Movement control is a primary application for brain-computer interfaces (BCIs), offering potential solutions for cursor control, robotic limb manipulation, and the reanimation of paralyzed limbs. However, transitioning these technologies from laboratory settings to real-world assistive tools requires "dependability"—a combination of safety, portability, minimal recalibration, and robustness against artifacts. This paper reviews recent advancements in non-invasive neural recording technologies, specifically focusing on four studies that address signal decoding, user training via neurofeedback, and feature optimization to bridge the gap between experimental prototypes and dependable assistive devices.
Introduction
While BCIs have demonstrated significant promise in controlled environments, they often lack the consistency required for daily use by end-users. Current systems struggle with signal noise, high inter-user variability, and the need for frequent, time-consuming recalibration. To address these challenges, researchers are exploring diverse signal modalities, including scalp-based electroencephalography (EEG) and surface electromyography (sEMG), alongside advanced decoding algorithms and closed-loop feedback mechanisms. The goal is to create systems that are not only accurate but also resilient to the dynamics of real-world movement tasks.
Advancing Gesture Recognition through sEMG
A critical component of restoring natural movement is understanding peripheral muscle activity. Peng et al. (2022) investigated the use of high-density sEMG to enhance the accuracy and stability of gesture recognition. By combining feature selection with an Ensemble Extreme Learning Machine (EELM), the researchers achieved over 75% accuracy in classifying 52 distinct hand movements across 10 participants. This study highlights the potential of sEMG to provide high-dimensional control signals that can complement or, in some cases, substitute for direct brain signals in movement execution tasks.
Enhancing User Performance via Neurofeedback
BCI performance is often limited by the user's ability to generate consistent neural patterns, particularly in motor imagery tasks. Zhou et al. (2022) explored whether training the user, rather than just the algorithm, could improve outcomes. Their study utilized alpha-band (mu-band) neurofeedback to help participants up-regulate their resting-state alpha power. The results indicated a positive correlation between increased alpha power and improved motor imagery BCI performance. This "closed-loop" approach suggests that neurofeedback training can be a vital tool in reducing user variability and enhancing the dependability of imagery-based systems.
Optimizing Feature Representation and Decoding Pipelines
Two studies in this collection focused on the mathematical frameworks used to interpret EEG signals:
Metric Learning for Feature Extraction: Plucknett et al. (2022) applied supervised dimensionality reduction (metric learning) to a binary movement task. While the technique did not significantly boost classification accuracy, it proved invaluable for identifying critical EEG channels. Their findings reinforced the importance of low-frequency activity (<5Hz) as a primary feature for movement discrimination.
Source-Space vs. Sensor-Space Decoding: Srisrisawang and Müller-Putz (2022) compared decoding hand trajectories directly from electrode data (sensor-space) against estimating activity at specific brain locations (source-space). Surprisingly, they found no significant performance difference between the two, nor between subject-specific and template head models. This suggests that for velocity decoding, simpler sensor-space models may offer a more computationally efficient and thus more "dependable" path forward for portable BCIs.
Discussion and Future Directions
The research presented here underscores a shift toward multi-modal and user-centric BCI design. By integrating peripheral signals (sEMG), training users through neurofeedback, and identifying the most robust neural features (low-frequency activity), the field is moving closer to dependable technology.
However, significant hurdles remain. Most studies still rely on small sample sizes, and the transition from correlation-based velocity decoding to discrete, high-accuracy classification in noisy environments is ongoing. Future work must prioritize artifact mitigation—ensuring that BCIs remain functional even when the user is moving their body or experiencing external stimulation. Ultimately, the path to dependability lies in creating systems that are as adaptable and resilient as the biological systems they aim to assist.
Summary updated
Abstract
Movement control is among the top brain-computer interface (BCI) application scenarios for potential end-users. Movement control could involve moving a computer cursor on a screen, controlling a robotic arm, or reanimating one’s own limbs. Current approaches vary in numerous factors, including signal modality (ranging from single unit spiking activity up to scalp-based electroencephalographic (EEG) recordings), control strategies, decoding algorithms, artifact mitigation, and feedback modalities (e.g. visual or somatosensory). In spite of this diversity, there are shared key challenges that must be solved in order to deliver dependable assistive technology that will meet the needs of end-users. In addition to safety, availability, and portability, a dependable BCI has to eliminate or alleviate time-consuming system(re-)calibration processes, be independent of artifacts introduced by body movements or external stimulations, and generalize to various movement tasks with different dynamics.The goal of this Research Topic was to share recent advances among diverse approaches and, thereby, facilitate exchange between groups and research directions. We aimed for a scope ranging from original neuroscientific studies investigating neural correlates of goal-directed movements to neural engineering applications towards dependable BCIs for movement control. Research Topic CoverageIn this Research Topic we collected four original scientific papers which shed light to specific aspects of the quest to develop dependable BCI technology. All of the articles use non-implanted neural recording technologies. Three manuscripts use EEG recordings of brain activity, while the fourth manuscript uses surface electromyography (sEMG) recordings of muscle activity during a large variety on hand movements.BCIs rely on an understanding of the neural correlates of movement at the level of the brain, however in order to restore natural movement, it is also important to understand the peripheral activity of these movements, i.e. the EMG activity, to be able to further connect movement generation with movement execution in all levels.To enhance accuracy and stability of gesture recognition based on sEMG, Peng et al. (Peng et al. 2022), proposed a method combining feature selection and Ensemble Extreme Learning Machine (EELM) to improve the recognition performance based on sEMG signals. The authors compared the suggested methods with classical ensemble classifiers on a data set with 52 different hand movements from 10 able-bodied participants. Ultimately, sEMG could be used to classify these different movements with >75% accuracy.Many BCIs rely on motor imagery to generate consistent patterns of neural activity that can be used as a control signal, however imagery performance can vary widely from person-to-person. In earlier studies, it was found that resting-state alpha-band power, also described as mu-band power, is positively correlated with motor imagery-based BCI performance. Zhou et al. (Zhou et al. 2022a)(Zhou et al. 2022b) conducted a study to evaluate whether neurofeedback training based on up-regulation of the alpha band relative power had an effect on motor imagery BCI performance. A general finding was that such alpha neurofeedback training could successfully help study participants increase their alpha rhythm power and improve their motor imagery BCI performance. Also, on an individual level, participants with increased alpha power had a larger performance improvement.The next two papers in the collection, Plucknett et al. (Plucknett, Sanchez Giraldo, and Bae 2022) and Srisrisawang (Srisrisawang and Müller-Putz 2022), studied feature representations to aid subsequent decoding pipelines with regard to performance increases and interpretation. Plucknett et al. (Plucknett, Sanchez Giraldo, and Bae 2022) investigated how various combinations of feature extraction and metric learning algorithms affected feature representations in a self-paced, binary center-out movement task. As a supervised dimensionality reduction technique, the goal in metric learning is to project data in a low-dimensional space that maintains discriminative information. In their study with a publicly available EEG dataset comprising two subjects, they report that the considered metric learning algorithms did not improve classification performance but were useful to identify the importance of EEG channels encoding discriminative information. Their findings confirmed the importance of low-frequency (< 5Hz) EEG activity in movement discrimination tasks.The majority of EEG BCI studies for hand trajectory decoding rely on low frequency activity using features in sensor space, rather than source space which estimates the activity at locations in the brain from the sensor activity using inverse modeling techniques. In the study of Srisrisawang and colleague (Srisrisawang and Müller-Putz 2022), they tackle the problem of co-linearity due to the higher number of dimensions in the source space by two folds: (i) they selected signals in predefined regions of interest (ROIs) and (ii) they applied dimensionality reduction techniques like computing the mean (Mean), principal component analysis (PCA), and locality preserving projections (LPP) in pre-defined ROIs. Also the effect of using individual head models compared to a template model was investigated. Statistical tests revealed no significant differences between the source-space and sensor-space approaches when decoding reach velocities. Similarly, there was no significant difference between subject-specific and template head models. While no significant differences in performance were noted for correlation-based decoding of velocity, further investigations are still necessary when it comes to other decoding approaches such as a classification task.All articles in this collection addressed important issues concerning BCIs for movement control. Three papers investigated feature extraction and classification methods from EMG (Peng et al. 2022) and EEG (Plucknett, Sanchez Giraldo, and Bae 2022)(Srisrisawang and Müller-Putz 2022) and one EEG-based neurofeedback training (Zhou et al. 2022a). Despite the progress made in these and other works, further research is required towards establishing dependable BCIs for movement control.
BibTeX
@inproceedings{mullerputz2023editorial,
author = {Gernot Müller-Putz and Jennifer L. Collinger and Reinmar J Kobler},
title = {{Editorial: Towards dependable brain computer/machine interfaces for movement control}},
booktitle = {Frontiers in Human Neuroscience},
year = {2023},
doi = {10.3389/fnhum.2023.1186423},
}