LIBRARYTurning EEG Into a Command: CSP & EEGNet.

How a noisy EEG window becomes one of a few commands: the classic Common Spatial Patterns + LDA pipeline, and the deep-learning alternative, EEGNet.

The last stage of a EEG BCI turns a window of signal into a decision, one of a small set of commands. The standard pipeline is a chain: cut the signal into epochs, bandpass to the task's frequencies, extract features (or spatially filter), then hand those to a classifier that outputs a class. Two eras of this exist side by side: the classic hand-engineered approach, and end-to-end deep learning.

Fig 1

Common Spatial Patterns (CSP), plainly

CSP is a supervised, two-class spatial filter (e.g. left vs right motor imagery). It learns combinations of electrodes that MAXIMIZE the band-power for one class while MINIMIZING it for the other, turning a spatial difference into an easily separated number. The usual feature is the log-variance of the CSP-filtered signals, fed to a simple linear classifier (LDA). It fits motor imagery beautifully, because the signal there IS a spatial band-power difference (Ramoser et al., 2000).

CSP's main limit is that it works on one chosen frequency band, so the popular Filter-Bank CSP (FBCSP) splits the EEG into sub-bands and automatically picks the discriminative ones per person. CSP is also supervised (it needs labeled calibration data), sensitive to noise and small sample sizes, and natively two-class; multi-class needs extensions.

EEGNet: the deep-learning route

The deep-learning route replaces the hand-built features with a network that learns them. EEGNet (Lawhern et al., 2018) is a deliberately compact convolutional neural net for EEG: a temporal convolution learns frequency filters, a depthwise convolution learns spatial filters per frequency (the network's own analog of CSP), a separable convolution summarizes, and a small layer classifies. It was shown to generalize across several BCI paradigms (P300, error-related negativity, movement-related potentials, and motor imagery) with few parameters.

CSP + LDA vs EEGNet

	CSP + LDA	EEGNet
Feature engineering	Hand-designed (bandpass + spatial filter)	Learned end-to-end
Data appetite	Low, works with little data	Higher (but EEGNet is designed to be data-efficient)
Interpretability	High (you can read the filters)	Lower (though it ships visualization tools)
Best for	2-class motor imagery, a strong baseline	Cross-paradigm, more data available

Maker reality

EEG datasets are small, brains differ, and the signal drifts between sessions, so expect a per-subject calibration step (often 15–30 minutes of recording) and some retraining over time. Start with CSP + LDA: it's fast, low-data, interpretable, and a strong baseline that EEGNet itself is measured against. Reach for EEGNet or FBCSP when you have more data or need more than two classes.

Checkpoint

Quick check

What is the job of the final classification stage of an EEG BCI?