Brain-computer interface provides a new communication bridge between the human mind and devices, depending largely on the accurate classification and identification of non-invasive EEG signals. Recently, the deep learning approaches have been widely used in many fields to extract features and classify various types of data successfully. However, the deep learning approach requires massive data to train its neural networks, and the amount of data impacts greatly on the quality of the classifiers. This paper proposes a novel approach that combines deep learning and data augmentation for EEG classification. We applied the empirical mode decomposition on the EEG frames and mixed their intrinsic mode functions to create new artificial EEG frames, followed by transforming all EEG data into tensors as inputs of the neural network by complex Morlet wavelets. We proposed two neural networks-convolutional neural network and wavelet neural network-to train the weights and classify two classes of motor imagery signals. The wavelet neural network is a new type of neural network using wavelets to replace the convolutional layers. The experimental results show that the artificial EEG frames substantially improve the training of neural networks, and both two networks yield relatively higher classification accuracies compared to prevailing approaches. Meanwhile, we also verified the performance of our new proposed wavelet neural network model in the classification of steady-state visual evoked potentials. INDEX TERMS Motor imagery classification, deep learning, convolutional neural network, wavelet neural network, empirical mode decomposition, artificial EEG frames.

EEG-based Brain-Computer Interfaces (BCIs) are becoming a new tool for neurorehabilitation. BCIs are used to help stroke patients to improve the functional capability of the impaired limbs, and to communicate and assess the level of consciousness in Disorder of Consciousness (DoC) patients. BCIs based on a motor imagery paradigm typically require a training period to adapt the system to each user's brain, and the BCI then creates and uses a classifier created with the acquired EEG. The quality of this classifier relies on amount of data used for training. More data can improve the classifier, but also increases the training time, which can be especially problematic for some patients. Training time might be reduced by creating new artificial frames by applying Empirical Mode Decomposition (EMD) on the EEG frames and mixing their Intrinsic Mode Function (IMFs). The purpose of this study is to explore the use of artificial EEG frames as replacements for some real ones by comparing classifiers trained with some artificial frames to classifiers trained with only real data. Results showed that, in some subjects, it is possible to replace up to 50% of frames with artificial data, which reduces training time from 720 to 360 s. In the remaining subjects, at least 12.5% of the real EEG frames could be replaced, reducing the training time by 90 s. Moreover, the method can be used to replace EEG frames that contain artifact, which reduces the impact of rejecting data with artifact. The method was also tested on an out of sample scenario with the best subjects from a public database, who yielded very good results using a frame collection with 87.5% artificial frames. These initial results with healthy users need to be further explored with patients' data, along with research into alternative IMF mixing strategies and using other BCI paradigms.

Introduction: Recent studies explored promising new quantitative methods to analyze electroencephalography (EEG) signals. This paper analyzes the correlation of two EEG parameters, Brain Symmetry Index (BSI) and Laterality Coefficient (LC), with established functional scales for the stroke assessment. Methods: Thirty-two healthy subjects and thirty-six stroke patients with upper extremity hemiparesis were recruited for this study. The stroke patients where subdivided in three groups according to the stroke location: Cortical, Subcortical, and Cortical + Subcortical. The participants performed assessment visits to record the EEG in the resting state and perform functional tests using rehabilitation scales. Then, stroke patients performed 25 sessions using a motor-imagery based Brain Computer Interface system (BCI). BSI was calculated with the EEG data in resting state and LC was calculated with the Event-Related Synchronization maps. Results: The results of this study demonstrated significant differences in the BSI between the healthy group and Subcortical group (P = 0.001), and also between the healthy and Cortical+Subcortical group (P = 0.019). No significant differences were found between the healthy group and the Cortical group (P = 0.505). Furthermore, the BSI analysis in the healthy group based on gender showed statistical differences (P = 0.027). In the stroke group, the correlation between the BSI and the functional state of the upper extremity assessed by Fugl-Meyer Assessment (FMA) was also significant, ρ = −0.430 and P = 0.046. The correlation between the BSI and the FMA-Lower extremity was not significant (ρ = −0.063, P = 0.852). Similarly, the LC calculated in the alpha band has significative correlation with FMA of upper extremity (ρ = −0.623 and P < 0.001) and FMA of lower extremity (ρ = −0.509 and P = 0.026). Other important significant correlations between LC and functional scales were observed. In addition, the patients showed an improvement in the FMA-upper extremity after the BCI therapy (FMA = 1 median [IQR: 0-8], P = 0.002). Conclusion: The quantitative EEG tools used here may help support our understanding of stroke and how the brain changes during rehabilitation therapy. These tools can help identify changes in EEG biomarkers and parameters during therapy that might lead to improved therapy methods and functional prognoses.

No abstract