A Boring Baseline That Is Hard to Beat (CSP + LDA)
Research · NeuroAI & peopleThe standard motor-imagery pipeline, on one subject, with the honest number — weak, noisy, and still useful as a reference point for everything built on top of it.
On this page
With the signal filtered, the next step is the obvious, unglamorous one: build the standard pipeline, on one subject, and see what number it produces. Not to be impressive — to have an honest reference point before asking anything harder of the data.
The pipeline
Left vs. right imagined fist movement, one subject, three runs concatenated (45 trials total — 23 left, 22 right). Filtered as in the last post, then epoched around each trial’s event marker.
The feature extraction method is CSP — Common Spatial Patterns, the standard spatial filter for motor-imagery EEG. CSP finds linear combinations of channels that maximize the variance difference between two classes; in practice, for left/right hand imagery, it tends to find filters that emphasize the motor cortex areas where the two imagined movements produce the clearest contrast. On top of that: LDA, a simple linear classifier. Lotte et al., 2018
Evaluation: 5-fold cross-validation, CSP fit fresh inside each training fold — never on the full 45 trials before splitting. It’s a small detail, but fitting CSP on everything first would let information about the test trials leak into the spatial filter itself, before the classifier even sees the labels.
The result
Epochs: 45 (left=23, right=22)
Balanced accuracy: 0.535 +/- 0.145
Fold scores: [0.475, 0.300, 0.525, 0.700, 0.675]
0.535 balanced accuracy, barely above the 0.5 chance line, with a standard deviation (0.145) almost as large as the signal itself. One fold scored 0.300 — worse than guessing — and another scored 0.700.
Why I’m not trying to fix this yet
It would be easy to try a different classifier, more CSP components, a different frequency band, and keep tuning until the number looks better. I’m deliberately not doing that here. This baseline exists to be a fixed reference point for comparison, not a target to optimize — and the next posts in this thread need exactly this number, weak as it is, to show what happens when the same pipeline gets evaluated across people instead of within one.
Sources & papers I used while learning this
-
A review of classification algorithms for EEG-based brain-computer interfaces: a 10-year update
The reference for CSP, LDA, and why this particular combination is the standard first thing to try for motor-imagery EEG before reaching for anything more complex.
Read paper