Madiraju, P; Yahyasoltani, N; Ahamed, SI. Data augmentation strategies for EEG-based motor imagery decoding. Heliyon; 2022; 8, pp. 1-14. [DOI: https://dx.doi.org/10.1016/j.heliyon.2022.e10240]

5. Chen, D. Improved empirical mode decomposition bagging RCSP combined with Fisher discriminant method for EEG feature extraction and classification. Heliyon; 2024; 10, pp. 1-10.1:CAS:528:DC%2BB2cXpsVOhtg%3D%3D [DOI: https://dx.doi.org/10.1016/j.heliyon.2024.e28235]

6. Barnova, K; Mikolasova, M; Kahankova, RV; Jaros, R; Kawala-Sterniuk, A; Snasel, V; Martinek, R. Implementation of artificial intelligence and machine learning-based methods in brain–computer interaction. Comput. Biol. Med.; 2023; 163, pp. 1-24. [DOI: https://dx.doi.org/10.1016/j.compbiomed.2023.107135]

7. Kew, SYN; Mok, S-Y; Goh, C-H. Machine learning and brain-computer interface approaches in prognosis and individualized care strategies for individuals with amyotrophic lateral sclerosis: A systematic review. MethodsX; 2024; 13, pp. 1-11. [DOI: https://dx.doi.org/10.1016/j.mex.2024.102765]

8. Wang, X; Liesaputra, V; Liu, Z; Wang, Yi; Huang, Z. An in-depth survey on deep learning-based motor imagery electroencephalogram (EEG) classification. Artif. Intell. Med.; 2024; 147, pp. 1-20. [DOI: https://dx.doi.org/10.1016/j.artmed.2023.102738]

9. Rekha, V; Samuel Manoharan, J; Hemalatha, R; Saravanan, D. Deep learning models for multiple face mask detection under a complex big data environment. Proc. Comput. Sci.; 2022; 215, pp. 706-712.1:STN:280:DC%2BB28nmslCrtg%3D%3D [DOI: https://dx.doi.org/10.1016/j.procs.2022.12.072]

10. Samuel Manoharan, J; Braveen, M; Ganesan Subramanian, G. A hybrid approach to accelerate the classification accuracy of cervical cancer data with class imbalance problems. Int. J. Data Min.; 2021; 25, 3/4 pp. 234-259. [DOI: https://dx.doi.org/10.1504/IJDMB.2021.122865]

11. Rao, Z; Zhu, J; Lu, Z; Zhang, R; Li, K; Guan, Z. A wearable brain-computer interface with fewer EEG channels for online motor imagery detection. IEEE Trans. Neural Syst. Rehabil. Eng.; 2024; 32, pp. 4143-4154. [DOI: https://dx.doi.org/10.1109/TNSRE.2024.3502135]

12. Zhou, B; Wang, L; Xu, W; Jiang, C. Multi-scale convolutional attention and Riemannian geometry network for EEG-based motor imagery classification. IEEE Access; 2024; 12, pp. 79731-79740. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3410036]

13. Dhiman, R. Motor imagery signal classification using Wavelet packet decomposition and modified binary grey wolf optimization. Meas. Sens.; 2024; 24, pp. 1-9. [DOI: https://dx.doi.org/10.1016/j.measen.2022.100553]

14. Jin, J; Qu, T; Xu, R; Wang, X; Cichocki, A. Motor imagery EEG classification based on riemannian sparse optimization and dempster-shafer fusion of multi-time-frequency patterns. IEEE Trans. Neural Syst. Rehabil. Eng.; 2023; 31, pp. 58-67. [DOI: https://dx.doi.org/10.1109/TNSRE.2022.3217573] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36288214]

15. Tuan, D. Pham, “classification of motor-imagery tasks using a large EEG dataset by fusing classifiers learning on wavelet-scattering features”. IEEE Trans. Neural Syst. Rehabil. Eng.; 2023; 31, pp. 1097-1107. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3241241]

16. Kardam, VS; Taran, S; Pandey, A. Motor imagery tasks based electroencephalogram signals classification using data-driven features. Neurosci. Inf.; 2023; 3, pp. 1-13. [DOI: https://dx.doi.org/10.1016/j.neuri.2023.100128]

17. Jana, GC; Swetapadma, A; Pattnaik, PK. Enhancing the performance of motor imagery classification to design a robust brain computer interface using feed forward back-propagation neural network. Ain Shams Eng. J.; 2018; 9, pp. 2871-2878. [DOI: https://dx.doi.org/10.1016/j.asej.2017.12.003]

18. Zhang, S; Zhu, Z; Zhang, B; Feng, B; Yu, T; Li, Z. Fused group lasso: A new EEG classification model with spatial smooth constraint for motor imagery-based brain-computer interface. IEEE Sens. J.; 2021; 21, 2 pp. 1764-1778.2021ISenJ.21.1764Z [DOI: https://dx.doi.org/10.1109/JSEN.2020.3016402]

19. Wu, R; Jin, J; Daly, I; Wang, X; Cichocki, A. Classification of motor imagery based on multi-scale feature extraction and the channel-temporal attention module. IEEE Trans. Neural Syst. Rehabil. Eng.; 2023; 31, pp. 3075-3085. [DOI: https://dx.doi.org/10.1109/TNSRE.2023.3294815] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37436869]

20. Duan, L; Li, J; Ji, H; Pang, Z; Zheng, X; Lu, R. Zero-shot learning for EEG classification in motor imagery-based BCI system. IEEE Trans. Neural Syst. Rehabil. Eng.; 2020; 28, 11 pp. 2411-2419. [DOI: https://dx.doi.org/10.1109/TNSRE.2020.3027004] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32986556]

21. Akuthota, S; RajKumar, K; Ravichander, J. Artifact removal and motor imagery classification in EEG using advanced algorithms and modified DNN. Heliyon; 2024; 10, pp. 1-16. [DOI: https://dx.doi.org/10.1016/j.heliyon.2024.e27198]

22. Jin, J; Sun, H; Daly, I; Li, S; Liu, C; Wang, X. A novel classification framework using the graph representations of electroencephalogram for motor imagery based brain-computer interface. IEEE Trans. Neural Syst. Rehabil. Eng.; 2022; 30, pp. 20-29. [DOI: https://dx.doi.org/10.1109/TNSRE.2021.3139095] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34962871]

23. Zhao, R; Wang, Y; Cheng, X; Zhu, W; Meng, X; Niu, H; Cheng, J; Liu, T. A mutli-scale spatial-temporal convolutional neural network with contrastive learning for motor imagery EEG classification. Med. Novel Technol. Dev.; 2023; 17, pp. 1-9. [DOI: https://dx.doi.org/10.1016/j.medntd.2023.100215]

24. Tarahi, O; Hamou, S; Moufassih, M; Agounad, S; Azam, HI. Decoding brain signals: A convolutional neural network approach for motor imagery classification, e-prime-advances in electrical engineering. Electron. Energy; 2024; 7, pp. 1-9. [DOI: https://dx.doi.org/10.1016/j.prime.2024.100451]

25. Chen, J; Yu, Z; Gu, Z; Li, Y. Deep temporal-spatial feature learning for motor imagery-based brain–computer interfaces. IEEE Trans. Neural Syst. Rehabil. Eng.; 2020; 28, 11 pp. 2356-2366. [DOI: https://dx.doi.org/10.1109/TNSRE.2020.3023417] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32956061]

26. Deny, P; Cheon, S; Son, H; Choi, KW. Hierarchical transformer for motor imagery-based brain computer interface. IEEE J. Biomed. Health Inf.; 2023; 27, 11 pp. 5459-5470. [DOI: https://dx.doi.org/10.1109/JBHI.2023.3304646]

27. https://bbci.de/competition/iv/download/

28. https://physionet.org/content/eegmmidb/1.0.0/