A lightweight methodology for Motor Imagery EEG classification utilizing step scaled wavelet fractals and Bi-LSTM architecture

A lightweight methodology for Motor Imagery EEG classification utilizing step scaled wavelet fractals and Bi-LSTM architecture | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A lightweight methodology for Motor Imagery EEG classification utilizing step scaled wavelet fractals and Bi-LSTM architecture Balendra ., Neeraj Sharma, Shiru Sharma This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5750495/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Brain-Computer Interface (BCI) represents a cutting-edge area of research that integrates neuroscience, engineering, and computer science to establish direct communication links between the human brain and external systems. BCI-based technologies hold significant promise, particularly in the development of prosthetic devices. However, the practical application of BCI in real-time scenarios faces several obstacles, including bulky models, noise interference, artifacts, and the complexity of motor imagery (MI) electroencephalogram (EEG) data, which exhibits both inter-subject and intra-subject variability. To address these challenges, the proposed algorithm utilizes loading and pre-processing of MI EEG, extracts their common spatial patterns, calculates the continuous wavelet transform (CWT) coefficients, computes their proposed step scaled mean wavelet fractals which exhibits robustness towards inherent noises and artifacts, calculates the cross-correlation matrix at different scale for all channels and observes the evolution in cross-correlation matrices with the help of customized Bi-long-short term memory (Bi-LSTM) neural network to classify MI EEG. The customized Bi-LSTM architecture had the size < 10MB showing the effectiveness of methodology for MI EEG classification utilizing embedded based devices. The best classification accuracies achieved with proposed step scaled mean wavelet fractals were 87.75% and 85.45% for intra-subject as well as 76.16% and 71.76% for inter-subject on BCI Competition IV 2b and 2a respectively; the comparative analysis with earlier state of the art methods showed an average improvement of 2.28% and 2.33% for intra-subject as well as 0.22% and 3.21% for inter-subject in accuracy. Electroencephalogram (EEG) Motor Imagery Continuous Wavelet Transform Cross correlation matrix Bidirectional long short term memory Figures Figure 1 Figure 2 Figure 3 1. Introduction Brain Computer Interface (BCI) technology represents a cutting-edge area of research that integrates neuroscience, engineering, and computer science to establish direct communication links between the human brain and external systems [ 1 ]. In contrast to traditional neuromuscular pathways that depend on peripheral nerves and muscles, BCIs provide an innovative solution, allowing individuals to control devices solely through their cognitive processes [ 2 ]. This shift in approach has the potential to be transformative, especially for individuals with motor disabilities, granting them greater independence and control over their actions [ 3 ]. At the core of BCI systems is the process of decoding neural activity, where sophisticated algorithms interpret brain signals and convert them into actionable commands for external devices [ 4 ]. A key application in this field focuses on decoding motor imagery (MI) movements from electroencephalography (EEG) signals captured during motor imagery tasks [ 5 ]. Motor imagery involves mentally visualizing physical movements without actual muscle involvement, producing unique brain activation patterns that can be detected through EEG [ 6 ]. The decoding of MI movements has significant implications for neuro-rehabilitation and assistive technologies, offering the potential to improve quality of life and restore functionality for individuals with motor impairments [ 7 ]. Additionally, in fields such as human-computer interaction and virtual reality, BCI-based MI decoding provides an intuitive way to interact, enabling effortless control of digital interfaces using only cognitive processes [ 8 ]. Achieving robust MI decoding in BCI systems necessitates expertise in EEG signal acquisition, feature extraction, and decoding algorithms [ 9 ]. Recent advancements have seen the application of fractal analysis and deep learning methods in motor imagery EEG classification [ 10 ]. Fractal analysis provides insights into the underlying neural mechanisms of motor imagery tasks by quantifying the self-similarity, irregularity, and nonlinearity in EEG signals [ 11 ]. This approach enhances understanding of neural oscillations, brain region synchronization, and motor network coordination, bolstering the reliability of BCI algorithms [ 12 ]. Furthermore, deep learning techniques, notably convolutional neural networks (CNNs) and recurrent neural networks (RNNs), offer powerful tools for decoding motor imagery EEG data [ 13 ]. By automatically learning hierarchical representations from raw data, these models bypass the need for handcrafted features, thereby enhancing adaptability to complex EEG datasets [ 14 ]. RNNs excel in modelling sequential dependencies, enabling analysis of temporal neural activity evolution during motor imagery tasks [ 15 ]. Optimized deep learning architectures facilitate real-time processing of EEG data, fostering interactive applications like BCIs and neuro-feedback systems [ 16 ]. In earlier studies on MI EEG classification, Zhang et. al [ 17 ] proposed an EEG inception architecture-based methodology to classify the MI EEG for intra-subject and inter-subject and achieved an accuracy of 88.4% and 88.6% for intra-subject as well as 65.88% and 77.44% for inter-subject on BCI Competition IV 2a and 2b respectively. Further, they checked the model size as well as model parameters for different layer depths with respect to accuracy and achieved best accuracy on 12- and 48-layer depths for 2 and 4 classes respectively. The model size of EEG inception for binary and multi-class were 10.83 MB and 34.1 MB respectively. Altaheri et. al [ 18 ] proposed a sliding window based ATCNet (Attention-Temporal Convolutional Network) for MI EEG classification and achieved accuracies of 85.38% and 70.97% for intra-subject and inter-subject on BCI Competition IV 2a respectively. Salami et. al [ 19 ] proposed an EEG-ITNet(Inception-Temporal Network) architecture and achieved accuracies of 76.74% and 78.74% for intra-subject and inter-subject classification on BCI Competition IV 2a. Zhao et. al [ 20 ] proposed a Convolutional Transformer Network (CTNet) architecture utilizing a convolutional layer in a Transformer architecture and achieved accuracies of 82.52% and 88.49% for intra-subject as well as 58.64% and 76.27% for inter-subject on BCI Competition IV 2a and 2b respectively. Zhi et. al [ 21 ] proposed a multi-domain convolution neural network utilizing spatial, frequency and time-frequency convolutions to design temporal-spatial-frequency convolution network (TSFCNet) and achieved 82.72% and 86.39% for intra-subject on BCI Competition IV 2a and 2b respectively. Jia et. al [ 22 ] proposed a temporal-spatial filtering and temporal-spatial convolution network and achieved an accuracy of 83% and 88% on BCI Competition IV 2a and 2b respectively. These studies on MI EEG classification showed that the methodologies were bulky, complex, time consuming and required high performing systems to deploy a model. To solve these limitations, we proposed a simple, lightweight, reliable, robust, and real time applicable methodology with inter-subject and intra-subject complexity handling capabilities for MI EEG classification. The proposed algorithm comprises several key stages, including pre-processing, common spatial pattern (CSP) filtering, wavelet analysis, fractal feature extraction, and the construction of a cross-correlation matrix [ 23 ]. Wavelet analysis, in particular, leverages two primary functions: the scaling function (φ, phi), which captures low-frequency components, and the wavelet function (ψ, psi), which highlights high-frequency details [ 24 ]. Furthermore, the wavelet coefficients obtained through this analysis can be utilized to extract fractal features, while cross-correlation matrices across multiple channels at various scales can be generated to form feature vectors for motor imagery (MI) EEG classification and a long-short term memory neural network [ 25 ] can be utilized for classification. For training and testing, the BCI Competition IV 2b [ 26 ] and 2a dataset [ 27 ] were utilized. Finally, we validated the proposed model and methodology by calculating the performance metrics (accuracy [ 28 ], precision [ 28 ], recall [ 28 ], f1-score [ 28 ], and Cohen’s kappa [ 29 ]). 2. Materials and Methods 2.1. Dataset Description In the present work, BCI competition data has been utilized (BCI competition data IV Dataset). BCI competition IV 2b dataset was acquired on 9 right-handed subjects, for five sessions per subject. The first two sessions of each subject were recorded without feedback, and the other three sessions were recorded with feedback from the subject [ 26 ]. BCI competition IV 2a dataset has 9 subjects Motor imagery EEG data. These data had been recorded by design a cue-based BCI paradigm protocol. There were two sessions for each subject on different days. Each session has 6 runs separated by short breaks. Each run has 48 trials (12 for each of the four possible classes) having total of 288 trials per session [ 27 ]. The sampling frequency was 250 Hz and the EEG data had been bandpass-filtered between 0.5 Hz and 100 Hz along with a 50 Hz notch filter. The BCI competition IV 2b and 2a has 3 and 22 channels of EEG along with 3 EOG channels respectively. The EOG Data was also sampled at 250Hz, and bandpass was filtered between 0.5 and 100 Hz with 50 Hz notch filtering [ 26 ][ 27 ]. 2.2. Pre-processing The pre-processing involved following operations: (i) band pass filtering of the data using Butterworth IIR filter of order 5 with frequency band 8 and 30 Hz, (ii) signal segmentation carried out to select the samples of signal in time frame of 3 to 6 seconds of data for each trial, (iii) baseline correction, (v) Normalization of data has been carried out using the Min–max normalization[ 30 ], as this is a good choice for normalizing where the bounds (maximum and minimum values) of the scores produced by a matcher are known. 2.3. Class Discrepancy Guided Sub-Band Filter-Based Common Spatial Pattern (CDGSBFCSP) To address the challenges in multi-class motor imagery classification in EEG signals, the Class Discrepancy-Guided Sub-Band Filter-Based Common Spatial Pattern (CDGSBFCSP) method was implemented [ 31 ]. This approach extended the traditional CSP by incorporating sub-band filtering and class discrepancy measures, which were particularly effective in extracting discriminative features across multiple classes. The method identified and utilized the most informative frequency sub-bands that differentiate between the motor imagery classes, allowing for better classification performance in multi-class settings [ 31 ]. After the pre-processing stage, the EEG signals were decomposed into multiple overlapping frequency sub-bands. This finer frequency resolution allowed to explore detailed neural oscillations across different bands, particularly important for multi-class motor imagery tasks [ 31 ]. For each sub-band, the class discrepancy was calculated to determine its ability to differentiate between the multiple motor imagery classes. The class discrepancy was measured by analysing the variance ratios of the classes within each sub-band [ 31 ]. $$\:{W}_{b}=\frac{{Var}_{class\:1,\:\:\:sub-band\:b}}{{Var}_{class\:2,\:\:\:sub-band\:b}}\dots\:..\frac{{Var}_{class\:1,\:\:\:sub-band\:b}}{{Var}_{class\:N,\:\:\:sub-band\:b}}$$ 2 Where W b calculated by (2) represented the discrepancy for sub-band b, and Var indicates the variance of signals across multiple classes. Sub-bands with the highest-class discrepancy values were selected for further processing, emphasizing those that provide the most significant distinctions among the multiple motor imagery tasks [ 31 ]. Given the multi-class nature of the data, One-vs-Rest CSP was employed, wherein binary CSP filters were learned for each class against all other classes. This was achieved by solving a generalized eigenvalue problem for each class, which optimized spatial filters to maximize variance for the target class while minimizing it for others [ 31 ]. The multi-class CSP filters were calculated by (3). $$\:{CSP}_{filters}=\frac{{W}_{b}^{T\:}{C}_{class\:i}{W}_{b}}{{W}_{b}^{T\:}{C}_{others}{W}_{b}}$$ 3 Where C class i was the covariance matrix of the target class, and C others ​corresponded to the covariance matrices of all other classes. 2.4. Continuous Wavelet Transform (CWT) The wavelet transform consists of expanding functions over wavelets, which are constructed from a single function by means of dilations and translations [ 32 ]. The formula for CWT is shown in (4). $$\:T\left(a,b\right)=\frac{1}{a}{\int\:}_{-\infty\:}^{\infty\:}Ѱ\left(\frac{t-b}{a}\right)f\left(t\right)dt,\:\text{a}>0,\:\text{b}\:\in\:R\:$$ 4 Where f(t) is signal and Ѱ(t) is a mother wavelet, a and b are dilation and translation factors respectively. 2.4.1. Morlet Wavelet (M) The Morlet wavelet is one of the most commonly used continuous wavelets. It is a complex-valued function and resembles a sinusoidal waveform modulated by a Gaussian envelope. The Morlet wavelet is particularly useful for analysing oscillatory components in signals, making it popular in the field of time-frequency analysis and neuroscience [ 33 ][ 34 ]. The equation of the Morlet wavelet is shown in (5). Ѱ M (t) = \(\:{e}^{-{t}^{2}/2}{e}^{i5t}\) (5) 2.4.2. Proposed step scaled wavelet fractals Wavelet fractals use the idea of wavelets to generate fractals. These combine the multi-resolution capabilities of wavelets with the self-similarity and scaling properties of fractals. The concept leverages wavelet transforms to analyse signals that exhibit fractal-like behaviour or structure. In this approach, Morlet wavelet was utilized to generate fractals by exploiting their self-similarity and recursive properties. By using wavelet to create fractals, we generated more complex, intricate patterns that are self-similar at different scales. Wavelet fractals represented a fascinating intersection of wavelet theory and fractal geometry, providing a rich framework for analysing and generating complex patterns that exhibit self-similarity across scales. Wavelet fractals were generated by taking mean statistical measure along time-axis of CWT coefficients at different scales between 1 and a maximum scale like 4, 8, 12, 16, 20 and 24 etc. Figure 1 is showing mean wavelet-fractals at different scale. Utilizing (5), the formulas for mean wavelet-fractal can be given by (6). Mean wavelet-fractal = µ wf (b) = \(\:\frac{\sum\:_{1}^{a}T(a,b)}{a-1}\) (6) Where, T(a, b) is CWT of EEG signal, ‘a’ is changing value from 4 to 32 with step size of 4, 6, 8, 10, 12, and 14. ‘b’ is a time point in CWT of EEG data across scales between 1 and a. 2.5. Cross correlation matrix (CCM) The equation for computing the Pearson correlation coefficient r between two time series X and Y to generate CCM is given by (7) [ 23 ]. $$\:r=\:\frac{{\sum\:}_{i=1}^{n}({X}_{i}-\stackrel{-}{X})({Y}_{i}-\stackrel{-}{Y})}{\sqrt{{\sum\:}_{i=1}^{n}({X}_{i}-\stackrel{-}{X}{)}^{2}}\:\sqrt{{\sum\:}_{i=1}^{n}({Y}_{i}-\stackrel{-}{Y}{)}^{2}}}$$ 7 Where X i , and Y i are the data points of the two-time series at time i , \(\:\stackrel{-}{X}\) and \(\:\stackrel{-}{Y}\) are the means of the respective time series, and n is total number of data points. The process of CCM generation explained and shown in Fig. 2 . 2.6. Proposed Bi-LSTM architecture LSTMs are a specialized type of Recurrent Neural Network (RNN) designed to effectively capture long-term dependencies by processing both past and future sequences, which is crucial for accurate predictions [ 25 ]. The customized LSTM architecture used in this study is outlined in Table I. The ' ReLU ' activation function was applied in all LSTM layers and the multi-perceptron layer, while the ' softmax ' activation function was utilized in the output layer of the model. The number of neurons in each layer was determined based on the number of classes (cl) in the dataset and the input shape, following a unique design that incorporates memory cells and gating mechanisms. These gates (input, forget, and output) enable the network to selectively retain, update, or discard information as it processes new data [ 25 ]. This structure helps LSTMs overcome the vanishing gradient problem and makes them more effective than traditional RNNs in learning long-term dependencies [ 25 ]. In addition, a Bi-directional LSTM (Bi-LSTM) was employed, which consists of two LSTM layers-one processing the input sequence in the forward direction (left to right) and the other in the reverse direction (right to left). This bidirectional processing is particularly beneficial for tasks requiring contextual information from both past and future time steps. The number of neurons in the architecture was further adjusted based on the number of channels (ch) present in the datasets to optimize performance. TABLE I Proposed Bi-LSTM architecture and model parameters Layer name Bidirectional Model Parameters Unidirectional Model Parameters Output shape (ch = 22, cl = 4) (ch = 3, cl = 2) Output shape (ch = 22, cl = 4) (ch = 3, cl = 2) Input layer (None, 3, ch×ch) 0 0 (None, 2, ch×ch) 0 0 LSTM Layer I (None, 3, 64×cl) 627712 37888 (None, 2, 64×cl) 758784 70656 LSTM layer II (None, 3, 32×cl) 164352 41216 (None, 2, 32×cl) 197120 49408 LSTM layer III (None, 3, 16×cl) 41216 10368 (None, 2, 16×cl) 49408 12416 Flatten layer (None, 3×16×cl) 0 0 (None, 2×16×cl) 0 0 Dense layer I (None, 32×cl) 24704 6208 (None, 32×cl) 16512 4160 Dense layer II (None, 8×cl) 4128 1040 (None, 8×cl) 4128 1040 Output layer (None, cl) 132 34 (None, cl) 132 34 Model Size(MB) 9.87 1.11 - 11.74 1.58 2.7. Training and evaluation scheme In the present study, the BCI Competition IV 2a and 2b dataset utilized for evaluation and validation. The EEG data was initially shuffled to get one of the best-balanced datasets and saved for further analysis. The training and testing data of each subject utilized for each wavelet was the same. The training data size was 90% of EEG data, the validation data was 5% of EEG data and 5% of EEG data was used for testing. The batch size of training was 144 for each case. The model was trained sequentially first without checkpoints for 100 epochs, afterward with checkpoints for 300 epochs to extract the model parameters corresponding to maximum validation accuracy. Finally, the proposed methodology was validated with 10-fold cross validation technique by calculating precision, recall, f1-score, and Cohen’s kappa. The number of testing samples, number validating samples and number of training samples were in the ratio of 5%, 5% and 90% respectively for 10-fold cross validation. 3. Results and Discussion This study proposed an innovative algorithm to classify MI EEG classification by utilizing CWT for generating fractal features and observing their evolution at different scales with the help of Bi-LSTM and LSTM architecture. The algorithm involved loading of EEG data, pre-processing, CDGSBFCSP features extraction, generating CWT coefficients at different scales, taking statistical measures to get a wavelet fractal along time axis, and calculating cross correlation matrix for these wavelet fractals of all channels as shown in Fig. 2 . These cross correlation matrices flattened to feed Bi-LSTM and LSTM architecture for classification. To understand the role of step size in proposed wavelet fractal, an ablation study was performed on subject 7 of BCI Competition IV 2a. The ablation study included 10-fold cross validation to check for testing accuracy and validation accuracy by varying the step size of proposed wavelet fractal. Additionally, the model size was also obtained after the final training and testing. The model size was varying due to the change in input shape of model. The input shape was obtained by calculating the number of steps required to achieve final scale value of 32 from initial scale value of 4 with a step size. The input shapes for each step size were (7, ch×ch), (5, ch×ch), (4, ch×ch), (3, ch×ch), (3, ch×ch) and (2, ch×ch) for 4, 6, 8, 10, 12 and 14 step size respectively. The average accuracy of each case was also calculated by obtained 10-fold validation and 10-fold testing accuracy. Further, the average accuracy, validation accuracy, testing accuracy and model size for each step size were plotted as shown in Fig. 3(a) for Bi-LSTM and Fig. 3(b) for LSTM. (b) Figure 3. 10-fold testing accuracy, 10-fold validation accuracy, 10-fold average accuracy and model size utilizing (a) Bi-LSTM, (b) LSTM and Morlet wavelet on subject 7 of BCI Competition IV 2a with respect to different step sizes of proposed wavelet fractal Next, the performance and model sizes of the Bi-LSTM and LSTM architectures were compared for different step sizes of wavelet fractals, as illustrated in Fig. 3. Figure 3(a) and Fig. 3(b) depict the testing accuracy, validation accuracy, and average accuracy (computed from testing and validation) for each step size of wavelet fractals, along with the corresponding model sizes of Bi-LSTM and LSTM, respectively. The best average accuracy of Bi-LSTM and LSTM architecture was obtained on step size of 10 and 14 respectively as shown in Fig. 3. The model sizes of Bi-LSTM and LSTM architecture were 9.87 and 12.07 MB respectively. The best performance was obtained by Bi-LSTM architecture with better model size than LSTM architecture. Further, the Bi-LSTM architecture with step size of 10 wavelet fractal was selected for analysis on BCI Competition IV 2a and 2b. TABLE II Performance metrics on BCI Competition IV 2b and 2a for intra-subject classification Datasets Performance Metrics Accuracy of Subjects (S) Mean 10-fold mean S1 S2 S3 S4 S5 S6 S7 S8 S9 BCI Competition IV 2b Accuracy 91.67 76.48 77.78 94.6 91.89 91.67 83.33 97.37 85 87.75 77.46 Mean Precision 91.88 76.84 77.71 92.06 92.06 92.86 82.86 97.73 84.76 88.53 78.04 Mean Recall 91.49 76.47 71.52 91.81 91.81 91.67 82.86 97.06 85 87.52 77.31 Mean F1-score 91.61 76.39 71.43 91.87 91.87 91.61 82.86 97.32 84.83 88.12 76.93 Cohen’s Kappa 0.83 0.53 0.55 0.84 0.84 0.83 0.66 0.95 0.68 0.73 0.55 BCI Competition IV 2a Accuracy 89.66 79.31 96.6 89.66 58.62 65.52 93.1 100 96.6 85.45 76.9 Mean Precision 87.5 84.97 96.43 88.54 66.62 79.69 93.33 100 97.73 88.31 78.48 Mean Recall 86.88 78.54 92.93 89.1 56.32 63.92 93.31 100 98.08 84.34 76.76 Mean F1-score 86.88 79.52 95.80 88.59 59.17 62.45 92.82 100 97.81 84.78 75.46 Cohen’s Kappa 0.86 0.71 0.95 0.86 0.41 0.52 0.91 1 0.95 0.8 0.69 Furthermore, the proposed algorithm evaluated with 10-fold cross validation technique to check for reliability and robustness of model and methodology by calculating accuracy, precision, recall, f1-score and Cohen’s kappa for each dataset as shown in Table II for intra-subject and Table III for inter-subject. The obtained performance metrics showed that the proposed methodology was reliable and robust. TABLE III Performance metrics on BCI Competition IV 2a and 2b for inter-subject classification Datasets n-fold mean Performance metrics Accuracy Mean precision Mean recall Mean f1-score Cohen’s Kappa BCI Competition IV 2b 1-fold 76.16 76.38 76.30 76.15 0.48 10-fold 74.01 74.18 73.88 71.77 0.44 BCI Competition IV 2a 1-fold 71.76 71.68 70.18 70.39 0.6 10-fold 67.39 68 67 67 0.55 Finally, we compared our results with earlier proposed state-of-the-art algorithms, the proposed algorithm achieved better classification accuracy through cross correlation matrix based on mean wavelet-fractal utilizing CWT coefficients at different scale followed by Bi-LSTM based algorithm that evaluates the evolution in CCMs and performs classification. The comparison of performance with the state-of-the-art algorithms and the proposed algorithm has been compiled in Table IV and Table V for intra-subject and inter-subject classification respectively. In the current work, the proposed methodology is lightweight, reliable, robust, and practically applicable to real-time scenarios for MI EEG classification. It has implications on embedded based device implementation for MI classification. The obtained results were optimum with CCMs based on mean wavelet-fractal utilizing Morlet. Besides it, the step scaled mean wavelet-fractals obtained were also unique as compared to each other as shown in Fig. 1 . TABLE IV Performance metrics on BCI Competition IV 2b and 2a for intra-subject classification Dataset Authors and year Methodology Model Mean accuracy (%) Improvement in accuracy BCI Competition IV 2b H.K. Lee et.al (2019) [ 35 ] CWT (Morlet) CNN 83 4.75 CWT (Mexican hat) CNN 81.2 6.55 X. Tang et.al (2020) [ 36 ] Conditional Empirical Mode Decomposition 1D Multi-scale CNN 82.6 5.15 Zhang et. al (2021) [ 17 ] Augmentation EEG inception 88.58 -0.83 Jia et. al (2022) [ 22 ] Temporal-spatial filtering and convolution CNN 88 -0.25 Zhi et. al (2023) [ 21 ] Multi-domain convolution CNN 86.39 1.36 Zhao et. al (2024) [ 20 ] CTNet 88.49 -0.74 Ours CDGSBFCSP + CWT + proposed wavelet fractals + CCMs Bi-LSTM 87.75 Mean = 2.28 BCI Competition IV 2a Zhang et. al (2021) [ 17 ] Augmentation EEG Inception 88.39 -2.94 Jia et. al (2022) [ 22 ] Temporal-spatial filtering and convolution CNN 83 2.45 Salami et. al (2022) [ 19 ] EEG-ITNet 76.74 8.71 Zhi et. al (2023) [ 21 ] Multi-domain convolution CNN 82.72 2.73 Altaheri et. al (2023) [ 18 ] Sliding window ATCNet 85.38 0.07 Zhao et. al (2024) [ 20 ] CTNet 82.52 2.93 Ours CDGSBFCSP + CWT + proposed wavelet fractals + CCMs Bi-LSTM 85.45 Mean = 2.33 In future research, we aim to further enhance the proposed methodology by addressing key challenges in multi-class motor imagery (MI) EEG classification. One of the primary directions will involve developing an attention-based framework to dynamically analyse and capture the evolution of discriminative features across multiple wavelet fractals simultaneously. This approach is expected to improve the identification of class-specific patterns, leading to more robust feature extraction and classification performance. Further, we plan to extend this methodology for implementation in a real-time brain-computer interface (BCI) system aimed at addressing the complexities involved in developing practical BCI devices. Specifically, the proposed approach will be optimized for commercialization in prosthetic limb control for individuals with paralysis. TABLE V Comparison with earlier studies for inter-subject classification Dataset Authors and year Methodology Model Accuracy (%) Improvement in accuracy BCI Competition IV 2b Hermosilla et.al (2021) [ 37 ] Temporal and spatial features ShallowConvNet 75.33 0.83 Zhang et. al (2021) [ 17 ] Augmentation EEG inception 77.44 -1.28 Mammone et.al (2023) [ 38 ] Autoencoder based FBCSP FNN 74.75 1.41 Zhao et. al (2024) [ 20 ] CTNet 76.27 -0.11 Ours CDGSBFCSP + CWT + proposed wavelet fractals + CCMs Bi-LSTM 76.16 Mean = 0.22 BCI Competition IV 2a Zhang et. al (2021) [ 17 ] Augmentation EEG inception 65.88 5.88 Salami et. al (2022) [ 19 ] Fine tuning EEG-ITNet 78.74 -6.98 Altaheri et. al (2023) [ 18 ] Sliding window ATCNet 70.97 0.79 Zhao et. al (2024) [ 20 ] CTNet 58.64 13.12 Ours CDGSBFCSP + CWT + proposed wavelet fractals + CCMs Bi-LSTM 71.76 Mean = 3.21 4. Conclusion In this study, we proposed a novel algorithm for multi-class motor imagery (MI) EEG classification using a lightweight, customized Bi-LSTM network. The methodology integrated CDGSBFCSP, continuous wavelet transform (CWT), wavelet fractal generation, and cross-correlation matrices, offering a practical and effective approach for MI EEG classification. The evolution of CCMs across all channels, based on mean wavelet-fractal features, was analysed and utilized to enhance classification performance. For intra-subject classification, the proposed algorithm achieved best accuracies of 85.45% and 87.75%, demonstrating improvements of 2.33% and 2.28% compared to earlier state-of-the-art methods on the BCI Competition IV 2a and 2b datasets, respectively. Similarly, for inter-subject classification, the best accuracies achieved were 71.76% and 76.16%, with improvements of 3.21% and 0.22% over prior methods on the same datasets respectively. The customized Bi-LSTM architecture exhibited a compact model size of 9.87 MB for BCI Competition IV 2a and 1.11 MB for BCI Competition IV 2b, highlighting its suitability for embedded device implementation in real-time MI EEG classification tasks. In the future, we plan to develop an attention-based approach to capture the evolution of wavelet-fractal features more effectively and to design a brain-computer interface (BCI) system suitable for real-time applications. Declarations Conflict of Interest The authors declare no conflict of interest. Ethical approval No ethical approval is required as there was no involvement of human /animal participation in this study. Funding The Science for Equity Empowerment and Development Division, Department of Science and Technology, govt. of India [SEED/TIDE/2019/340], has provided funds for the necessary systems and support for completing this implementation. Author Contribution All authors contributed to the study's conception and design. Balendra, Neeraj Sharma, and Shiru Sharma performed implementation and analysis. Balendra wrote the manuscript's first draft, and all authors commented on previous versions. All authors read and approved the final manuscript. Acknowledgement All authors would like to thank the Science for Equity Empowerment and Development Division (SEED), Department of Science and Technology (DST), Government of India, for providing funds for the necessary systems and support for completing this implementation. The authors are also very obliged to the Prime Minister Research Fellowship (PMRF) for providing financial motivation to work on this study. References Wolpaw, J.R., 2013. Brain–computer interfaces. In Handbook of clinical neurology (Vol. 110, pp. 67–74). Lebedev, M.A. and Nicolelis, M.A., 2006. Brain–machine interfaces: past, present and future. TRENDS in Neurosciences, 29 (9), pp.536–546. Hochberg, L.R., Bacher, D., Jarosiewicz, B., Masse, N.Y., Simeral, J.D., Vogel, J., Haddadin, S., Liu, J., Cash, S.S., Van Der Smagt, P. and Donoghue, J.P., 2012. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, 485 (7398), pp.372–375. Nicolelis, M.A., 2003. Brain–machine interfaces to restore motor function and probe neural circuits. Nature Reviews Neuroscience, 4 (5), pp.417–422. Pfurtscheller, G. and Neuper, C., 1997. Motor imagery activates primary sensorimotor area in humans. Neuroscience letters, 239 (2–3), pp.65–68. Jeannerod, M., 1994. The representing brain: Neural correlates of motor intention and imagery. Behavioral and Brain sciences, 17 (2), pp.187–202. Daly, J.J. and Wolpaw, J.R., 2008. Brain–computer interfaces in neurological rehabilitation. The Lancet Neurology, 7 (11), pp.1032–1043. Charte, D., Charte, F., García, S., del Jesus, M.J. and Herrera, F., 2018. A practical tutorial on autoencoders for nonlinear feature fusion: Taxonomy, models, software and guidelines. Information Fusion, 44 , pp.78–96. Guger, C., Edlinger, G., Harkam, W., Niedermayer, I. and Pfurtscheller, G., 2003. How many people are able to operate an EEG-based brain-computer interface (BCI)?. IEEE transactions on neural systems and rehabilitation engineering, 11 (2), pp.145–147. Bashashati, A., Fatourechi, M., Ward, R.K. and Birch, G.E., 2007. A survey of signal processing algorithms in brain–computer interfaces based on electrical brain signals. Journal of Neural engineering , 4 (2), p.R32. Kantz, H. and Schreiber, T., 2003. Nonlinear time series analysis . Cambridge university press. Wronkiewicz, M., Larson, E. and Lee, A.K., 2016. Incorporating modern neuroscience findings to improve brain–computer interfaces: tracking auditory attention. Journal of neural engineering , 13 (5), p.056017. Schirrmeister, R.T., Springenberg, J.T., Fiederer, L.D.J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W. and Ball, T., 2017. Deep learning with convolutional neural networks for EEG decoding and visualization. Human brain mapping, 38 (11), pp.5391–5420. She, Q., Hu, B., Luo, Z., Nguyen, T. and Zhang, Y., 2019. A hierarchical semi-supervised extreme learning machine method for EEG recognition. Medical & biological engineering & computing, 57 , pp.147–157. León, J., Escobar, J.J., Ortiz, A., Ortega, J., González, J., Martín-Smith, P., Gan, J.Q. and Damas, M., 2020. Deep learning for EEG-based Motor Imagery classification: Accuracy-cost trade-off. Plos one, 15 (6), p .e 0234178. Lotte, F., Congedo, M., Lécuyer, A., Lamarche, F. and Arnaldi, B., 2007. A review of classification algorithms for EEG-based brain–computer interfaces. Journal of neural engineering, 4 (2), p.R1. Zhang, C., Kim, Y.K. and Eskandarian, A., 2021. EEG-inception: an accurate and robust end-to-end neural network for EEG-based motor imagery classification. Journal of Neural Engineering , 18 (4), p.046014. Altaheri, H., Muhammad, G. and Alsulaiman, M., 2022. Physics-informed attention temporal convolutional network for EEG-based motor imagery classification. IEEE transactions on industrial informatics, 19 (2), pp.2249–2258. Salami, A., Andreu-Perez, J. and Gillmeister, H., 2022. EEG-ITNet: An explainable inception temporal convolutional network for motor imagery classification. IEEE Access, 10 , pp.36672–36685. Zhao, W., Jiang, X., Zhang, B., Xiao, S. and Weng, S., 2024. CTNet: a convolutional transformer network for EEG-based motor imagery classification. Scientific Reports , 14 (1), p.20237. H. Zhi, Z. Yu, T. Yu, Z. Gu and J. Yang, "A Multi-Domain Convolutional Neural Network for EEG-Based Motor Imagery Decoding," in IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 3988–3998, 2023 Jia, X., Song, Y., Yang, L. and Xie, L., 2022. Joint spatial and temporal features extraction for multi-classification of motor imagery EEG. Biomedical Signal Processing and Control , 71 , p.103247. Pearson, K., 1920. Notes on the history of correlation. Biometrika, 13 (1), pp.25–45. Daubechies, I., 1992. Ten lectures on wavelets. Society for industrial and applied mathematics . Hochreiter, S., 1997. Long Short-term Memory. Neural Computation MIT-Press . Leeb, R., Brunner, C., Müller-Putz, G., Schlögl, A. and Pfurtscheller, G.J.G.U.O.T., 2008. BCI Competition 2008–Graz data set B. Graz University of Technology, Austria, 16 , pp.1–6. Brunner, C., Leeb, R., Müller-Putz, G., Schlögl, A. and Pfurtscheller, G., 2008. BCI Competition 2008–Graz data set A. Institute for knowledge discovery (laboratory of brain-computer interfaces), Graz University of Technology , 16 , pp.1–6. Hossin, M. and Sulaiman, M.N., 2015. A review on evaluation metrics for data classification evaluations. International journal of data mining & knowledge management process , 5 (2), p.1. Vieira, S.M., Kaymak, U. and Sousa, J.M., 2010, July. Cohen's kappa coefficient as a performance measure for feature selection. In International conference on fuzzy systems (pp. 1–8). IEEE. Jain, A., Nandakumar, K. and Ross, A., 2005. Score normalization in multimodal biometric systems. Pattern recognition, 38 (12), pp.2270–2285. Luo, J., Wang, J., Xu, R. and Xu, K., 2019. Class discrepancy-guided sub-band filter-based common spatial pattern for motor imagery classification. Journal of neuroscience methods, 323 , pp.98–107. Strang, G., 1994. Wavelets. American Scientist, 82 (3), pp.250–255. Büssow, R., 2007. An algorithm for the continuous Morlet wavelet transform. Mechanical Systems and Signal Processing, 21 (8), pp.2970–2979. Torrence, C. and Compo, G.P., 1998. A practical guide to wavelet analysis. Bulletin of the American Meteorological society, 79 (1), pp.61–78. Lee, H.K. and Choi, Y.S., 2019. Application of continuous wavelet transform and convolutional neural network in decoding motor imagery brain-computer interface. Entropy , 21 (12), p.1199. Tang, X., Li, W., Li, X., Ma, W. and Dang, X., 2020. Motor imagery EEG recognition based on conditional optimization empirical mode decomposition and multi-scale convolutional neural network. Expert Systems with Applications , 149 , p.113285. Hermosilla, D.M., Codorniú, R.T., Baracaldo, R.L., Zamora, R.S., Rodriguez, D.D., Albuerne, Y.L. and Álvarez, J.R.N., 2021. Shallow convolutional network excel for classifying motor imagery EEG in BCI applications. IEEE Access, 9 , pp.98275–98286. Mammone, N., Ieracitano, C., Adeli, H. and Morabito, F.C., 2023. AutoEncoder filter bank common spatial patterns to decode motor imagery from EEG. IEEE journal of biomedical and health informatics, 27 (5), pp.2365–2376. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5750495","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":398790774,"identity":"300f657f-7771-4977-8da2-53eeb6908f20","order_by":0,"name":"Balendra .","email":"data:image/png;base64,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","orcid":"","institution":"Indian Institute of Technology Varanasi","correspondingAuthor":true,"prefix":"","firstName":"Balendra","middleName":"","lastName":".","suffix":""},{"id":398790775,"identity":"9e9212f5-5a05-477e-a75e-d86848a641de","order_by":1,"name":"Neeraj Sharma","email":"","orcid":"","institution":"Indian Institute of Technology Varanasi","correspondingAuthor":false,"prefix":"","firstName":"Neeraj","middleName":"","lastName":"Sharma","suffix":""},{"id":398790776,"identity":"d2c36058-1ab8-4f34-9e55-e1d34c023aa4","order_by":2,"name":"Shiru Sharma","email":"","orcid":"","institution":"Indian Institute of Technology Varanasi","correspondingAuthor":false,"prefix":"","firstName":"Shiru","middleName":"","lastName":"Sharma","suffix":""}],"badges":[],"createdAt":"2025-01-02 09:23:09","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-5750495/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5750495/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":73386068,"identity":"518456d1-cf18-4907-b88f-934321aa83a6","added_by":"auto","created_at":"2025-01-09 12:02:06","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":265861,"visible":true,"origin":"","legend":"\u003cp\u003eStep scaled mean wavelet-fractals of step size 2 utilizing Morlet\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-5750495/v1/500d4bd3844b3ef7edd4a35e.png"},{"id":73387371,"identity":"5598e585-b1fa-4a76-9297-3e81a09d4adb","added_by":"auto","created_at":"2025-01-09 12:18:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":399849,"visible":true,"origin":"","legend":"\u003cp\u003eProposed methodology and FCM generation from Wavelet-fractals utilizing Morlet\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-5750495/v1/3de68af7ce06eb18feb4446d.png"},{"id":73386918,"identity":"9510ef12-d213-4e40-a431-3fb198c5f78e","added_by":"auto","created_at":"2025-01-09 12:10:07","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":80812,"visible":true,"origin":"","legend":"\u003cp\u003e10-fold testing accuracy, 10-fold validation accuracy, 10-fold average accuracy and model size utilizing (a) Bi-LSTM, (b) LSTM and Morlet wavelet on subject 7 of BCI Competition IV 2a with respect to different step sizes of proposed wavelet fractal\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5750495/v1/657e75cdad915832852a0679.jpg"},{"id":73389036,"identity":"74566dd9-1610-4f5b-ad56-5acb4bd73d47","added_by":"auto","created_at":"2025-01-09 12:34:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1959679,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5750495/v1/1a80dd7c-31d2-43ed-8668-c2a49aec6c1c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A lightweight methodology for Motor Imagery EEG classification utilizing step scaled wavelet fractals and Bi-LSTM architecture","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBrain Computer Interface (BCI) technology represents a cutting-edge area of research that integrates neuroscience, engineering, and computer science to establish direct communication links between the human brain and external systems [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. In contrast to traditional neuromuscular pathways that depend on peripheral nerves and muscles, BCIs provide an innovative solution, allowing individuals to control devices solely through their cognitive processes [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. This shift in approach has the potential to be transformative, especially for individuals with motor disabilities, granting them greater independence and control over their actions [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. At the core of BCI systems is the process of decoding neural activity, where sophisticated algorithms interpret brain signals and convert them into actionable commands for external devices [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. A key application in this field focuses on decoding motor imagery (MI) movements from electroencephalography (EEG) signals captured during motor imagery tasks [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Motor imagery involves mentally visualizing physical movements without actual muscle involvement, producing unique brain activation patterns that can be detected through EEG [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. The decoding of MI movements has significant implications for neuro-rehabilitation and assistive technologies, offering the potential to improve quality of life and restore functionality for individuals with motor impairments [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Additionally, in fields such as human-computer interaction and virtual reality, BCI-based MI decoding provides an intuitive way to interact, enabling effortless control of digital interfaces using only cognitive processes [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Achieving robust MI decoding in BCI systems necessitates expertise in EEG signal acquisition, feature extraction, and decoding algorithms [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Recent advancements have seen the application of fractal analysis and deep learning methods in motor imagery EEG classification [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Fractal analysis provides insights into the underlying neural mechanisms of motor imagery tasks by quantifying the self-similarity, irregularity, and nonlinearity in EEG signals [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. This approach enhances understanding of neural oscillations, brain region synchronization, and motor network coordination, bolstering the reliability of BCI algorithms [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Furthermore, deep learning techniques, notably convolutional neural networks (CNNs) and recurrent neural networks (RNNs), offer powerful tools for decoding motor imagery EEG data [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. By automatically learning hierarchical representations from raw data, these models bypass the need for handcrafted features, thereby enhancing adaptability to complex EEG datasets [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. RNNs excel in modelling sequential dependencies, enabling analysis of temporal neural activity evolution during motor imagery tasks [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Optimized deep learning architectures facilitate real-time processing of EEG data, fostering interactive applications like BCIs and neuro-feedback systems [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn earlier studies on MI EEG classification, Zhang et. al [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] proposed an EEG inception architecture-based methodology to classify the MI EEG for intra-subject and inter-subject and achieved an accuracy of 88.4% and 88.6% for intra-subject as well as 65.88% and 77.44% for inter-subject on BCI Competition IV 2a and 2b respectively. Further, they checked the model size as well as model parameters for different layer depths with respect to accuracy and achieved best accuracy on 12- and 48-layer depths for 2 and 4 classes respectively. The model size of EEG inception for binary and multi-class were 10.83 MB and 34.1 MB respectively. Altaheri et. al [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] proposed a sliding window based ATCNet (Attention-Temporal Convolutional Network) for MI EEG classification and achieved accuracies of 85.38% and 70.97% for intra-subject and inter-subject on BCI Competition IV 2a respectively. Salami et. al [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] proposed an EEG-ITNet(Inception-Temporal Network) architecture and achieved accuracies of 76.74% and 78.74% for intra-subject and inter-subject classification on BCI Competition IV 2a. Zhao et. al [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] proposed a Convolutional Transformer Network (CTNet) architecture utilizing a convolutional layer in a Transformer architecture and achieved accuracies of 82.52% and 88.49% for intra-subject as well as 58.64% and 76.27% for inter-subject on BCI Competition IV 2a and 2b respectively. Zhi et. al [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] proposed a multi-domain convolution neural network utilizing spatial, frequency and time-frequency convolutions to design temporal-spatial-frequency convolution network (TSFCNet) and achieved 82.72% and 86.39% for intra-subject on BCI Competition IV 2a and 2b respectively. Jia et. al [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] proposed a temporal-spatial filtering and temporal-spatial convolution network and achieved an accuracy of 83% and 88% on BCI Competition IV 2a and 2b respectively. These studies on MI EEG classification showed that the methodologies were bulky, complex, time consuming and required high performing systems to deploy a model.\u003c/p\u003e \u003cp\u003eTo solve these limitations, we proposed a simple, lightweight, reliable, robust, and real time applicable methodology with inter-subject and intra-subject complexity handling capabilities for MI EEG classification. The proposed algorithm comprises several key stages, including pre-processing, common spatial pattern (CSP) filtering, wavelet analysis, fractal feature extraction, and the construction of a cross-correlation matrix [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Wavelet analysis, in particular, leverages two primary functions: the scaling function (φ, phi), which captures low-frequency components, and the wavelet function (ψ, psi), which highlights high-frequency details [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Furthermore, the wavelet coefficients obtained through this analysis can be utilized to extract fractal features, while cross-correlation matrices across multiple channels at various scales can be generated to form feature vectors for motor imagery (MI) EEG classification and a long-short term memory neural network [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] can be utilized for classification. For training and testing, the BCI Competition IV 2b [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] and 2a dataset [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] were utilized. Finally, we validated the proposed model and methodology by calculating the performance metrics (accuracy [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], precision [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], recall [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], f1-score [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], and Cohen\u0026rsquo;s kappa [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]).\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Dataset Description\u003c/h2\u003e \u003cp\u003eIn the present work, BCI competition data has been utilized (BCI competition data IV Dataset). BCI competition IV 2b dataset was acquired on 9 right-handed subjects, for five sessions per subject. The first two sessions of each subject were recorded without feedback, and the other three sessions were recorded with feedback from the subject [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. BCI competition IV 2a dataset has 9 subjects Motor imagery EEG data. These data had been recorded by design a cue-based BCI paradigm protocol. There were two sessions for each subject on different days. Each session has 6 runs separated by short breaks. Each run has 48 trials (12 for each of the four possible classes) having total of 288 trials per session [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The sampling frequency was 250 Hz and the EEG data had been bandpass-filtered between 0.5 Hz and 100 Hz along with a 50 Hz notch filter. The BCI competition IV 2b and 2a has 3 and 22 channels of EEG along with 3 EOG channels respectively. The EOG Data was also sampled at 250Hz, and bandpass was filtered between 0.5 and 100 Hz with 50 Hz notch filtering [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e][\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Pre-processing\u003c/h2\u003e \u003cp\u003eThe pre-processing involved following operations: (i) band pass filtering of the data using Butterworth IIR filter of order 5 with frequency band 8 and 30 Hz, (ii) signal segmentation carried out to select the samples of signal in time frame of 3 to 6 seconds of data for each trial, (iii) baseline correction, (v) Normalization of data has been carried out using the Min\u0026ndash;max normalization[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], as this is a good choice for normalizing where the bounds (maximum and minimum values) of the scores produced by a matcher are known.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Class Discrepancy Guided Sub-Band Filter-Based Common Spatial Pattern (CDGSBFCSP)\u003c/h2\u003e \u003cp\u003eTo address the challenges in multi-class motor imagery classification in EEG signals, the Class Discrepancy-Guided Sub-Band Filter-Based Common Spatial Pattern (CDGSBFCSP) method was implemented [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. This approach extended the traditional CSP by incorporating sub-band filtering and class discrepancy measures, which were particularly effective in extracting discriminative features across multiple classes. The method identified and utilized the most informative frequency sub-bands that differentiate between the motor imagery classes, allowing for better classification performance in multi-class settings [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. After the pre-processing stage, the EEG signals were decomposed into multiple overlapping frequency sub-bands. This finer frequency resolution allowed to explore detailed neural oscillations across different bands, particularly important for multi-class motor imagery tasks [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. For each sub-band, the class discrepancy was calculated to determine its ability to differentiate between the multiple motor imagery classes. The class discrepancy was measured by analysing the variance ratios of the classes within each sub-band [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{W}_{b}=\\frac{{Var}_{class\\:1,\\:\\:\\:sub-band\\:b}}{{Var}_{class\\:2,\\:\\:\\:sub-band\\:b}}\\dots\\:..\\frac{{Var}_{class\\:1,\\:\\:\\:sub-band\\:b}}{{Var}_{class\\:N,\\:\\:\\:sub-band\\:b}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere \u003cem\u003eW\u003c/em\u003e\u003csub\u003e\u003cem\u003eb\u003c/em\u003e\u003c/sub\u003e calculated by (2) represented the discrepancy for sub-band b, and \u003cem\u003eVar\u003c/em\u003e indicates the variance of signals across multiple classes. Sub-bands with the highest-class discrepancy values were selected for further processing, emphasizing those that provide the most significant distinctions among the multiple motor imagery tasks [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eGiven the multi-class nature of the data, One-vs-Rest CSP was employed, wherein binary CSP filters were learned for each class against all other classes. This was achieved by solving a generalized eigenvalue problem for each class, which optimized spatial filters to maximize variance for the target class while minimizing it for others [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. The multi-class CSP filters were calculated by (3).\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{CSP}_{filters}=\\frac{{W}_{b}^{T\\:}{C}_{class\\:i}{W}_{b}}{{W}_{b}^{T\\:}{C}_{others}{W}_{b}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere C\u003csub\u003eclass i\u003c/sub\u003e was the covariance matrix of the target class, and C\u003csub\u003eothers\u003c/sub\u003e ​corresponded to the covariance matrices of all other classes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Continuous Wavelet Transform (CWT)\u003c/h2\u003e \u003cp\u003eThe wavelet transform consists of expanding functions over wavelets, which are constructed from a single function by means of dilations and translations [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. The formula for CWT is shown in (4).\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:T\\left(a,b\\right)=\\frac{1}{a}{\\int\\:}_{-\\infty\\:}^{\\infty\\:}Ѱ\\left(\\frac{t-b}{a}\\right)f\\left(t\\right)dt,\\:\\text{a}\u0026gt;0,\\:\\text{b}\\:\\in\\:R\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere f(t) is signal and Ѱ(t) is a mother wavelet, a and b are dilation and translation factors respectively.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1. Morlet Wavelet (M)\u003c/h2\u003e \u003cp\u003eThe Morlet wavelet is one of the most commonly used continuous wavelets. It is a complex-valued function and resembles a sinusoidal waveform modulated by a Gaussian envelope. The Morlet wavelet is particularly useful for analysing oscillatory components in signals, making it popular in the field of time-frequency analysis and neuroscience [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e][\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. The equation of the Morlet wavelet is shown in (5).\u003c/p\u003e \u003cp\u003eѰ\u003csub\u003eM\u003c/sub\u003e(t) = \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{e}^{-{t}^{2}/2}{e}^{i5t}\\)\u003c/span\u003e\u003c/span\u003e (5)\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2. Proposed step scaled wavelet fractals\u003c/h2\u003e \u003cp\u003eWavelet fractals use the idea of wavelets to generate fractals. These combine the multi-resolution capabilities of wavelets with the self-similarity and scaling properties of fractals. The concept leverages wavelet transforms to analyse signals that exhibit fractal-like behaviour or structure. In this approach, Morlet wavelet was utilized to generate fractals by exploiting their self-similarity and recursive properties. By using wavelet to create fractals, we generated more complex, intricate patterns that are self-similar at different scales. Wavelet fractals represented a fascinating intersection of wavelet theory and fractal geometry, providing a rich framework for analysing and generating complex patterns that exhibit self-similarity across scales. Wavelet fractals were generated by taking mean statistical measure along time-axis of CWT coefficients at different scales between 1 and a maximum scale like 4, 8, 12, 16, 20 and 24 etc. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e is showing mean wavelet-fractals at different scale. Utilizing (5), the formulas for mean wavelet-fractal can be given by (6).\u003c/p\u003e \u003cp\u003eMean wavelet-fractal\u0026thinsp;=\u0026thinsp;\u0026micro;\u003csub\u003ewf\u003c/sub\u003e(b) = \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{\\sum\\:_{1}^{a}T(a,b)}{a-1}\\)\u003c/span\u003e\u003c/span\u003e (6)\u003c/p\u003e \u003cp\u003eWhere, T(a, b) is CWT of EEG signal, \u003cem\u003e\u0026lsquo;a\u0026rsquo;\u003c/em\u003e is changing value from 4 to 32 with step size of 4, 6, 8, 10, 12, and 14. \u0026lsquo;b\u0026rsquo; is a time point in CWT of EEG data across scales between 1 and a.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Cross correlation matrix (CCM)\u003c/h2\u003e \u003cp\u003eThe equation for computing the Pearson correlation coefficient r between two time series X and Y to generate CCM is given by (7) [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:r=\\:\\frac{{\\sum\\:}_{i=1}^{n}({X}_{i}-\\stackrel{-}{X})({Y}_{i}-\\stackrel{-}{Y})}{\\sqrt{{\\sum\\:}_{i=1}^{n}({X}_{i}-\\stackrel{-}{X}{)}^{2}}\\:\\sqrt{{\\sum\\:}_{i=1}^{n}({Y}_{i}-\\stackrel{-}{Y}{)}^{2}}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere X\u003csub\u003ei\u003c/sub\u003e, and Y\u003csub\u003ei\u003c/sub\u003e are the data points of the two-time series at time \u003cem\u003ei\u003c/em\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{-}{X}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{-}{Y}\\)\u003c/span\u003e\u003c/span\u003e are the means of the respective time series, and n is total number of data points. The process of CCM generation explained and shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.6. Proposed Bi-LSTM architecture\u003c/h2\u003e \u003cp\u003eLSTMs are a specialized type of Recurrent Neural Network (RNN) designed to effectively capture long-term dependencies by processing both past and future sequences, which is crucial for accurate predictions [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The customized LSTM architecture used in this study is outlined in Table I. The '\u003cem\u003eReLU\u003c/em\u003e' activation function was applied in all LSTM layers and the multi-perceptron layer, while the '\u003cem\u003esoftmax\u003c/em\u003e' activation function was utilized in the output layer of the model. The number of neurons in each layer was determined based on the number of classes (cl) in the dataset and the input shape, following a unique design that incorporates memory cells and gating mechanisms. These gates (input, forget, and output) enable the network to selectively retain, update, or discard information as it processes new data [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. This structure helps LSTMs overcome the vanishing gradient problem and makes them more effective than traditional RNNs in learning long-term dependencies [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. In addition, a Bi-directional LSTM (Bi-LSTM) was employed, which consists of two LSTM layers-one processing the input sequence in the forward direction (left to right) and the other in the reverse direction (right to left). This bidirectional processing is particularly beneficial for tasks requiring contextual information from both past and future time steps. The number of neurons in the architecture was further adjusted based on the number of channels (ch) present in the datasets to optimize performance.\u003c/p\u003e \u003cp\u003eTABLE I\u003c/p\u003e \u003cp\u003eProposed Bi-LSTM architecture and model parameters\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLayer name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eBidirectional Model Parameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eUnidirectional Model Parameters\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOutput shape\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(ch\u0026thinsp;=\u0026thinsp;22, cl\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(ch\u0026thinsp;=\u0026thinsp;3, cl\u0026thinsp;=\u0026thinsp;2)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOutput shape\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e(ch\u0026thinsp;=\u0026thinsp;22, cl\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e(ch\u0026thinsp;=\u0026thinsp;3, cl\u0026thinsp;=\u0026thinsp;2)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInput layer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 3, ch\u0026times;ch)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 2, ch\u0026times;ch)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM Layer I\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 3, 64\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e627712\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e37888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 2, 64\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e758784\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e70656\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM layer II\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 3, 32\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e164352\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e41216\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 2, 32\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e197120\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e49408\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM layer III\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 3, 16\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41216\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10368\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 2, 16\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e49408\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e12416\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlatten layer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 3\u0026times;16\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 2\u0026times;16\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDense layer I\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 32\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6208\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 32\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e16512\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4160\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDense layer II\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, 8\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1040\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, 8\u0026times;cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1040\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOutput layer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(None, cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e132\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(None, cl)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e132\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e34\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eModel Size(MB)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e11.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.58\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.7. Training and evaluation scheme\u003c/h2\u003e \u003cp\u003eIn the present study, the BCI Competition IV 2a and 2b dataset utilized for evaluation and validation. The EEG data was initially shuffled to get one of the best-balanced datasets and saved for further analysis. The training and testing data of each subject utilized for each wavelet was the same. The training data size was 90% of EEG data, the validation data was 5% of EEG data and 5% of EEG data was used for testing. The batch size of training was 144 for each case. The model was trained sequentially first without checkpoints for 100 epochs, afterward with checkpoints for 300 epochs to extract the model parameters corresponding to maximum validation accuracy. Finally, the proposed methodology was validated with 10-fold cross validation technique by calculating precision, recall, f1-score, and Cohen\u0026rsquo;s kappa. The number of testing samples, number validating samples and number of training samples were in the ratio of 5%, 5% and 90% respectively for 10-fold cross validation.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cp\u003eThis study proposed an innovative algorithm to classify MI EEG classification by utilizing CWT for generating fractal features and observing their evolution at different scales with the help of Bi-LSTM and LSTM architecture. The algorithm involved loading of EEG data, pre-processing, CDGSBFCSP features extraction, generating CWT coefficients at different scales, taking statistical measures to get a wavelet fractal along time axis, and calculating cross correlation matrix for these wavelet fractals of all channels as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. These cross correlation matrices flattened to feed Bi-LSTM and LSTM architecture for classification. To understand the role of step size in proposed wavelet fractal, an ablation study was performed on subject 7 of BCI Competition IV 2a. The ablation study included 10-fold cross validation to check for testing accuracy and validation accuracy by varying the step size of proposed wavelet fractal. Additionally, the model size was also obtained after the final training and testing. The model size was varying due to the change in input shape of model. The input shape was obtained by calculating the number of steps required to achieve final scale value of 32 from initial scale value of 4 with a step size. The input shapes for each step size were (7, ch\u0026times;ch), (5, ch\u0026times;ch), (4, ch\u0026times;ch), (3, ch\u0026times;ch), (3, ch\u0026times;ch) and (2, ch\u0026times;ch) for 4, 6, 8, 10, 12 and 14 step size respectively. The average accuracy of each case was also calculated by obtained 10-fold validation and 10-fold testing accuracy. Further, the average accuracy, validation accuracy, testing accuracy and model size for each step size were plotted as shown in Fig.\u0026nbsp;3(a) for Bi-LSTM and Fig.\u0026nbsp;3(b) for LSTM.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e(b)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;3. 10-fold testing accuracy, 10-fold validation accuracy, 10-fold average accuracy and model size utilizing (a) Bi-LSTM, (b) LSTM and Morlet wavelet on subject 7 of BCI Competition IV 2a with respect to different step sizes of proposed wavelet fractal\u003c/p\u003e \u003cp\u003eNext, the performance and model sizes of the Bi-LSTM and LSTM architectures were compared for different step sizes of wavelet fractals, as illustrated in Fig.\u0026nbsp;3. Figure\u0026nbsp;3(a) and Fig.\u0026nbsp;3(b) depict the testing accuracy, validation accuracy, and average accuracy (computed from testing and validation) for each step size of wavelet fractals, along with the corresponding model sizes of Bi-LSTM and LSTM, respectively. The best average accuracy of Bi-LSTM and LSTM architecture was obtained on step size of 10 and 14 respectively as shown in Fig.\u0026nbsp;3. The model sizes of Bi-LSTM and LSTM architecture were 9.87 and 12.07 MB respectively. The best performance was obtained by Bi-LSTM architecture with better model size than LSTM architecture. Further, the Bi-LSTM architecture with step size of 10 wavelet fractal was selected for analysis on BCI Competition IV 2a and 2b.\u003c/p\u003e \u003cp\u003eTABLE II\u003c/p\u003e \u003cp\u003ePerformance metrics on BCI Competition IV 2b and 2a for intra-subject classification\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"13\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c13\" colnum=\"13\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eDatasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003ePerformance Metrics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"9\" nameend=\"c11\" namest=\"c3\"\u003e \u003cp\u003eAccuracy of Subjects (S)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMean\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e10-fold\u003c/p\u003e \u003cp\u003emean\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eS1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eS2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eS3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eS4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eS5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eS6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eS7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eS8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003eS9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2b\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e91.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e83.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e97.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e\u003cb\u003e87.75\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e77.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean Precision\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e92.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e92.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e82.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e97.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e84.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e88.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e78.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean Recall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e71.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e91.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e91.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e82.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e97.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e87.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e77.31\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean F1-score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e71.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e91.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e91.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e82.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e97.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e84.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e88.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e76.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eCohen\u0026rsquo;s Kappa\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2a\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e89.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e79.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e96.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e89.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e58.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e65.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e93.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e96.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e\u003cb\u003e85.45\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e76.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean Precision\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e87.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e84.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e96.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e88.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e66.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e79.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e93.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e97.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e88.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e78.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean Recall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e86.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e92.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e89.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e56.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e63.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e93.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e98.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e84.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e76.76\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMean F1-score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e86.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e79.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e88.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e59.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e62.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e92.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e97.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e84.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e75.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eCohen\u0026rsquo;s Kappa\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c13\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFurthermore, the proposed algorithm evaluated with 10-fold cross validation technique to check for reliability and robustness of model and methodology by calculating accuracy, precision, recall, f1-score and Cohen\u0026rsquo;s kappa for each dataset as shown in Table II for intra-subject and Table III for inter-subject. The obtained performance metrics showed that the proposed methodology was reliable and robust.\u003c/p\u003e \u003cp\u003eTABLE III\u003c/p\u003e \u003cp\u003ePerformance metrics on BCI Competition IV 2a and 2b for inter-subject classification\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eDatasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003en-fold mean\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c7\" namest=\"c3\"\u003e \u003cp\u003ePerformance metrics\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean precision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMean recall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMean f1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCohen\u0026rsquo;s Kappa\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2b\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1-fold\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e76.16\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e76.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e76.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e10-fold\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e74.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e74.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e73.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e71.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2a\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1-fold\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e71.76\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e71.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e70.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e70.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e10-fold\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFinally, we compared our results with earlier proposed state-of-the-art algorithms, the proposed algorithm achieved better classification accuracy through cross correlation matrix based on mean wavelet-fractal utilizing CWT coefficients at different scale followed by Bi-LSTM based algorithm that evaluates the evolution in CCMs and performs classification. The comparison of performance with the state-of-the-art algorithms and the proposed algorithm has been compiled in Table IV and Table V for intra-subject and inter-subject classification respectively. In the current work, the proposed methodology is lightweight, reliable, robust, and practically applicable to real-time scenarios for MI EEG classification. It has implications on embedded based device implementation for MI classification. The obtained results were optimum with CCMs based on mean wavelet-fractal utilizing Morlet. Besides it, the step scaled mean wavelet-fractals obtained were also unique as compared to each other as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eTABLE IV\u003c/p\u003e \u003cp\u003ePerformance metrics on BCI Competition IV 2b and 2a for intra-subject classification\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAuthors and year\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMethodology\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMean accuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eImprovement in accuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2b\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eH.K. Lee et.al (2019) [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCWT (Morlet)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCWT (Mexican hat)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6.55\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eX. Tang et.al (2020) [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eConditional Empirical Mode Decomposition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1D Multi-scale CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhang et. al (2021) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG inception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eJia et. al (2022) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTemporal-spatial filtering and convolution\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.25\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhi et. al (2023) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti-domain convolution\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e86.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.36\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhao et. al (2024) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCTNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eOurs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eCDGSBFCSP\u0026thinsp;+\u0026thinsp;CWT\u0026thinsp;+\u0026thinsp;proposed wavelet fractals\u0026thinsp;+\u0026thinsp;CCMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eBi-LSTM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e87.75\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eMean\u0026thinsp;=\u0026thinsp;2.28\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"6\" rowspan=\"7\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2a\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhang et. al (2021) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG Inception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-2.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eJia et. al (2022) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTemporal-spatial filtering and convolution\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.45\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSalami et. al (2022) [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG-ITNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e76.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8.71\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhi et. al (2023) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti-domain convolution\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.73\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAltaheri et. al (2023) [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSliding window\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eATCNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e85.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.07\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhao et. al (2024) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCTNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eOurs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eCDGSBFCSP\u0026thinsp;+\u0026thinsp;CWT\u0026thinsp;+\u0026thinsp;proposed wavelet fractals\u0026thinsp;+\u0026thinsp;CCMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eBi-LSTM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e85.45\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eMean\u0026thinsp;=\u0026thinsp;2.33\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn future research, we aim to further enhance the proposed methodology by addressing key challenges in multi-class motor imagery (MI) EEG classification. One of the primary directions will involve developing an attention-based framework to dynamically analyse and capture the evolution of discriminative features across multiple wavelet fractals simultaneously. This approach is expected to improve the identification of class-specific patterns, leading to more robust feature extraction and classification performance. Further, we plan to extend this methodology for implementation in a real-time brain-computer interface (BCI) system aimed at addressing the complexities involved in developing practical BCI devices. Specifically, the proposed approach will be optimized for commercialization in prosthetic limb control for individuals with paralysis.\u003c/p\u003e \u003cp\u003eTABLE V\u003c/p\u003e \u003cp\u003eComparison with earlier studies for inter-subject classification\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabe\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAuthors and year\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMethodology\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eImprovement in accuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2b\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHermosilla et.al (2021) [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTemporal and spatial features\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eShallowConvNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e75.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhang et. al (2021) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG inception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-1.28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMammone et.al (2023) [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAutoencoder based FBCSP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e74.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhao et. al (2024) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCTNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e76.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eOurs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eCDGSBFCSP\u0026thinsp;+\u0026thinsp;CWT\u0026thinsp;+\u0026thinsp;proposed wavelet fractals\u0026thinsp;+\u0026thinsp;CCMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eBi-LSTM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e76.16\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eMean\u0026thinsp;=\u0026thinsp;0.22\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003e\u003cb\u003eBCI Competition IV 2a\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhang et. al (2021) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG inception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e65.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSalami et. al (2022) [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFine tuning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEEG-ITNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e78.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-6.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAltaheri et. al (2023) [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSliding window\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eATCNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e70.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZhao et. al (2024) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCTNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e58.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e13.12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eOurs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eCDGSBFCSP\u0026thinsp;+\u0026thinsp;CWT\u0026thinsp;+\u0026thinsp;proposed wavelet fractals\u0026thinsp;+\u0026thinsp;CCMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eBi-LSTM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e71.76\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eMean\u0026thinsp;=\u0026thinsp;3.21\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eIn this study, we proposed a novel algorithm for multi-class motor imagery (MI) EEG classification using a lightweight, customized Bi-LSTM network. The methodology integrated CDGSBFCSP, continuous wavelet transform (CWT), wavelet fractal generation, and cross-correlation matrices, offering a practical and effective approach for MI EEG classification. The evolution of CCMs across all channels, based on mean wavelet-fractal features, was analysed and utilized to enhance classification performance. For intra-subject classification, the proposed algorithm achieved best accuracies of 85.45% and 87.75%, demonstrating improvements of 2.33% and 2.28% compared to earlier state-of-the-art methods on the BCI Competition IV 2a and 2b datasets, respectively. Similarly, for inter-subject classification, the best accuracies achieved were 71.76% and 76.16%, with improvements of 3.21% and 0.22% over prior methods on the same datasets respectively. The customized Bi-LSTM architecture exhibited a compact model size of 9.87 MB for BCI Competition IV 2a and 1.11 MB for BCI Competition IV 2b, highlighting its suitability for embedded device implementation in real-time MI EEG classification tasks. In the future, we plan to develop an attention-based approach to capture the evolution of wavelet-fractal features more effectively and to design a brain-computer interface (BCI) system suitable for real-time applications.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eConflict of Interest\u003c/strong\u003e \u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eEthical approval\u003c/h2\u003e \u003cp\u003eNo ethical approval is required as there was no involvement of human /animal participation in this study.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe Science for Equity Empowerment and Development Division, Department of Science and Technology, govt. of India [SEED/TIDE/2019/340], has provided funds for the necessary systems and support for completing this implementation.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors contributed to the study's conception and design. Balendra, Neeraj Sharma, and Shiru Sharma performed implementation and analysis. Balendra wrote the manuscript's first draft, and all authors commented on previous versions. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e \u003cp\u003eAll authors would like to thank the Science for Equity Empowerment and Development Division (SEED), Department of Science and Technology (DST), Government of India, for providing funds for the necessary systems and support for completing this implementation. The authors are also very obliged to the Prime Minister Research Fellowship (PMRF) for providing financial motivation to work on this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWolpaw, J.R., 2013. Brain\u0026ndash;computer interfaces. In \u003cem\u003eHandbook of clinical neurology\u003c/em\u003e (Vol. 110, pp. 67\u0026ndash;74).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLebedev, M.A. and Nicolelis, M.A., 2006. Brain\u0026ndash;machine interfaces: past, present and future. TRENDS in Neurosciences, \u003cem\u003e29\u003c/em\u003e(9), pp.536\u0026ndash;546.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHochberg, L.R., Bacher, D., Jarosiewicz, B., Masse, N.Y., Simeral, J.D., Vogel, J., Haddadin, S., Liu, J., Cash, S.S., Van Der Smagt, P. and Donoghue, J.P., 2012. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, \u003cem\u003e485\u003c/em\u003e(7398), pp.372\u0026ndash;375.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicolelis, M.A., 2003. Brain\u0026ndash;machine interfaces to restore motor function and probe neural circuits. Nature Reviews Neuroscience, \u003cem\u003e4\u003c/em\u003e(5), pp.417\u0026ndash;422.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePfurtscheller, G. and Neuper, C., 1997. Motor imagery activates primary sensorimotor area in humans. Neuroscience letters, \u003cem\u003e239\u003c/em\u003e(2\u0026ndash;3), pp.65\u0026ndash;68.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeannerod, M., 1994. The representing brain: Neural correlates of motor intention and imagery. Behavioral and Brain sciences, \u003cem\u003e17\u003c/em\u003e(2), pp.187\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDaly, J.J. and Wolpaw, J.R., 2008. Brain\u0026ndash;computer interfaces in neurological rehabilitation. The Lancet Neurology, \u003cem\u003e7\u003c/em\u003e(11), pp.1032\u0026ndash;1043.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCharte, D., Charte, F., Garc\u0026iacute;a, S., del Jesus, M.J. and Herrera, F., 2018. A practical tutorial on autoencoders for nonlinear feature fusion: Taxonomy, models, software and guidelines. Information Fusion, \u003cem\u003e44\u003c/em\u003e, pp.78\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuger, C., Edlinger, G., Harkam, W., Niedermayer, I. and Pfurtscheller, G., 2003. How many people are able to operate an EEG-based brain-computer interface (BCI)?. IEEE transactions on neural systems and rehabilitation engineering, \u003cem\u003e11\u003c/em\u003e(2), pp.145\u0026ndash;147.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBashashati, A., Fatourechi, M., Ward, R.K. and Birch, G.E., 2007. A survey of signal processing algorithms in brain\u0026ndash;computer interfaces based on electrical brain signals. \u003cem\u003eJournal of Neural engineering\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(2), p.R32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKantz, H. and Schreiber, T., 2003. \u003cem\u003eNonlinear time series analysis\u003c/em\u003e. Cambridge university press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWronkiewicz, M., Larson, E. and Lee, A.K., 2016. Incorporating modern neuroscience findings to improve brain\u0026ndash;computer interfaces: tracking auditory attention. \u003cem\u003eJournal of neural engineering\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(5), p.056017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchirrmeister, R.T., Springenberg, J.T., Fiederer, L.D.J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W. and Ball, T., 2017. Deep learning with convolutional neural networks for EEG decoding and visualization. Human brain mapping, \u003cem\u003e38\u003c/em\u003e(11), pp.5391\u0026ndash;5420.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShe, Q., Hu, B., Luo, Z., Nguyen, T. and Zhang, Y., 2019. A hierarchical semi-supervised extreme learning machine method for EEG recognition. Medical \u0026amp; biological engineering \u0026amp; computing, \u003cem\u003e57\u003c/em\u003e, pp.147\u0026ndash;157.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLe\u0026oacute;n, J., Escobar, J.J., Ortiz, A., Ortega, J., Gonz\u0026aacute;lez, J., Mart\u0026iacute;n-Smith, P., Gan, J.Q. and Damas, M., 2020. Deep learning for EEG-based Motor Imagery classification: Accuracy-cost trade-off. Plos one, \u003cem\u003e15\u003c/em\u003e(6), p\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e.e\u003c/span\u003e\u003cspan address=\"http://.e\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e0234178.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLotte, F., Congedo, M., L\u0026eacute;cuyer, A., Lamarche, F. and Arnaldi, B., 2007. A review of classification algorithms for EEG-based brain\u0026ndash;computer interfaces. Journal of neural engineering, \u003cem\u003e4\u003c/em\u003e(2), p.R1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, C., Kim, Y.K. and Eskandarian, A., 2021. EEG-inception: an accurate and robust end-to-end neural network for EEG-based motor imagery classification. \u003cem\u003eJournal of Neural Engineering\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(4), p.046014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAltaheri, H., Muhammad, G. and Alsulaiman, M., 2022. Physics-informed attention temporal convolutional network for EEG-based motor imagery classification. IEEE transactions on industrial informatics, \u003cem\u003e19\u003c/em\u003e(2), pp.2249\u0026ndash;2258.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalami, A., Andreu-Perez, J. and Gillmeister, H., 2022. EEG-ITNet: An explainable inception temporal convolutional network for motor imagery classification. IEEE Access, \u003cem\u003e10\u003c/em\u003e, pp.36672\u0026ndash;36685.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, W., Jiang, X., Zhang, B., Xiao, S. and Weng, S., 2024. CTNet: a convolutional transformer network for EEG-based motor imagery classification. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(1), p.20237.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH. Zhi, Z. Yu, T. Yu, Z. Gu and J. Yang, \"A Multi-Domain Convolutional Neural Network for EEG-Based Motor Imagery Decoding,\" in IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 3988\u0026ndash;3998, 2023\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJia, X., Song, Y., Yang, L. and Xie, L., 2022. Joint spatial and temporal features extraction for multi-classification of motor imagery EEG. \u003cem\u003eBiomedical Signal Processing and Control\u003c/em\u003e, \u003cem\u003e71\u003c/em\u003e, p.103247.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePearson, K., 1920. Notes on the history of correlation. Biometrika, \u003cem\u003e13\u003c/em\u003e(1), pp.25\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDaubechies, I., 1992. Ten lectures on wavelets. \u003cem\u003eSociety for industrial and applied mathematics\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHochreiter, S., 1997. Long Short-term Memory. \u003cem\u003eNeural Computation MIT-Press\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeeb, R., Brunner, C., M\u0026uuml;ller-Putz, G., Schl\u0026ouml;gl, A. and Pfurtscheller, G.J.G.U.O.T., 2008. BCI Competition 2008\u0026ndash;Graz data set B. Graz University of Technology, Austria, \u003cem\u003e16\u003c/em\u003e, pp.1\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrunner, C., Leeb, R., M\u0026uuml;ller-Putz, G., Schl\u0026ouml;gl, A. and Pfurtscheller, G., 2008. BCI Competition 2008\u0026ndash;Graz data set A. \u003cem\u003eInstitute for knowledge discovery (laboratory of brain-computer interfaces), Graz University of Technology\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e, pp.1\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHossin, M. and Sulaiman, M.N., 2015. A review on evaluation metrics for data classification evaluations. \u003cem\u003eInternational journal of data mining \u0026amp; knowledge management process\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(2), p.1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVieira, S.M., Kaymak, U. and Sousa, J.M., 2010, July. Cohen's kappa coefficient as a performance measure for feature selection. In \u003cem\u003eInternational conference on fuzzy systems\u003c/em\u003e (pp. 1\u0026ndash;8). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJain, A., Nandakumar, K. and Ross, A., 2005. Score normalization in multimodal biometric systems. Pattern recognition, \u003cem\u003e38\u003c/em\u003e(12), pp.2270\u0026ndash;2285.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo, J., Wang, J., Xu, R. and Xu, K., 2019. Class discrepancy-guided sub-band filter-based common spatial pattern for motor imagery classification. Journal of neuroscience methods, \u003cem\u003e323\u003c/em\u003e, pp.98\u0026ndash;107.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStrang, G., 1994. Wavelets. American Scientist, \u003cem\u003e82\u003c/em\u003e(3), pp.250\u0026ndash;255.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eB\u0026uuml;ssow, R., 2007. An algorithm for the continuous Morlet wavelet transform. Mechanical Systems and Signal Processing, \u003cem\u003e21\u003c/em\u003e(8), pp.2970\u0026ndash;2979.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTorrence, C. and Compo, G.P., 1998. A practical guide to wavelet analysis. Bulletin of the American Meteorological society, \u003cem\u003e79\u003c/em\u003e(1), pp.61\u0026ndash;78.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, H.K. and Choi, Y.S., 2019. Application of continuous wavelet transform and convolutional neural network in decoding motor imagery brain-computer interface. \u003cem\u003eEntropy\u003c/em\u003e, \u003cem\u003e21\u003c/em\u003e(12), p.1199.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang, X., Li, W., Li, X., Ma, W. and Dang, X., 2020. Motor imagery EEG recognition based on conditional optimization empirical mode decomposition and multi-scale convolutional neural network. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e149\u003c/em\u003e, p.113285.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHermosilla, D.M., Codorni\u0026uacute;, R.T., Baracaldo, R.L., Zamora, R.S., Rodriguez, D.D., Albuerne, Y.L. and \u0026Aacute;lvarez, J.R.N., 2021. Shallow convolutional network excel for classifying motor imagery EEG in BCI applications. IEEE Access, \u003cem\u003e9\u003c/em\u003e, pp.98275\u0026ndash;98286.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMammone, N., Ieracitano, C., Adeli, H. and Morabito, F.C., 2023. AutoEncoder filter bank common spatial patterns to decode motor imagery from EEG. IEEE journal of biomedical and health informatics, \u003cem\u003e27\u003c/em\u003e(5), pp.2365\u0026ndash;2376.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Electroencephalogram (EEG), Motor Imagery, Continuous Wavelet Transform, Cross correlation matrix, Bidirectional long short term memory","lastPublishedDoi":"10.21203/rs.3.rs-5750495/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5750495/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBrain-Computer Interface (BCI) represents a cutting-edge area of research that integrates neuroscience, engineering, and computer science to establish direct communication links between the human brain and external systems. BCI-based technologies hold significant promise, particularly in the development of prosthetic devices. However, the practical application of BCI in real-time scenarios faces several obstacles, including bulky models, noise interference, artifacts, and the complexity of motor imagery (MI) electroencephalogram (EEG) data, which exhibits both inter-subject and intra-subject variability. To address these challenges, the proposed algorithm utilizes loading and pre-processing of MI EEG, extracts their common spatial patterns, calculates the continuous wavelet transform (CWT) coefficients, computes their proposed step scaled mean wavelet fractals which exhibits robustness towards inherent noises and artifacts, calculates the cross-correlation matrix at different scale for all channels and observes the evolution in cross-correlation matrices with the help of customized Bi-long-short term memory (Bi-LSTM) neural network to classify MI EEG. The customized Bi-LSTM architecture had the size\u0026thinsp;\u0026lt;\u0026thinsp;10MB showing the effectiveness of methodology for MI EEG classification utilizing embedded based devices. The best classification accuracies achieved with proposed step scaled mean wavelet fractals were 87.75% and 85.45% for intra-subject as well as 76.16% and 71.76% for inter-subject on BCI Competition IV 2b and 2a respectively; the comparative analysis with earlier state of the art methods showed an average improvement of 2.28% and 2.33% for intra-subject as well as 0.22% and 3.21% for inter-subject in accuracy.\u003c/p\u003e","manuscriptTitle":"A lightweight methodology for Motor Imagery EEG classification utilizing step scaled wavelet fractals and Bi-LSTM architecture","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-01-09 12:02:02","doi":"10.21203/rs.3.rs-5750495/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5b6414c8-7900-4885-af4a-755d9e215de2","owner":[],"postedDate":"January 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-01-09T12:02:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-01-09 12:02:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5750495","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5750495","identity":"rs-5750495","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}