EEG motor imagery classifier — CSP + LDA (or small CNN) for left/right/feet/rest #19

New issue

Open

opened 2026-04-26 21:47:44 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-26 21:47:44 -07:00

Owner

Phase C learned classifier for EEG motor imagery. Classifies 4 classes: left hand imagine, right hand imagine, feet imagine, rest.

Feature extraction:

Band power per channel: delta (1–4Hz), theta (4–8Hz), alpha (8–13Hz), beta (13–30Hz) — via scipy.signal
Common Spatial Pattern (CSP) filters: maximizes variance for one class vs. rest
ICA (Independent Component Analysis) artifact removal: eye blink (delta burst in frontal channels), EMG contamination (broadband high-freq)

Classifier: CSP + LDA (fast, interpretable, trains in seconds). Optional: small CNN on raw EEG windows for higher accuracy.

Training session UI (in Avocet): guided paradigm — visual cue (arrow/icon), imagine movement for 4 seconds, rest for 3 seconds. 5 minutes per class. Avocet quality scorer flags epochs with eye blink or high-frequency artifact contamination.

BSL boundary: CSP filter computation and LDA classifier are BSL. brainflow bindings are MIT.

See avocet issue for EEG epoch labeling pipeline.

Phase C learned classifier for EEG motor imagery. Classifies 4 classes: left hand imagine, right hand imagine, feet imagine, rest. **Feature extraction:** - Band power per channel: delta (1–4Hz), theta (4–8Hz), alpha (8–13Hz), beta (13–30Hz) — via `scipy.signal` - Common Spatial Pattern (CSP) filters: maximizes variance for one class vs. rest - ICA (Independent Component Analysis) artifact removal: eye blink (delta burst in frontal channels), EMG contamination (broadband high-freq) **Classifier:** CSP + LDA (fast, interpretable, trains in seconds). Optional: small CNN on raw EEG windows for higher accuracy. **Training session UI (in Avocet):** guided paradigm — visual cue (arrow/icon), imagine movement for 4 seconds, rest for 3 seconds. 5 minutes per class. Avocet quality scorer flags epochs with eye blink or high-frequency artifact contamination. **BSL boundary:** CSP filter computation and LDA classifier are BSL. brainflow bindings are MIT. See avocet issue for EEG epoch labeling pipeline.