Total Perspective Vortex

PhysioNet EEG Motor Movement/Imagery: 109 subjects, 64-channel EEG at 160 Hz; runs 3–14, left/right fist and both fists/both feet.

Solo

1. Implemented Common Spatial Patterns from scratch — trace-normalised per-class covariances, a generalised eigenvalue decomposition and log-variance features — as a scikit-learn transformer.
2. Built the pipeline: the custom CSP feeding logistic regression, with k-fold cross-validation and model save and load.
3. Wrote the MNE preprocessing layer: EDF loading, an 8–30 Hz band-pass, event extraction and epoching.
4. Built the command-line tool (train, predict, stream) and a batch loop over all 109 subjects.

CSP written from scratch, not the library one

The subject requires a hand-implemented dimensionality reduction, so the covariances, the generalised eigenvalue problem and the projection matrix are computed directly; MNE's CSP is kept only as a reference to check against.

A scikit-learn-compatible transformer

The CSP subclasses BaseEstimator and TransformerMixin, so it composes with logistic regression in one pipeline and runs through cross-validation unchanged.

An 8–30 Hz band and a sensorimotor focus

Filtering is restricted to the mu and beta bands where motor-imagery activity lives, with epochs cut from −0.5 s to 4 s around each event.

Across all 109 subjects, the six experiment configurations average 62.5% accuracy (chance = 50%); the best (both fists vs both feet, real movement) reaches 71%, while left/right hand stays around 58–60%. The from-scratch CSP and the scikit-learn integration are the core of the project.

• Stream mode is offline replay, not live acquisition: it loops over pre-computed epochs one at a time and prints the processing time.
• One global model per task type, overwritten on each training run; there are no per-subject saved models.
• The evaluation is within-subject: each model is trained and tested on the same subject, with no generalisation to new subjects.
• Accuracy varies widely by subject and task (from ~0.58 on left/right hand to ~0.71 on fists/feet) — the from-scratch CSP and the scikit-learn integration remain the substance.