Table 1.
EEG dataset description. All EEG recordings were obtained from non-disabled participants.
Fig 1.
Overview of the EEG analysis pipeline.
After preprocessing, including ICA artifact removal, the cleaned EEG data were resampled and bandpass filtered to isolate the alpha/mu range. EEG data was then processed using an EKF-based oscillator tracking algorithm to estimate instantaneous frequency and magnitude for each channel. Robust linear regression was then applied to compute the slope of each feature across time, yielding topographical maps of alpha/mu frequency and magnitude changes (topographical slope maps). To ensure that these linear slopes captured sustained trends rather than transient fluctuations, a slope change likelihood curve was derived for both frequency and magnitude. The optimal filter configuration (cutoff frequencies and filter order) was selected by computing the correlation between EKF-estimated mu-band magnitude and task-related reference signals (e.g., binary motor labels or continuous kinematic traces), and identifying the band that maximized average absolute correlation across channels. These correlations were also visualized as topographical maps (task correlation maps), highlighting the spatial distribution of task-relevant modulation and aiding interpretation of the functional relevance of alpha/mu dynamics.
Fig 2.
Representative example of EKF-tracked mu (a) frequency and (b) magnitude from channel C3 of a participant in the Schalk2004 dataset.
The top panels display the raw estimated trajectories (black) along with the best-fit linear trend (cyan). The estimated slope of the mu frequency was 1.128 Hz/hour (p < 0.05, after correction for multiple comparisons), while the mu magnitude increased at a rate of 2.458 a.u./hour (p < 0.05, after correction for multiple comparisons). The bottom panels show the slope change likelihood curves, quantifying the probability of abrupt trend shifts over time based on binary change point detection smoothed with a noncausal Gaussian kernel. In the mu frequency panel (bottom left), the slope change likelihood sharply peaks near the session start (~0–2 minutes), indicating a prominent early change in slope. This is consistent with the visible shift in the frequency trajectory shown in the top-left panel. In contrast, the mu magnitude panel (bottom right) exhibits a delayed and more gradual peak in slope change likelihood around 14–16 minutes. Visually inspecting (b) suggests that mu magnitude started stabilizing toward the session’s end.
Fig 3.
(a) Grand-average EEG topographical maps of alpha/mu-band frequency slopes (top row), the percentage of subjects exhibiting a positive slope at each channel (middle row), and task correlations (bottom row), and (b) the corresponding maps for alpha/mu-band magnitude for the Schalk2004, Dreyer2023, Schwarz2020, and Pulferer2022 datasets (left to right; columnwise).
In the slope maps (top rows in (a) and (b)), red hues indicate increasing trends over time (i.e., rising frequency in Hz/hour or magnitude in a.u./hour), while blue hues indicate decreasing trends. The percentage maps (middle rows in (a) and (b)) depict, for each electrode, the fraction of participants exhibiting a statistically significant positive slope (i.e., increasing frequency or magnitude over time), providing a channel-wise measure of how consistently the direction of change is observed across individuals. The task correlation maps (bottom rows in (a) and (b)) show the spatial distribution of the absolute correlation between EKF-estimated alpha/mu-band dynamics and task-related reference signals (e.g., task vs. rest, movement labels, kinematics), averaged across participants. This correlation analysis was used to identify the frequency range most strongly modulated by the task, representing the functionally relevant mu rhythm. Topoplots were generated using the MATLAB toolbox from Víctor Martínez-Cagigal (2025): Topographic EEG/MEG plot (https://www.mathworks.com/matlabcentral/fileexchange/72729-topographic-eeg-meg-plot).
Fig 4.
(a) Scatter plots of estimated initial versus final alpha/mu frequency for all participants across the four datasets (Schalk2004, Dreyer2023, Schwarz2020, and Pulferer2022) and (b) PSDs from representative participants and channels, comparing the first and last 5 minutes of the BCI session.
In (a) initial and final frequencies were derived from the fitted robust linear regression of EKF-estimated frequency trajectories by evaluating the regression model at the beginning and at the end of each recording, rather than using instantaneous frequency estimates. For visualization, channels were grouped into anatomically defined regions of interest (ROIs: frontal, sensorimotor, and posterior; with the sensorimotor ROI comprising central and centroparietal electrodes). The identity line is shown for reference; points above the line indicate an increase in alpha/mu frequency over the session, whereas points below indicate a decrease. In (b) the top row illustrates cases where both mu peak frequency and magnitude increase over time, while the bottom row shows a decrease in mu peak frequency accompanied by a magnitude increase. Gray shaded regions indicate the dataset-specific alpha/mu frequency bands used for bandpass filtering.
Fig 5.
ROI-level correlation matrices summarizing inter-individual coupling of session-long alpha/mu dynamics across cortical regions for (a) alpha/mu frequency slopes and (b) alpha/mu magnitude slopes.
Rows correspond to datasets, and columns indicate pairwise correlations between regional slopes (frontal–central, posterior–central, frontal–centroparietal, and posterior–centroparietal). Only correlations surviving multiple-comparison correction were retained. Significant channel-wise correlations were subsequently averaged within ROIs. Color encodes the Pearson correlation coefficient across participants, with red indicating positive coupling and blue indicating inverse coupling. The matrices reveal consistent cross-dataset patterns, including inverse relationships between sensorimotor (central/centroparietal) frequency acceleration and posterior or frontal alpha dynamics.
Fig 6.
Comparison of absolute task correlations across spectral feature representations and datasets.
Boxplots summarize the distribution of subject-wise maximum absolute Pearson correlations between task reference signals and four EEG feature types: EKF-tracked alpha/mu frequency, EKF-tracked alpha/mu magnitude, the Frobenius norm of the EKF state covariance matrix (reflecting estimator uncertainty and dynamical stability), and conventional alpha/mu log-power, computed within the same dataset-specific frequency range used for EKF tracking by squaring and applying a base-10 logarithm to the bandpass-filtered signal. For each subject and feature, correlations were computed separately for all channels and the maximum absolute value across channels was retained to provide a single representative measure per participant. Results are shown for all four datasets, demonstrating that EKF-derived features exhibit consistently stronger task correlations than fixed-band log-power features (p < 0.05 after correction for multiple comparisons in all datasets).
Fig 7.
Slope maps of EKF-estimated (a) frequency (top row) and (b) magnitude (bottom row) changes across a wide range of bandpass filtered frequency bands, averaged over central EEG channels (i.e., Ci) and participants.
Columns correspond to different datasets (Schalk2004, Dreyer2023, Schwarz2020 and Pulferer2022 from left to right). Each pixel represents the mean slope (in Hz/hour) for a specific band defined by [fmin, fmax]. Positive slopes (red) indicate rising frequencies, while negative slopes (blue) indicate decreasing trends. Green boxes highlight the frequency bands selected to generate the topographical maps of Fig 3.