2.6.2. Spatiotemporal Beamformer Classifier

Before calculating the spatiotemporal beamformer (STBF), the signal was downsampled to 32 Hz or twice the low-pass frequency 16 Hz, resulting in 17 time samples between 0.1 s and 0.6 s. According to the Nyquist theorem, more samples would not contain more information; hence, the minimum temporal resolution was chosen to reduce the dimensionality of the covariance. The activation pattern is the difference between the averages of epochs in response to cued targets and the averages of those in response to non-cued targets. We constructed three variations of the spatiotemporal beamformer: STBF with empirical covariance estimation (STBF-EMP) as in Section 2.3.1, STBF with LOOCV-shrunk covariance estimation (STBF-SHRUNK) as in Section 2.3.2, and STBF with Kronecker–Toeplitz-structured covariance estimation (STBF-STRUCT) with LOOCV shrinkage for the Kronecker factors as in Section 2.3.4.

## 2.6.3. Riemannian Geometry Classifier

We opted for a Riemannian geometry-based classifier to compare our results. The Riemannian model (xDAWN+RG) uses the xDAWN spatial filter combined with Riemannian geometry in tangent space as implemented by Barachant et al. [58]. This classifier uses four xDAWN spatial filters and each epoch's empirical spatial covariance matrix. The target with the maximum score is the prediction of the cued target. xDAWN+RG was trained and validated without downsampling using epochs at the original sample rate of 2048 Hz.
