Decoding of imagined speech electroencephalography neural signals using transfer learning method

The use of brain-computer interfaces to produce imagined speech from brain waves has the potential to assist individuals with difficulty producing speech or communicating silently. The decoding of covert speech has been observed to have limited efficacy due to the diverse nature of the associated measured brain waves and the limited number of covert speech databases. As a result, traditional machine learning algorithms for learning and inference are challenging, and one of the real alternatives could be to leverage transfer of learning. The main goals of this research were to create a new deep learning (DL) framework for decoding imagined speech electroencephalography (EEG) signals tasks using transfer learning and to transfer the model learning of the source task of an imagined speech EEG dataset to the model training on the target task of another imagined speech EEG dataset, essentially the cross-task learning transfer of discriminative characteristics of the source task to the target task of imagined speech. The experiment was carried out using two distinct open-access EEG datasets, FEIS and KaraOne, that recorded the imagined speech classes of neural signals from multiple individuals. The target FEIS model and the target KaraOne model for multiclass classification exhibit overall accuracy of 89.01% and 82.35%, respectively, according to the proposed transfer learning. The experiment results indicate that the cross-task deep transfer learning design reliably classifies the imagined speech EEG signals by applying the source task learning to the target task learning. The findings suggest the feasibility of a consistent strategy for classifying multiclass imagined speech with transfer learning, which could thereby open up the possibility of future investigation into cross-task imagined speech classification knowledge usability for generalization of new imagined speech prompts.


Introduction
A brain-computer interface (BCI) application is a type of human-computer interface based on neural activity in the brain.A BCI application allows the individual to participate in communication in a different way by decoding the brain signals that represent the imagined speech activity [1][2][3][4].This study considers the BCI based on noninvasive electroencephalography (EEG) sensors to record brain activity because of the advantages of being portable, affordable, and having good temporal resolution.A psychological activity in which a participant imagines speech prompts such as syllables, phonemes, words, and etc without moving any articulators is denoted as imagined speech.Though research on alternative BCI paradigms, such as steady-state visual evoked potential, motor imagery, event-related potential, and others is advancing, imagined speech has emerged as one of the most common BCI models since verbal communication via speech is the most preferred means of communication for individuals.
In earlier studies, Fujimaki et al [5] investigated the evoked potential of the imagined speech prompt vowel using EEG signals.Another experiment by Porbadnigk et al [6] used the Hidden Markov machine learning learning classifier was applied to the five words and three vowels of imagined speech.The majority of recent research has been focused on the decoding of imagined speech prompts employing cross-subject or crosssession based knowledge sharing.The studies lack an investigation of the cross-task transfer of training from one task of decoding imagined speech to another task of decoding imagined speech with a different set of speech prompts using the same or different EEG devices, or inter-dataset knowledge transfer.
This work intends to examine the cross-domain and cross-task classifier transfer learning approach for generalizing and applying the learnt model knowledge from the source imagined speech task multiclass classification to the target model for the imagined speech task multiclass classification with efficient performance.This studyʼs findings might contribute to future research towards deep transfer learning in the EEG-based imagined speech BCI system extension.

Methods
The cross-domain as well as cross-task transfer learning framework is depicted in figure 1, where the source domain pre-trained ImageNet model learning is transferred to the cross-domain target model of the set of imagined speech classification tasks, and the learned imagined speech classification model knowledge is transferred to the cross-task learning target model of the distinct set of imagined speech classification tasks.

Mathematical depiction
In supervised learning, a record is represented by the tuple (E i , p i ), where E i stands for the feature vector of the imagined speech signal and p i for the speech class.The EEG signal is represented as a two-dimensional vector E i ä R n× m , where R represents the rational number, n represents the number of channels, and m represents the number of time domain sampling points.The learning functions for the imagined speech classification can be defined as in equation (1).
The following definitions provide instances of two transfer learning scenarios for the source domain and target domain: • In the cross-domain transfer learning, the source domain is denoted as D Img y ,  objective is to learn the classifier function  E p : i t i t ⟶ for the target context C t from the source context C s learning function.
• Zhao et al [9] from the Toronto Rehabilitation Institute published the KaraOne open-access database used in the research.The EEG imagined speech from the 14 subjects in the KaraOne database were acquired by placing electrodes on the scalp of the head using the 10-20 international standard [33] and utilizing the 64channel Neuroscan Qick-Cap device.The properties of the EEG signal set are summarized in table 2. Additionally, the KaraOne dataset is uniformly distributed among the various speech prompts.

EEG signal processing
The KaraOne signals were downsampled from 1000 Hz to 256 Hz to maintain the KaraOne signal sampling rate consistent with the FEIS signal sampling rate.The KaraOne and FEIS signals were both passed through a bandpass filter (a fifth-order Butterworth infinite impulse response filter) with a frequency range of 0.01 Hz to 120 Hz, retaining all five EEG frequency bands.Furthermore, the signals from both the databases were notch filtered at a 60 Hz cutoff frequency to remove electrical line noise.
As the KaraOne EEG signals were obtained with 64 channels for the synchronization with the FEIS EEG signals of 14 channels for the cross-task knowledge transfer, the identical FEIS 14 channels were selected from the KaraOne signals.
The signal-to-noise ratio (SNR) of EEG signals is low primarily because there is lots of neural background activity that is probably unrelated to speech imagery and can be considered noise, which causes recorded electrical activity to be contaminated with artifacts and makes it difficult to extract information from the EEG signal.As a result, artifacts in the EEG signal need to be eliminated.According to studies, the ICA technique is one of the preferable approaches for removing artifacts from EEG signals [34][35][36].Consequently, a widely used method, FastICA (fast independent component analysis), was employed to automatically remove artifacts from the KaraOne and FEIS signals.
In this study, DWT is employed for the non-stationary time-frequency analysis of EEG signals, and the deep transfer learning model further leverages the use of the wavelet coefficient characteristics of the signals.The DWT approach was used in various studies to obtain time-frequency analysis features from the non-stationary EEG signals [4,37,38].The wavelet transform function is defined in equation (2) as F WT for the EEG signal E(n).The DWT approach decomposes the signal into a number of mutually orthogonal wavelets.
where the wavelet function is denoted by ψ(n) and the scale and shift factors are a and b, respectively.The EEG signals were represented in nonlinear time-frequency domains to capture signal variation by DWT in terms of approximation coefficients (AC) and detailed coefficients (DC).The first level DWT decomposes a signal into approximation coefficients (AC 1 ) and detail coefficients (DC 1 ), and further, at each subsequent level, the approximation coefficients (AC n ) are further divided into detail coefficients (DC n+1 ) and approximation coefficients (AC n+1 ).The order 4 Daubechies (db4) wavelet was used in the experiment because its smoothing property made it more suited to detecting changes in EEG signals [39] and also because research has suggested that the db4 wavelet basis is useful in imagined speech [18,19,40].The DWT was employed at level 7 of signal decomposition using the db4 mother wavelet.The obtained wavelet coefficients (DC 1 : 128-256Hz, DC 2 : 64-127 Hz, DC 3 : 32-63 Hz, DC 4 : 16-32 Hz, DC 5 : 8-15 Hz, DC 6 : 4-8 Hz, DC 7 : 0-1 Hz, and AC 7 : 2-3 Hz) were used to construct the signal feature matrix.The EEG signal feature matrix was normalised for computationally efficient deep learning model training using the z-score technique, as shown in equation (3).
where σ denotes the mean, μ the standard deviation, and E i the signal for the record i. Figure 2 highlights the various signal processing approaches used to prepare the KaraOne and FEIS data sets for feeding into deep transfer learning models.

Model frameworks
The  transferred layers, a small number of fully connected (FC) or dense layers were added for the task-I imagined speech specfic feature information learning.The rectified linear unit (ReLU) [41] activation function was used in the FC layers, as defined in equation (4).Finally, after the FC layer, the output layer was added to claasify the imagined speech signal.

ReLU y y
where y is the output of the neuron that was fed into the ReLU function.
Similarly, the cross-task deep transfer learning method for task-II was to retain the convolution and dense layers while excluding the output layer of the learned task-I imagined speech classification model.In addition to the transferred layers, a small number of FC layers with activation function ReLU were added for the imagined speech task-II special information extraction.The output layer was then added to claasify the imagined speech signal after the FC layer.
Considering the source domain image input size being different compared to the target domain EEG signal set input size, the input dimension of the target domain imagined speech signal sets was resized to match the input dimension of the pre-trained model architecture.
The target domain or context modelʼs framework, which consists of transferred layers from the source model, the new dense layers, and an output layer for the classification, reused the obtained layer weights in the training by freezing transferred layers, and the trainable layers were fine-tuned and trained on target imagined speech EEG signal sets.
The modelʼs output layer has a number of neurons, each of which corresponds to a single imagined speech class, as the target imagined speech classes were one-hot encoded.The activation function is a softmax activation function as defined in equation (5), which generates a probability distribution for each neuron and interprets the highest probability as the classified imagined speech class as defined in equation (6).And the cross entropy loss function, which is described in equation (7), was employed in the multiclass model learning process.

Pre-trained imagenet model
Several tests were performed with the source pre-trained well-known deep convolutional network models used in ImageNet classification that have great performance, such as ResNet152V2 [42] and DenseNet-201 [43], to evaluate the effective convolutional-based transfer learning model with the best performance.The DenseNet2001 (dense convolutional network) architecture employs dense connections between layers through the use of dense blocks and is based on the idea of residual learning by performing a depth concatenation of the current layerʼs output and the outputs of all the preceding layers so that the combined prior feature maps produced by the preceding layers are included in the succeeding layer.The mathematical representation of the feature map concatenation is defined as in equation (8).
x H x x x , , , , 8 where x j represents the output of layer j.The nonlinear transformation H j () for layer j is defined as a composite function that combines batch normalization, activation function (ReLU), and convolution.The feature map concatenation or dense connection from layer 0 to layer j − 1 is denoted by x x x , , , j The current ResNet152V2 (residual networks) convolutional network architecture transmits features leveraging identity mapping to reduce gradient dispersion, and several residual blocks are stacked up to build the network structure.The residual block is made up of connections that skip one or more levels in order to transfer the feature data as indicated in equation (9).
x H x x , 9 where x j and x j−1 represent the output of layers j and j − 1, respectively.Hj() denotes the trainable nonlinear residual function for layer j, which is mainly composed of convolution, normalization, and activation layers.The pre-trained ResNet152V2 and DenseNet201 models' architecture input shape tuple has a 3-dimensional shape of (width: 224, height: 224, channel: 3), and the input must be an accurate 3-dimensional tuple with the restriction that width and height cannot be less than 32.Hence, in this experiment, the transfer learning framework integration required resizing the EEG signal feature matrix of both the datasets, KaraOne and FEIS, to align with these pre-trained models.The feature matrix is resampled to the higher sampling rate with a dimension of (channel: 14 x sampling points: 1920) from the original (channel: 14 x sampling points: 1280).In order to maintain the feature matrix aligned with the pre-trained receptive field (width, height, and channel), the sample points of each channel were then resampled into three distinct parts, and the resulting feature matrix tuple dimension is (14, 640, 3).Further, each of the three split parts of the feature matrix, which has windows of a two-dimensional size (channel: 14, sampling point: 640), was reshaped into a one-dimensional array of a size of 8960.The one-dimensional array was then randomly rearranged with regard to the dimensions into twodimensional windows with the restriction of a minimum width and minimum height of 32.Finally, the resultant feature matrix is aligned with the pre-trained modelʼs input.Figure 4 depicts several steps involved in aligning the input signal with the input form of pretrained transfer learning models.In the case of the FEIS, the tuple shape (width: 80, height: 112, channel: 3) and KaraOneʼs tuple shape (width: 280, height: 32, channel: 3) are selected with the best classification performance.

Model training and optimization
The tuning of the hyperparameters is one of the most critical elements influencing the efficiency of a deep learning model, hence the accuracy of the imagined speech EEG signal decoding.The CNN-based transfer learning model has training parameters (such as batch size and learning rates) and the number of trainable layers (fully connected layers) that directly affect performance.In this experiment, the random search algorithm is used for the hyperparameter selection.The Adam optimization [44] algorithm with cross-entropy loss function was employed for the model training to discover optimal hyperparameters while keeping focus on the accuracy of the outcome of the model.More precisely, the learning rate is configured to search in the discrete search space list [0.0001, 0.0005, 0.001, 0.005, 0.01], the batch size is configured to search a discrete space of [64,128,192,256], and the number of FC layers is configured to search in the discrete space of (1, 2, 3, 4).The model training and validation were carried out with respect to the selected hyperparameters using the training and validation imagined speech feature matrix sets, and the best combination of hyperparameters was sorted with the highest overall accuracy.Furthermore, the classifierʼs performance on the test imagined speech feature matrix sets was employed as an impartial measure of its efficacy on left-over imagined speech.Table 3 summarises the model framework parameters as well as the training parameters.

Evaluation of experiments
In order to assess the proposed framework, two experiments were conducted with exchanging the imagined speech data sets for the cross-task knowledge transfer.In the experiment-1 the pre-trained source model was ImageNet, which was transferred to the cross-domain target model task-I to classify the imagined speech signal set FEIS, and the task-I model learning was transferred to the cross-task target model task-II to classify the imagined speech signal set KaraOne.In the contrast experiment-2, task-I was to classify the imagined speech signal set KaraOne, whereas task-II was to classify the imagined speech signal set FEIS.In essence, in the two experiments, FEIS and KaraOne imagined speech signal sets exchanged between task-I and task-II.
The evaluation matrices for models used to evaluate the performance of transfer learning models on target signal sets for imagined speech prompts classification were overall accuracy, precision (confidence), recall (sensitivity), and f1-score (harmonic mean of recall and precision), as defined in equations (10), (11), (12), and (13), respectively.

accuracy TP TN TP TN FN FP
, 1 0 where TN, TP, FN, and FP stand for true negatives, true positives, and false negatives, respectively.The experiment was designed to be population-based, with imagined speech from all subjects compiled before deep transfer learning model training, validation, and performance evaluation.The imagined speech signal database has been separated into training (80%), validation (10%), and testing (10%) sets, with training and validation signal sets employed in model training and model hyperparameter fine tuning.Furthermore, the unused test signal set was used to evaluate the performance of the proposed frameworks.

Results
Table 4 shows the findings of experiment-1, the cross-task transfer learning to the target imagined speech task of the KaraOne signal set, presentated the multiple pretrained ImageNet modelʼs validation overall accuracy outcomes in percentage.
The results of experiment-2, which involved cross-task transfer learning to the target imagined speech task of the FEIS signal set, are displayed in table 5, with the validation classification results for the overall accuracy of the multiple pretrained ImageNet model noted in percentage.
On the basis of the Densenet201 pre-trained model, the better overall accuracy demonstration of cross-task transfer learning from the source model FEIS task-I to the target model KaraOne task-II in experiment-1 was accomplished.The confusion matrix for the results of the successful Densenet201-based experiment-1 (target task-II KaraOne) model validation is shown in figure 5. Recall, precision, and f1-score metrics from the Densenet201-based experiment-1 (target task-II KaraOne) model validation are presented in table 6.
Cross-task transfer learning from the source model KaraOne task-I to the target model FEIS task-II had overall better accuracy in experiment-2 based on the ResNet152V2 pre-trained model.The confusion matrix for the model validation outcomes for experiment-2 (target task-II FEIS) based on ResNet152V2 is shown in figure 6.Table 7 displays the model validation results for the ResNet152V2-based experiment-2 (target task-II FEIS), including recall, precision, and f1-score metrics.

Transfer learning and knowledge transferability
The well-established ImageNet CNN types of models, such as DenseNet201 and ResNet152V2, include millions of parameters.Millions of data records would be required to ensure that the network could discriminate features with all the model parameters learned from the beginning.Moreover, due to the constraints of specific criteria for the experimental configuration and accessible participants, it really is challenging to acquire large EEG datasets in BCI applications.As a result, utilization of the knowledge of the pre-trained model with the knowledge type in the transfer learning framework was the model parameters that transferred from the source domain or context to the target domain or context.In this experiment, DenseNet201 has 18,321,984 parameters frozen or non-trainable, whereas ResNet152V2 has 58,331,648 parameters used.Furthermore, the lower layers of the deep neural network model learn how to represent general feature patterns, whereas the higher layer learns individual feature representations relevant to a specific imagined speech class.The pre-trained ImageNet parameters were transferred to the cross-domain imagined speech task-I model to extract universal features, and the fully connected layers and output layer parameters were trained from scratch using task-I imagined speech signal sets.Subsequently, the learned parameters of the imagined speech task-I model were transferred to the imagined speech task-II model to extract generic features, and the fully connected layers and output layer parameters were trained from scratch using task-II imagined speech.As indicated in tables 4 and 5, the proposed transfer learning framework multiclass classification result of the 11 KaraOne imagined speech overall accuracy is 82.35% for DenseNet201, and the multiclass classification result of the 16 FEIS imagined speech overall accuracy is 89.01%for ResNet152V2.According to the preceding, the transfer learning model can learn the discriminative feature selection complexity of EEG signals.
As indicated in table 5, the multiclass classification result for the target FEIS imagined speech overall accuracy for DenseNet201 was 79.79%, with performance degrading.The pre-trained model DenseNet201ʼs knowledge of the source domain has a negative direct or indirect impact on the target imagined speech task-II.This negative impact could be induced by differences between the source data and target data or by transferable components or parameters that do not function effectively during the transfer.As revealed in previous research, this degradation of performance could be mostly the impact of the negative transfer learning of the model [45].

Comparison with the state-of-the-art
As shown in tables 4 and 5, the proposed cross-task transfer learning model achieves competitive performance accuracy of 82.35% for the target KaraOne model and 89.01%for the target FEIS model for multiclass classification, in comparison to prior studies with other transfer learning models such as Cooney et al [28] crosssubject learning of accuracy of 35.68%, Vorontsova et al [29] cross-domain learning of accuracy of 84.5%, Tamm et al [30] cross-subject learning of accuracy of 24.77%, and Panachakel et al [31] cross-domain learning of accuracy of 79.7%-95.5%.The results show that transfer learning procedures are significant for the generalizability of decoding imagined speech EEG signals, despite the difficulties of comparison across state-ofthe-art transfer learning research where the investigations were with distinct datasets.
Furthermore, in order to assess the efficacy of the proposed cross-domain and cross-task transfer learning approach, the results of the proposed model are compared to the most significant results of the state-of-the-art imagined speech recognition techniques using the KaraOne and FEIS databases.The proposed methodology differs from earlier research with the use of EEG signal preprocessing and feature extraction methods prior to employing the classification algorithm.In the comparison of multiclass classification performance for the 11 imagined speech signal of the KaraOne database, the proposed methodology seems to perform better than the other approaches, as shown in table 8. Similarly, table 9 shows that the proposed methodology significantly improves the previous state-of-the-art in multiclass classification of FEIS imagined speech EEG signals.

Limitation of the studies
The individual differences of EEG signals potentially include additional complexities in the discriminative feature extraction, transformation, and classification of the signals as the trained model of the imagined speech task-I gets applied to the imagined speech task-II, which could have contributed to the limiting of the classification performance improvement of the target imagined speech task-II, resulting in a marginal overall accuracy increase.
The imagined speech multiclass classification task is utilized in this studyʼs experiments using the transfer learning approach with a single source of information transmitted to a single target.Additional investigation is needed into the requirements for generalization learning in imagined speech decoding, taking individual signal variances and performance improvement into consideration, by using the most comprehensive information from multiple source knowledge domains transferred to a single target.More investigations will concentrate on multiple source task knowledge sharing for the imagined speech decoding, as multiple source transfer learning has attracted a considerable interest in real world applications [50,51].4.5.Current challenge and future work One difficulty has been that the EEG signal sets for the source and target brain signals are frequently different in terms of the number of scalp electrodes placed, which indirectly influences the signal properties.The deep transfer learning model architecture experiences challenges as the input sizes are different for source and target, so the parameters of the hidden layer shape and size would be different as well.Transfer learning is not achievable until the signal input dimensions are aligned with the input dimensions of the pre-trained models.The challenge was addressed by resizing the target domain signal dimension to match the size of the pre-trained framework.Although the approach proved effective, there was a possibility of important information being lost when reducing a neural imagined speech signal to a smaller size.Future research will investigate ways to align the target signals with the pre-trained model without having to scale the signals for the imgined speech cross-task transfer learning.

Conclusion
The proposed study illustrates the possibility of decoding imagined speech by utilizing transfer learning approaches based on deep learning algorithms.In these experiments, the EEG signal was processed using independent component analysis for artifact removal and wavelet transformation time-frequency analysis for feature extraction to prepare input signals before classifying imagined speech using the cross-domain and crosstask deep transfer learning model.The studies were conducted using a variety of transfer learning models that had been pretrained.The models could indeed be trained with smaller signal sets by freezing the transferred layers, reducing the number of network parameters to learn.The outcomes showed that multiclass imagined speech EEG signals could be classified using knowledge from other tasks or contexts, though the overall performance needs to improve further for the real-time BCI application.

Figure 1 .
Figure 1.The figure depicts the basis of the proposed deep cross-task transfer learning framework.On the left, a classifier trained on the source domain of image classification transfers weights to the target domain of imagined speech signal classification, implying inhomogeneous information sharing as the source domain and target domain are dissimilar.On the right, weights from the source task-I imagined speech signal classifier are transferred to the target task-II imagined speech signal classification, eventually homogeneous knowledge transfer.
proposed CNN model learning transfer architecture is divided into two phases.The first phase includes transferring the weights of a pre-trained ImageNet classification model to train the intermediate target domain imagined speech classification CNN model task-I.The second phase is imagined speech cross-task transfer learning, in which the trained task-I model weights are used to train the model for task-II of imagined speech classification.figure 3 depicts the proposed modelʼs framework.The task-I cross-domain deep transfer learning approach was used by freezing the convolution as well as dense layers and excluding the output layers of the pre-trained ImageNet classification model.After the

Figure 2 .
Figure 2. The figure illustrates the signal processing techniques used on the KaraOne and FEIS imagined speech EEG signals.

=
where x( k) denotes the outcomes of the networkʼs full connected output layer with k = 1,2,KN, and N is the number of imagined speech classes.The network x ( k) outcomes are determined to predict the likelihood of the interpretation result o ( k) corresponding to a specific imagined speech class.The networkʼs predicted imagined speech label is represented by the p ˆof signal E. The softmax value of the modelʼs predicted speech class is denoted by o ( k) , while the actual speech class of signal E is denoted categorically by p = [p (0) , p(1) ,...,p( N) ].

Figure 3 .
Figure 3.The proposed deep transfer learning model framework is depicted in the figure.The cross-domain transfer of the pretrained ImageNet model, in which the output layer is replaced by the few fully connected(FC) layers, followed by the softmax output layer, with the number of neurons equating to the number of imagined speech classes of task-I.The cross-task transfer of the trained task-I model, in which the output layer is replaced by the few fully connected layers, followed by the softmax output layer, with the number of neurons equating to the number of imagined speech classes task-II.

Figure 4 .
Figure 4.The figure illustrates KaraOne and FEISʼs imagined speech EEG signal reshaping aligned with the pretrained transfer learning model input form.

Figure 5 .
Figure 5.The figure depicts the confusion matrix of the proposed modelʼs experiment-1 classification findings of the target KaraOne signals.The predicted imagined speech class is shown by the x-axis, while the actual imagined speech class is represented by the y-axis.

Figure 6 .
Figure 6.The classification results of the target FEIS signals from experiment-2 of the proposed model are illustrated in the figure as a confusion matrix.The x-axis represents the predicted imagined speech class, whereas the y-axis represents the actual imagined speech class.

Table 1 .
The brief describes the characteristics of the FEIS imagined speech open-access EEG dataset.

Table 2 .
The brief outlines the characteristics of the KaraOne imagined speech open-access EEG dataset.

Table 3 .
Transfer learning model framework parameters.

Table 4 .
Overall validation accuracy for experiment-1 cross-task transfer learning model classification performance on target KaraOne dataset.

Table 5 .
The

Table 6 .
The proposed modelʼs classification performance for the distinct imagined speech classes of signals in KaraOne experiment-1 included metrics f1-score (f1), recall (R), and precision (P).The confusion matrix, which analyzes the effectiveness of the proposed model classification with regard to each imagined speech, is displayed in figure 5 of the experiment-1 DenseNet-201 based transfer learning model task-II KaraOne signals findings.The classification accuracy ranges from 0.72 to 0.94, which correlates to the metric precision and f1-score presented in table 6.The precision results range from 0.74 to 1.00, and the f1-score results range from 0.75 to 0.95.The observation shows that the accuracy of KaraOneʼs classification predictions for each simulated speech class varies, which could be due to the smaller number of records in KaraOne impacting the learning of the model.Also in the experiment-2 ResNet152V2 based transfer learning model task-II FEIS signals findings, the confusion matrix is shown in figure6.The classification accuracy varies from 0.53 to 0.96, which further coincides with the metric precision and f1-score shown in table7.The f1-score findings range from 0.72 to 0.98,

Table 7 .
The classification performance of the proposed model for the different imagined speech class types of the signals in FEIS experiment-2 comprised metrics f1score (f1), recall (R), and precision (P).outcomes range from 0.62 to 0.96.The results revealed that the accuracy of FEIS classification predictions differed or ranged considerably with respect to each class of imagined speech.Although cross-task transfer learning led to imagined speech signal discrimination in both cases of the KaraOne and FEIS signals, the findings indicate that the transfer learning models varied in the multiclass classification of each EEG imagined speech cue.

Table 8 .
State-of-the-art performance comparison of classification performance on the KaraOne dataset.

Table 9 .
State-of-the-art performance comparison of classification performance on the FEIS dataset.