A public EEG decoding benchmark
A simple classifier trained on 16 people correctly labeled 71 of 120 left- or right-fist imagery epochs from 8 different people.
· EEG Motor Movement/Imagery Dataset v1.0.0
| Recorded cue | Predicted left fist | Predicted right fist |
|---|---|---|
| Left fist | 40 | 22 |
| Right fist | 27 | 31 |
Accuracy 59.2%. Balanced accuracy 59%. The 95% interval from resampling the eight test participants was 49.8%–68%, which includes the 50% balanced baseline.
Method and limits
We used the first unilateral imagery run (R04) from S001–S016 for training and the next repetition (R08) from S017–S024 for evaluation. T1 and T2 identify the two cued movements. We analyzed two seconds from half a second after each cue, using the difference in log band power between C3 and C4 in 8–12 and 13–26 Hz. Feature normalization and the class centroids used training data only.
Participant and run changes occur together in this split. The test is small and uses cues that can also elicit eye or muscle responses. These recordings provide imagery cues and have no feedback-error labels or dopamine measurements. The next checks are cue and movement artifact controls, a larger held-out sample, and a separate dataset with feedback-error labels.
Reproduce the result
Download RECORDS and the 24 files linked below. Put their basenames in one folder. Compare the SHA-256 values in the result JSON before running the script. The recorded NumPy version is 2.5.4.
python3 audit-eeg.py ./recordsDownload audit scriptResult and input hashesPhysioNet EEGMMIDB
Input files
Data by Gerwin Schalk and colleagues, distributed by PhysioNet under the Open Data Commons Attribution License v1.0.