Motor-imagery brain-computer interface electroencephalography and behavioural assessment datasets in prolonged disorders of consciousness

Data collection method:

Data were collected for this registered (ClinicalTrials.gov, NCT03827187; 30th January, 2019), ethically approved, multi-site clinical study investigating awareness detection and communication using electroencephalography (EEG)-based brain–computer interface (BCI) methods in individuals with prolonged disorders of consciousness (PDoC) and locked-in syndrome (LIS). Participants were recruited through clinical teams at NHS and partner rehabilitation centres following consultant screening for eligibility. Inclusion criteria comprised a diagnosis of PDoC, low awareness state, or (complete) locked-in syndrome, while exclusion criteria included progressive neurological disease, uncontrolled epilepsy, excessive movement interfering with EEG acquisition, or medications impairing cognition. Able-bodied participants were also recruited as benchmark controls using the same experimental protocol. Informed consent was obtained directly or via proxy where required, in accordance with approved ethical procedures. EEG data were acquired during structured motor imagery BCI sessions conducted at the bedside or in quiet clinical environments. Participants were instructed to perform specific motor imagery tasks (e.g. imagined arm or foot movements) in response to auditory cues. Movement imagery tasks were selected on an individual basis in consultation with clinicians or family members to avoid injured brain regions and maximise task feasibility. The experimental paradigm comprised three sequential phases: (i) an initial assessment phase to evaluate task-related EEG responses; (ii) training and feedback phases involving repeated imagery trials with real-time auditory feedback; and (iii) a question-and-answer paradigm for participants demonstrating reliable task-related neural modulation. Sessions typically lasted 1–2 hours, with multiple sessions recorded per participant depending on health status and task performance. EEG was recorded using a 16-channel wireless system with active electrodes positioned according to the international 10–20 system, referenced to the right earlobe and grounded at AFz. Signals were band-pass filtered (0.5–100 Hz) and sampled at 250 Hz (subsequently down-sampled to 125 Hz for analysis). Continuous EEG recordings were stored together with event triggers marking task cue onset, task offset, and rest periods, enabling trial-based segmentation during offline analysis. In parallel with EEG acquisition, standardised behavioural assessments were conducted during each recording session, comprising the Coma Recovery Scale–Revised (CRS-R) and the Wessex Head Injury Matrix (WHIM). These behavioural measures were collected at the session level to enable comparison with EEG-based BCI performance. All data were processed and stored using study-specific, pseudonymous participant identifiers, and no personal identifiers were retained in the research dataset. The resulting dataset therefore consists of continuous EEG recordings with associated event markers and session-level behavioural assessment scores, organised by participant and session.

Data processing and preparation activities:

No third-party datasets were used in this study. All data were collected by the research team. Following analysis, the dataset was prepared for publication by applying approved anonymisation procedures, including removal of all personal identifiers, replacement of participant identifiers with study-specific pseudonymous IDs, and reporting age in ranges rather than exact values. No other modifications were made to the research data prior to deposition.

Technical details and requirements:

EEG acquisition hardware EEG data were acquired using a g.Nautilus Wireless Research EEG system (g.tec medical engineering)[1] with 16 active electrodes positioned according to the international 10–20 system. The reference electrode was placed on the right earlobe and the ground electrode at AFz. Signals were hardware band-pass filtered (eighth-order Butterworth, 0.5–100 Hz) and sampled at 250 Hz, with data subsequently down-sampled to 125 Hz for offline processing and storage. Online acquisition and experimental control software EEG acquisition and online signal processing were implemented in MATLAB Simulink[2], which communicated with the experimental protocol controller via User Datagram Protocol (UDP). Experimental paradigms and auditory stimuli were implemented in the Unity 3D Game Engine[3], which also generated event markers corresponding to task cue onset, task offset, and rest periods that were recorded synchronously with EEG data. Offline analysis and deposited code The repository includes MATLAB/Simulink code to perform offline analysis and calibration for a single motor imagery BCI run (i.e. one participant, one recording session, and one motor imagery task pairing). The deposited code implements a filter-bank common spatial patterns (FBCSP) framework with mutual information–based feature ranking and classification[4]. EEG signals are filtered into six standard frequency bands using Simulink finite impulse response (FIR) band-pass filters: delta (0.5–4 Hz), theta (4–8 Hz), mu (8–12 Hz), low beta (12–18 Hz), high beta (18–28 Hz), and low gamma (28–40 Hz). Filter parameters were configured with 0 dB passband attenuation and 60 dB stopband attenuation. I. For each analysed run, the calibration procedure evaluates combinations of: two EEG channel sets (16-channel full montage; 9-channel motor-cortex-focused subset); II. three frequency band groupings, each comprising four adjacent bands (delta–low beta, theta–high beta, mu–low gamma); III. classification windows of 1 s and 2 s duration; IV. Between 4–10 of the highest ranked MI-selected features (i.e., feature subsets consisting of the top-ranked 4 to 10 mutual-information-ranked features). The purpose of the deposited code is to reproduce the single-run calibration and feature selection process used in the study and to enable methodological transparency and reuse. It is not intended to represent a complete end-to-end analysis pipeline for the full dataset. A readme.txt file is provided with the code, detailing software dependencies, required input formats, parameter settings, and instructions for running the analysis.. Additional technical details and analysis software Statistical analyses were performed using R (version 4.4.2, R Foundation for Statistical Computing) [5]. Topographical analyses and exploratory source-space visualisation were performed using sLORETA (standardised low-resolution brain electromagnetic tomography). These analyses were qualitative and intended for visualisation rather than as a quantitative inverse solution. Functional connectivity analyses were conducted in MATLAB using the FieldTrip toolbox, with eLORETA employed for source-space connectivity estimation where applicable. References [1] g.tec medical engineering, “g.NAUTILUS RESEARCH | Wearable EEG Headset,” 2020. https://www.gtec.at/product/gnautilus-pro/ (accessed May 24, 2020). [2] “Simulink for Matlab (The MathWorks, Inc.),” 2020. . [3] Unity Technologies, “Unity Real-Time Development Platform | 3D, 2D VR & AR Visualizations,” Unity Technologies, 2020. https://unity.com/ (accessed May 24, 2020). [4] A. Korik et al., “Competing at the Cybathlon championship for people with disabilities: long-term motor imagery brain–computer interface training of a cybathlete who has tetraplegia,” J. Neuroeng. Rehabil., 2022, doi: 10.1186/s12984-022-01073-9. [5] R Core Team, “R: A language and environment for statistical computing. R Foundation for Statistical Computing.” 2024, [Online]. Available: https://www.r-project.org/.

Additional information:

EEG data are organised hierarchically in a directory structure ordered by participant, recording session, and run. Each participant has a top-level directory, within which subdirectories correspond to individual recording sessions and runs. Each run directory contains a single MATLAB file (EEG_rec.mat) containing the continuous EEG recording and associated event markers for that run. Each participant is assigned a study-specific, pseudonymous identifier. Recording sessions and runs are denoted as ssn (where n indicates the session number), followed by the run ID (for assessment, training, feedback, and Q&A runs the ID contains the letter a, t, f, and qa respectively). EEG data are stored as continuous recordings in MATLAB (.mat) format and include embedded event markers indicating task cue onset and task offset for each run. Rest periods are defined as the −1000 to 0 ms interval preceding task-cue onset. Event markers enable trial-based segmentation of the EEG data during offline analysis. Behavioural data, comprising Coma Recovery Scale–Revised (CRS-R) and Wessex Head Injury Matrix (WHIM) scores, are provided in a separate spreadsheet file. Each behavioural record is linked to the corresponding EEG data via participant and session identifiers. Data collection occurred across two recruitment phases that employed different internal participant coding schemes; however, all identifiers used in the deposited dataset are non-identifying and carry no intrinsic meaning outside the study. No key enabling re-identification is included with the shared data. The repository also includes code required to reproduce the single-run calibration and feature selection process used in the study. A README file is provided to describe the directory structure, file naming conventions, software dependencies, and steps required to run the analysis code. The template documentation and administration manuals are for the behavioural assessments used in this study. – The Coma Recovery Scale-Revised (CRS-R) material can be found at https://www.sralab.org/rehabilitation-measures/coma-recovery-scale-revised. - The Wessex Head Injury Matrix (WHIM) can be requested via the Royal Hospital for Neuro-disability: info@rhn.org.uk. These materials are provided for transparency regarding the assessment procedures applied during data collection and to support interpretation of the behavioural scores included in the dataset.

Methodology link: