Motor imagery (MI) is the most studied BCI paradigm. The user imagines moving the left hand, the right hand, both feet or the tongue, and each produces a different spatial pattern of power change over the sensorimotor cortex. Because the signal comes from the user rather than from a flashing stimulus, MI is called an active or endogenous BCI. That makes it attractive for control and for stroke rehabilitation, and harder to learn than the stimulus-driven paradigms.
The effect is event-related desynchronisation (ERD): a drop in mu (8–12 Hz) and beta (13–30 Hz) power that begins during imagery, strongest over the hemisphere opposite the imagined hand. After the imagery ends, beta often rebounds above baseline. This is event-related synchronisation (ERS).
Where to look
| Imagined movement | Strongest ERD | Typical electrodes |
|---|---|---|
| Right hand | Left sensorimotor cortex | C3 and neighbours |
| Left hand | Right sensorimotor cortex | C4 and neighbours |
| Both feet | Midline, deep in the central sulcus | Cz |
| Tongue | Lateral, lower sensorimotor cortex | Broad, harder to localise |
A standard motor imagery pipeline
Bandpass filter to roughly 8–30 Hz, epoch from the cue onset over the imagery period, then extract spatial and spectral features. Common Spatial Patterns with log-variance features and a linear classifier is the classic baseline, and its filter-bank extension (FBCSP) won the BCI Competition IV four-class dataset. Riemannian classifiers on covariance matrices and compact CNNs such as EEGNet are the modern alternatives.
In BCILattice each of those is available. The analysis suite has epoching with baseline correction and rejection, an ERD/ERS feature family with its own baseline window, and a time-frequency step whose Summary view reports peak ERD/ERS in dB, peak time and peak frequency per class and band. On the ML canvas, CSP (with regularized and sparse variants as channel selection methods), the Riemannian family and EEGNet, ShallowConvNet, DeepConvNet, ATCNet and EEGConformer are blocks.
Public datasets to start with
BCI Competition IV 2a
Nine subjects, four classes (left hand, right hand, feet, tongue), 22 EEG channels at 250 Hz, two sessions each. The standard four-class benchmark.
BCI Competition IV 2b
Nine subjects, left versus right hand, three bipolar channels over C3, Cz and C4. A small-montage benchmark.
PhysioNet EEG Motor Movement/Imagery
109 subjects, 64 channels at 160 Hz, real and imagined movement. Large enough for cross-subject work.
On BCINexus
The dataset library lists published datasets with paradigm metadata, and benchmarks compare models on shared motor imagery tasks.
Why motor imagery is hard
Not everyone produces a clear, decodable ERD. A substantial minority of users never reach reliable control even with training. The literature calls this "BCI inefficiency". Performance also drifts between sessions as cap placement and impedance change, and transfers poorly between people.
That is why the split matters. A within-session random split can report excellent accuracy for a decoder that fails the next day. Evaluate across sessions or subjects (see EEG cross-validation), and for cross-subject work start with Euclidean Alignment and domain adaptation.
Questions
What is motor imagery in BCI?
Motor imagery is a BCI paradigm in which the user imagines a movement, such as moving the left or right hand, without actually moving. The imagined movement changes mu and beta rhythm power over the motor cortex, and a classifier decodes which movement was imagined.
What is ERD/ERS?
Event-related desynchronisation (ERD) is a drop in band power relative to a baseline, and event-related synchronisation (ERS) is a rise. In motor imagery, mu and beta ERD appear over the motor cortex opposite the imagined limb.
Which electrodes are used for motor imagery?
C3, Cz and C4 and the electrodes around them over the sensorimotor cortex. C3 reflects right-hand imagery, C4 left-hand imagery, and Cz foot imagery.
What is the best classifier for motor imagery EEG?
CSP with a linear classifier is the classic baseline, Riemannian tangent-space classifiers are strong and fast, and compact CNNs such as EEGNet perform well with enough data. Compare them under the same subject- or session-grouped evaluation.
How many motor imagery classes can a BCI decode?
Two classes (left versus right hand) is the most reliable. Four classes (left hand, right hand, feet, tongue) is the common benchmark and noticeably harder.
Read the reference
Related
Common Spatial Patterns (CSP) for EEG
CSP finds channel combinations whose variance differs most between two classes. Log-variance of those few filtered signals is often all a linear classifier needs.
EEGNet: The Compact CNN for EEG Explained
EEGNet (Lawhern et al., 2018) is a small convolutional network whose layers mirror classical EEG processing: frequency filters, then spatial filters, then a summary over time. It has a few thousand parameters and trains on the data you actually have.
Riemannian Geometry for BCI: MDM and Tangent Space
Instead of filtering channels and computing features, Riemannian methods treat each trial's covariance matrix as a point on a curved space and classify it there. They are simple, fast and hard to beat.
BCI Domain Adaptation
CORAL, MMD alignment, Transfer Component Analysis, a DANN discriminator, feature alignment and subject adaptation — assembled on the same canvas as the model they feed.