Motor imagery and SSVEP BCIs

"They asked me to read a mind. I settled for reading a mind that was busy imagining its own left hand, which turned out to be a far more cooperative subject."

A Pragmatic AI Agent

Motor imagery: reading the sensorimotor idle

What is the signal? Over the hand area of motor cortex, neurons at rest oscillate together in the mu (roughly 8 to 12 Hz) and beta (13 to 30 Hz) bands. When you plan, execute, or merely imagine a movement, that local population desynchronizes and the band power drops. This is event-related desynchronization (ERD); the rebound after the imagined movement is event-related synchronization (ERS). Crucially the effect is somatotopic and contralateral: imagined left-hand movement suppresses mu power at the right-hemisphere electrode (C4), and imagined right-hand movement suppresses it at C3. A motor-imagery BCI is, at heart, a detector of where and in which band the power dropped.

Why is raw band power not enough? Because scalp EEG is badly volume-conducted: one cortical source smears across many electrodes, and the ERD at C3 is buried under unrelated activity picked up by the same sensor. The classical breakthrough is Common Spatial Patterns (CSP), a supervised spatial filter that learns linear combinations of channels which maximize the variance (band power) for one class while minimizing it for the other. Formally, given the average per-class spatial covariance matrices \(\Sigma_1\) and \(\Sigma_2\) of band-pass-filtered trials, CSP solves the generalized eigenproblem

and keeps the filters \(\mathbf{w}\) with the most extreme eigenvalues \(\lambda\). The largest \(\lambda\) give filters whose output variance is high for class 1 and low for class 2; the smallest give the mirror image. Project a trial through these filters, take the log-variance of each projection, and you have a compact, discriminative feature vector that a simple linear classifier can separate. Because ERD lives in a specific band, the standard production recipe is Filter Bank CSP (FBCSP): run CSP inside several sub-bands (for example 4 to 8, 8 to 12, ... 32 to 40 Hz), then select the informative band-and-filter combinations. FBCSP won the BCI Competition IV datasets and is still the reference baseline.

SSVEP: when you plant the signal yourself

What is the signal? The visual cortex is a frequency-faithful amplifier. Present a stimulus flickering at frequency \(f\) and the EEG over the occipital pole (Oz, O1, O2) develops a sharp spectral peak at \(f\) and its harmonics \(2f, 3f\). Give the user several on-screen targets, each flickering at a distinct frequency, and their gaze choice is written directly into the spectrum. There is no imagined intent to learn and, for many users, no calibration at all: the decoder just asks which reference frequency the occipital signal correlates with best.

How do we decode it robustly? A naive peak-picking FFT works but wastes the multi-channel structure and the harmonics. The standard method is Canonical Correlation Analysis (CCA). For each candidate frequency \(f_k\) you build a reference matrix \(Y_k\) of sines and cosines at \(f_k\) and its harmonics, then find the linear combinations of the EEG channels \(X\) and of the references \(Y_k\) whose canonical correlation \(\rho_k\) is maximal. The predicted target is \(\arg\max_k \rho_k\). CCA needs no training data because the references are synthetic. Its stronger cousins, Filter Bank CCA (FBCCA) and the calibration-based Task-Related Component Analysis (TRCA), push spellers past 100 to 200 bits per minute, the highest information-transfer rate (ITR) of any non-invasive BCI. That speed is exactly why SSVEP dominates practical spellers even though staring at flicker is tiring and unusable for anyone who cannot direct their gaze.

The listing below is a complete, dependency-light CCA decoder. It generates a short synthetic occipital trial that flickers at one of three target frequencies, then recovers the target purely by canonical correlation against synthetic sine-cosine references, with no training whatsoever.

import numpy as np
from scipy.stats import pearsonr

fs, T = 256, 2.0                       # 256 Hz, 2-second trials
t = np.arange(0, T, 1 / fs)
targets = [9.25, 11.25, 13.25]         # three flicker frequencies (Hz)

def reference(freq, n_harmonics=3):
    """Sine/cosine reference bank at freq and its harmonics."""
    cols = []
    for h in range(1, n_harmonics + 1):
        cols += [np.sin(2 * np.pi * h * freq * t),
                 np.cos(2 * np.pi * h * freq * t)]
    return np.column_stack(cols)       # shape (samples, 2*n_harmonics)

def cca_corr(X, Y):
    """Top canonical correlation between multichannel X and reference Y."""
    Xc, Yc = X - X.mean(0), Y - Y.mean(0)
    # whiten then take largest singular value of the cross-covariance
    Qx, _ = np.linalg.qr(Xc)
    Qy, _ = np.linalg.qr(Yc)
    return np.linalg.svd(Qx.T @ Qy, compute_uv=False)[0]

def decode(eeg):
    rhos = [cca_corr(eeg, reference(f)) for f in targets]
    return int(np.argmax(rhos)), rhos

rng = np.random.default_rng(0)
true_k = 1                             # user is looking at the 11.25 Hz target
clean = np.sin(2 * np.pi * targets[true_k] * t)
eeg = np.column_stack([clean, 0.6 * clean]) + rng.normal(0, 2.0, (len(t), 2))
pred, rhos = decode(eeg)
print("predicted target:", pred, " correlations:", np.round(rhos, 3))

Choosing a paradigm, and paying for it in calibration

When do you pick which? The trade is between autonomy and burden. SSVEP is fast, near-zero-calibration, and robust, but it commandeers the eyes and needs a screen of flickering targets, which rules it out for gaze-impaired users and for eyes-free control. Motor imagery is self-paced, screen-free, and works with the eyes closed, which makes it the paradigm of choice for the most severely paralyzed users and for continuous control such as steering a wheelchair or a cursor; the price is heavy per-subject calibration and a stubborn population of "BCI-illiterate" users (commonly cited around 15 to 30 percent) who never produce a separable ERD. A pragmatic middle ground is the hybrid BCI, which combines an SSVEP "switch" to arm the system with motor imagery for the actual command, or fuses both feature streams.

The evaluation trap. The single most common way to publish an inflated motor-imagery number is to shuffle trials from one recording session into train and test. EEG has slow drifts (electrode impedance, alertness, cap position) so trials minutes apart are correlated; a random split leaks that shared state and flatters the model. The honest protocols are leave-one-session-out and, hardest of all, leave-one-subject-out, the same leakage-safe discipline demanded in Chapter 5. Report the operating point with calibrated confidence rather than a bare accuracy, because a BCI that abstains when unsure is far safer than one that guesses; the calibration and abstention machinery of Chapter 18 applies directly.

What's Next

In Section 31.4, we move from these hand-designed spatial filters to learned ones: EEG foundation models (LaBraM, EEGPT, BIOT, CBraMod) that pretrain on massive unlabeled recordings and then fine-tune to motor imagery, SSVEP, and clinical tasks with a fraction of the per-subject calibration that CSP and TRCA demand.