Motor-imagery decoding

I want to investigate whether EEG recordings can distinguish instructed imagined movements, such as imagining a left-hand versus right-hand movement. This is planned; I have not built or evaluated the decoder.

Motivation

A useful initial question is: given a recording window from a specified task, which task label is most consistent with it? This is narrower than decoding arbitrary thoughts, predicting a limb's position, or demonstrating control of a device.

Candidate data

The PhysioNet EEG Motor Movement/Imagery dataset includes recordings during performed and imagined motor tasks. It is a candidate, not yet a selected experimental dataset.

Its event codes depend on the run: a code can refer to a unilateral fist task in one run and a bilateral hand/foot task in another. An analysis must choose the task family and interpret labels with that context.

Proposed method

I plan to begin with a simple classification baseline before considering more complex sequence models. A classifier would assign probabilities to the selected classes:

$$\hat y=\underset{k}{\operatorname{argmax}}\;p_\theta(k\mid x),\qquad \sum_k p_\theta(k\mid x)=1.$$

Here $x$ is a defined EEG window and $k$ indexes task labels. A high probability is a model output, not proof that a person's intention has been read correctly. Calibration and error analysis need separate evaluation.

The target classes, windows, preprocessing and participant split must be established before choosing an architecture. Evaluating unseen people asks a different question from evaluating more windows from people already seen during training.

Results

No results are available. The future time-series display will align reference labels with predicted class probabilities. A scrubbable window view should make mistakes inspectable alongside aggregate scores.

Limitations and next steps

Scalp recordings can contain non-neural artifacts. A classifier may exploit acquisition or task cues; that possibility must be examined before interpreting behavior as neural decoding. Recorded task labels do not imply unconstrained intention recognition.

Next: choose executed movement or imagery, fix a task family, inspect candidate data and define a baseline. SNNs remain a separate learning topic, not an implemented method here.

Projects · Research direction

Project workflow

Figures and evaluation

Evaluation exports are not available on this site.

Project links

No public repository or demo is linked here.