Non-invasive decoding of imagined movements has significant potential in developing neuro-prosthetics and assistive technology for people with motor disabilities. While there has been extensive research conducted on classifying bilateral movement imaginations, classification of the kinematics associated with movement imaginations of unilateral limb is still an open research problem. Decoding the movement imaginations of a single limb can enhance human computer interaction by enabling more natural and intuitive movement control. This work aims to classify the direction of imagined hand movement in a four directional motor imagery (MI) task. Electroencephalogram (EEG) is recorded from 14 healthy subjects while they imagined center out movement of their right hand in four orthogonal directions (right, left, up and down). A Wavelet Phase Common Spatial Pattern (WPCSP) method is proposed to extract useful features from EEG to decode the imagined movement directions. The proposed method extracts informative features from the instantaneous phase signal using Common Spatial Pattern and are then classified using Linear Discriminant classifier, resulting in a mean binary direction classification accuracy of 68.54±7.5% and four-class direction classification accuracy of 38.48±8.1% among 14 healthy subjects. The results highlight the significance of phase-based features in decoding imagined kinematics of unilateral limb movements. This outcome is a step towards achieving higher degrees of freedom of movement and enhancing the efficacy of rehabilitation strategies.

Motor imagery (MI)-based Brain-Computer Interface (BCI) systems decode the neuronal patterns during motor movement imagination tasks and induce neuroplasticity to assist patients with stroke rehabilitation. Deep Learning methods achieve better generalization but require a large dataset for which longer calibration sessions are required. Transfer Learning methods learn the shift in data distribution from source to the target domain without the need for a large dataset. The existing cross-subject and cross-session transfer learning methods learn the shift in input data distribution alone, ignoring label distribution. This work proposes a cross-task transfer learning to learn new MI tasks by realigning the parameters of the deep learning model. The parameters of the pre-trained EEGNet model, trained on binary MI tasks, are re-aligned to decode a new MI task using Conditional Domain Adversarial Network (CDAN), while improving the generalizability across existing and new MI tasks. The proposed method is evaluated on the BCI Competition IV 2a dataset and compared with the popular fine-tuning approach for transfer learning in deep learning. The classification results achieved using the proposed CDAN-based cross-task transfer learning approach generalize better compared to conventional training and the fine-tuning approaches for transfer learning. The proposed method achieved an improvement in classification accuracy of around 6% for learning a third MI task from a pre-trained binary MI task classifier using EEGNet. The proposed CDAN-based cross-task transfer learning approach can significantly reduce the calibration session to learn new MI tasks, as it offers to learn the shift in input and output label distribution due to new MI tasks. Further, the proposed method can be scaled to learn dexterous MI tasks and kinematics relation information from MI tasks.

Motor Imagery (MI)-based Brain-Computer Interface (MI-BCI) systems induce neuroplasticity, promoting rehabilitation in stroke-affected patients. Decoding of kinematics information such as speed from unilateral hand MI tasks provides more natural control of the BCI systems. However, decoding speed related information from unilateral MI tasks is challenging due to the significant spatial overlap of neuronal sources and the inherently low spatial resolution of EEG. To address this, we propose EEGScaler, an end-to-end deep learning framework designed to decode slow v/s fast MI tasks by adaptively scaling EEG samples and electrodes with high discriminative value. The work proposes electrode-scaling and sample-scaling blocks to learn the importance of electrodes (i.e., spatial) and samples (i.e., temporal) in decoding speed from MI tasks. Further, spatiotemporal features are extracted using temporal and depth-wise convolution filters. In this work, subject-independent data is used to learn the weights of the EEGScaler and then subject-specific data is used to fine-tune the weights of the proposed electrode-scaling and sample-scaling blocks. The proposed method is implemented on 14 healthy subjects’ EEG data to classify slow v/s fast MI tasks performed using their dominant hand. The cross-validated subject-specific classification accuracy achieved using the proposed method outperformed the existing methods by ~7%. EEGScaler is a novel end-to-end learning model designed to assign importance to spatial and temporal information in EEG via electrode- and sampling-scaling blocks, respectively. Decoding of kinematics such as speed of MI tasks increases the degree of freedom in BCI systems, paving the way for more intuitive and efficient neurorehabilitation applications. This advancement has the potential to improve motor rehabilitation strategies by enabling more precise and adaptive BCI-driven therapy tailored to individual recovery needs.

Attention is a critical cognitive function in daily life, and its impairment can lead to serious consequences. Electroencephalogram (EEG) enables detection of brain activity related to attention, but decoding attentional states from EEG signals remains challenging due to noise artefacts and low spatial resolution. While deep learning approaches have shown promise, their generalizability across subjects remain limited. This work proposes a novel end-to-end hybrid deep learning architecture, STESA-Net, for classifying attention versus inattention states from multichannel EEG signal. STESA-Net integrates spatiotemporal convolutional layers, self-attention mechanism, and Bidirectional Long Short-Term Memory (Bi-LSTM) units to effectively capture and weight relevant spatial and temporal features. The design enhances the model’s ability to learn subject-specific patterns while maintaining generalizability across individuals. The proposed method is validated using 11 subjects EEG data recorded during a simulated driving task, employing leave-one-out-validation (LOOV) method. The proposed method achieved an average attention v/s inattention classification accuracy, precision, recall and F1 score of 78. 47%, 80. 16%, 78. 46%, and 78. 14%, respectively. The improvement in average classification accuracy achieved by the proposed method over the existing methods is ~7%. The results highlight the effectiveness of combining convolutional layers, self-attention mechanisms, and Bi-LSTM networks for robust cross-subject attentional state classification. The proposed framework offers a promising solution for real-world applications in attention monitoring and may aid the early detection of attention-related cognitive impairments with reduced computational overhead and enhanced generalization.

Brain-Computer Interface (BCI) systems decode specific neuronal signatures from the brain and can provide an alternative communication pathway for patients suffering from neuromuscular disorders. Decoding complex neuronal signatures such as direction, speed, orientation from Electroencephalogram (EEG) is challenging due to complex network connectivity in the brain. In the existing work, phase and amplitude-related features are used to decode the direction information during motor execution task. In this paper, we aim to decode the binary direction (left v/s right direction) during hand motor imagination task. We use the phase-related feature known as Wavelet-Phase Locking Value (W-PLV) to decode the hand motor imagery direction on calibration session data of 20 healthy subjects. Further, an online session is conducted on 13 subjects that use the trained model from the calibration session to decode imagined movement direction. The average classification accuracy achieved on calibration and online session data using phase related features (W-PLV) is 64. 39 ± 4. 23% and 58. 67 ± 4. 06%, respectively. The phase related features (W-PLV) achieve an improvement of 4. 82% and 3. 04% on calibration and online sessions, respectively compared to the amplitude related features alone. Online decoding of user’s intentions such as direction can aid in achieving more natural control of a BCI system.

To improve the spatial resolution of the electroen-cephalogram (EEG) signal, it is conventional to use a large number of scalp electrode while recording oscillatory rhythms in motor imagery based brain-computer interface (MI-BCI). However, this increases the dimension of data and might fail to generalize and thus over-fit. Therefore, it is required to reduce the dimension of input data in an optimal way. In this paper, we propose a method using an artificial neural network to reduce the dimension of EEG for MI-BCI. We train an under-complete sparse autoencoder neural network for each subject separately to encode the EEG data optimally. The optimally encoded EEG trials are then used by the Filter Bank Common Spatial Pattern (FBCSP) method to decode the imagined motor movement. In similar lines, autoencoder was trained on subject independent data. We achieved improved motor imagery classification accuracy when the dimension of the data was almost reduced by half compared to the state-of art FBCSP. The performance of the proposed method is also compared with Sparse Common Spatial Pattern (SCSP) based channel selection method. The average classification accuracy obtained for 10 subjects is 74.3±8.06 % with only 13 encoded channels. Also, for the autoencoder trained to be subject independent we obtained an average classification accuracy of 66.64±3.93% with only 11 encoded channels after cross-validation. The study extends the use of autoencoder neural networks in motor imagery based brain-computer interface and shows significant improvement in performance with reduced data dimension.