Anirban Chowdhury

Brain–computer interface (BCI)-based neuromotor training is widely recognized as a promising approach for post-stroke motor rehabilitation. However, its practical deployability remains uncertain due to challenges related to EEG headset wearability and the limited validation of emerging algorithms in noisy, real-world environments. These settings demand low-density, affordable EEG systems, robust multimodal neurofeedback paradigms, and validation with a sufficient number of participants in online scenarios. To address these challenges, we developed a BCI system using an 8-channel g.tec Unicorn headset and conducted experiments in a real-world noisy environment with 20 participants for offline analysis and another 20 for online evaluation. The system includes a hand exoskeleton to deliver multimodal concomitant neurofeedback. We evaluated both a traditional method (Common Spatial Pattern+Support Vector Machine(CSP+SVM)) and a deep learning architecture (Filter-Bank Convolutional Neural Network(FBCNet)) across this setup. Our results show that incorporating data preprocessing and a longest consecutive repetition (LCR) based post-processing step significantly (p-value < 0.05) improves FBCNet’s performance in cross-subject decoding. In within-subject scenarios, CSP+SVM remains competitive with FBCNet, offering comparable accuracy with reduced training time. This study sets a comprehensive benchmark for the real-world deployment of wearable BCI systems for neuromotor training and provides a valuable testbed for evaluating future algorithms in practical, noisy environments.

BCIs facilitate seamless engagement between individuals with motor disabilities and their surrounding environment by translating electroencephalography (EEG) signals generated from Motor Imagery (MI). Crucial to this process is the accurate classification of different types of MI tasks - a challenge that calls for the consistent evolution and refinement of reliable methodologies for EEG signal classification. This paper introduces three innovative approaches: M1, employing a temporal block technique combined with Filter Bank Common Spatial Pattern (FBCSP) and mutual information-based feature selection with a Random Forest classifier; and M2 and M3, extending M1 using Temporal Probability Fusion (TPF) and Probability Difference-based Temporal Fusion (PDTF) respectively. These methods aim to enhance MI EEG signal classification. The effectiveness of M1, M2, and M3 was scrutinized under differing scenarios including changing overlap sizes and channel choices. The analysis highlights that our methods exhibit enhanced performance under particular conditions, underlining the crucial role of temporal information and channel selection. Comparison with established methodologies verifies the superior efficiency of our proposed strategies. This study foregrounds the considerable potential of TPF and PDTF in MI EEG classification tasks, with significant implications for the future development of BCI systems.

Common spatial pattern (CSP) is a widely used method for feature extraction in motor imagery (MI)-based brain-computer interface (BCI) development. However, the performance of traditional CSP features often lacks robustness against inter-session and inter-subject variabilities present in MI-related electroencephalogram (EEG) signals. To address this limitation, we propose a novel approach to CSP-based feature extraction, combining spectral information obtained from Welch power-spectrum (PS) estimation with temporal variations which we named here as SCSP-3. Our SCSP-3 method employs independent learning paths for the temporal and spectral features extracted through CSP. We introduce a postprocessing step that crosses the classification probabilities from these pathways using element-wise products, deriving linearly separable features. The performance of SCSP-3 is evaluated and compared to the traditional CSP approach utilizing a support vector machine (SVM) for classification following a within-subject evaluation scheme. The results demonstrate a significant improvement in average accuracy for SCSP-3 with more generalizability, as it performs equally well with datasets from healthy subjects and stroke patients. This enhanced robustness and generalizability highlight the potential of SCSP-3 as a superior alternative to traditional CSP-based feature extraction methods for achieving consistent performance across different subject categories.

Assisting paralysed people for conducting activities of daily living (ADL) is a major challenge in the area of collaborative robotics. The first predicament of wheelchair mounted robotic arms and/or exoskeletons is to identify and localize the objects for ADL in a 3-dimensional (3D) space so that they can be manipulated according to the user’s need. This requires robust integration and adaptation of state-of-the-art algorithms in a seamless framework for faster, accurate, and reliable manner ensuring the user’s safety. To this end, the current work focuses on the adaptation of the popular YOLO algorithm for image based object identification for detecting objects for ADL. In order to facilitate 3D localization of these objects we integrated the YOLO with a stereo vision based algorithm for depth perception. Thus the integrated system can provide both object identification and 3D localization in real-time so that a wheelchair mounted robotic arm and/or exoskeleton can be semi-autonomously guided to perform ADL for the disabled people. The proposed solution is resource efficient as it can be implemented using low-cost hardware so that the overall product can be affordable, environmentally efficient (as it may consume less power due to low computational complexity). This will lead to enhanced environment perception for the robots which can drastically improve the human robot interaction models for optimal assistance.

The need for repeated calibration and accounting for inter subject variability is a major challenge for the practical applications of a brain-computer interface. The problem becomes more severe when it is applied for the neurorehabilitation of stroke patients where the brain-activation pattern may differ quite a bit as compared to healthy individuals due to the altered neurodynamics because of a lesion. There were several approaches to handle this problem in the past that depend on creating customized features that can be generalized among the individual subjects. Recently, several deep learning architectures came into the picture although they often failed to produce superior accuracy as compared to the traditional approaches and mostly do not follow an end-to-end architecture as they depend on custom features. However, a few of them have the promising ability to create more generalizable features in an end-to-end fashion such as the popular EEGNet architecture. Although EEGNet was applied for decoding stroke patient’s motor imagery (MI) data with limited success it failed to achieve superior performance over the traditional methods. In this study, we have augmented the EEGNet based decoding by introducing a post-processing step called the longest consecutive repetition (LCR) in a sliding window-based approach and named it EEGNet+LCR. The proposed approach was tested on a dataset of 10 hemiparetic stroke patients’ MI data set yielding superior performance against the only EEGNet and a more traditional approach such as common spatial pattern (CSP)+support vector machine (SVM) for both within- and cross-subject decoding of MI signals. We also observed comparable and satisfactory performance of the EEGNet+LCR in both the within- and cross-subject categories which are rarely found in literature making it a promising candidate to realize practically feasible BCI for stroke rehabilitation.

Brain-computer interface (BCI) performance is often impacted due to the inherent non-stationarity in the recorded EEG signals coupled with a high variability across subjects. This study proposes a novel method using Logistic Regression with Tangent Space-based Transfer Learning (LR-TSTL) for motor imagery (MI)-based BCI classification problems. The single-trial covariance matrix (CM) features computed from the EEG signals are transformed into a Riemannian geometry frame and tangent space features are computed by considering the lower triangular matrix. These are then further classified using the logistic regression model to improve classification accuracy. The performance of LR-TSTL is tested on healthy subjects’ dataset as well as on stroke patients’ dataset. As compared to existing within-subject learning approaches the proposed method gave an equivalent or better performance in terms of average classification accuracy (78.95±11.68%), while applied as leave-one-out cross-subject learning for healthy subjects. Interestingly, for the patient dataset LR-TSTL significantly (p<0.05) outperformed the current benchmark performance by achieving an average classification accuracy of 81.75±6.88%. The results show that the proposed method for cross-subject learning has the potential to realize the next generation of calibration-free BCI technologies with enhanced practical usability especially in the case of neurorehabilitative BCI designs for stroke patients.B.

In the field of brain-computer interface (BCI) research, the availability of high-quality open-access datasets is essential to benchmark the performance of emerging algorithms. The existing open-access datasets from past competitions mostly deal with healthy individuals’ data, while the major application area of BCI is in the clinical domain. Thus the newly proposed algorithms to enhance the performance of BCI technology are very often tested against the healthy subjects’ datasets only, which doesn’t guarantee their success on patients’ datasets which are more challenging due to the presence of more non-stationarity and altered neurodynamics. In order to partially mitigate this scarcity, Clinical BCI Challenge aimed to provide an open-access rich dataset of stroke patients recorded similar to a neurorehabilitation paradigm. Another key feature of this challenge is that unlike many competitions in the past, it was designed for algorithms in both with-in subject and cross-subject categories as a major thrust area of current BCI technology is to realize calibration-free BCI designs. In this paper, we have discussed the winning algorithms and their performances across both competition categories which may help develop advanced algorithms for reliable BCIs for real-world practical applications.

Accurate binary classification of electroencephalography (EEG) signals is a challenging task for the development of motor imagery (MI) brain-computer interface (BCI) systems. In this study, two sliding window techniques are proposed to enhance the binary classification of MI. The first one calculates the longest consecutive repetition (LCR) of the sequence of prediction of all the sliding windows and is named SW-LCR. The second calculates the mode of the sequence of prediction of all the sliding windows and is named SW-Mode. Common spatial pattern (CSP) is used for extracting features with linear discriminant analysis (LDA) used for classification of each time window. Both the SW-LCR and SW-Mode are applied on publicly available BCI Competition IV-2a data set of healthy individuals and on a stroke patients' data set. Compared with the existing state of the art, the SW-LCR performed better in the case of healthy individuals and SW-Mode performed better on stroke patients' data set for left- versus right-hand MI with lower standard deviation. For both the data sets, the classification accuracy (CA) was approximately 80% and kappa (κ) was 0.6. The results show that the sliding window-based prediction of MI using SW-LCR and SW-Mode is robust against intertrial and intersession inconsistencies in the time of activation within a trial and thus can lead to a reliable performance in a neurorehabilitative BCI setting.

The effect of corticomuscular coactivation based hybrid brain-computer interface (h-BCI) on post-stroke neurorehabilitation has not been explored yet. A major challenge in this area is to find an appropriate corticomuscular feature which can not only drive an h-BCI but also serve as a biomarker for motor recovery monitoring. Our previous study established the feasibility of a new method of measuring corticomuscular co-activation called correlation of band-limited power time-courses (CBPT) of EEG and EMG signals, outperforming the traditional EEG-EMG coherence in terms of accurately controlling a robotic hand exoskeleton device by the stroke patients. In this paper, we have evaluated the neurophysiological significance of CBPT for motor recovery monitoring by conducting a 5-week long longitudinal pilot trial on 4 chronic hemiparetic stroke patients. Results show that the CBPT variations correlated significantly ( ${p}$ -value< 0.05) with the dynamic changes in motor outcome measures during the therapy for all the patients. As the bandpower based biomarkers are popular in literature, a comparison with such biomarkers has also been made to cross-verify whether the changes in CBPT are indeed neurophysiological. Thus the study concludes that CBPT can serve as a biomarker for motor recovery monitoring while serving as a corticomuscular co-activation feature for h-BCI based neurorehabilitation. Despite an observed significant positive change between pre- and post-intervention motor outcomes, the question of the clinical effectiveness of CBPT is subject to further controlled trial on a larger cohort.

Brain-machine interface (BMI) driven robot-assisted neurorehabilitation intervention has demonstrated improvement in upper-limb (UL) motor function, specifically, with post-stroke hemiparetic patients. However, neurophysiological patterns related to such interventions are not well understood. This study examined the longitudinal changes in band-limited resting-state (RS) functional connectivity (FC) networks in association with post-stroke UL functional recovery achieved by a multimodal intervention involving motor attempt (MA) based BMI and robotic hand-exoskeleton. Four adults were rehabilitated with the intervention for a period lasting upto 6 weeks. RS magnetoencephalography (MEG) signals, Action Research Arm Test (ARAT), and grip strength (GS) measures were recorded at five equispaced sessions over the intervention period. An average post-interventional increase of 100.0% (p=0.00028) and 88.0% were attained for ARAT and GS, respectively. A cluster-based statistical test involving correlation estimates between beta-band (15–26 Hz) RS-MEG FCs and UL functional recovery provided positively correlated sub-networks in both contralesional and ipsilesional motor cortices. The fronto-parietal FC exhibited hemispheric lateralisation wherein majority of the positively and negatively correlated connections were found in contralesional and ipsilesional hemispheres, respectively. Our findings are consistent with the theory of bilateral motor cortical association with UL recovery and predict novel FC patterns that can be important for higher level cognitive functions.