Geometry-Conditioned Hyperalignment for Generalizable EEG Decoding Across Datasets and Paradigms

hTSMNet achieves strict zero-shot generalizable EEG decoding on a leave-one-dataset-out left-/right-hand motor-imagery benchmark across three datasets (172 subjects), matching the strongest comparator.

• Why it matters: Cross-subject and cross-dataset EEG decoding is limited because different electrode montages and subject-specific leadfields create non-overlapping observation spaces and residual distribution shifts.
• What they did: The method uses a geometry-conditioned hyperalignment frontend that maps any montage to fixed scalp anchors from electrode coordinates, followed by an SPD-manifold backbone; it is evaluated with leave-one-dataset-out transfer and eight comparators.
• The result: With only final-layer calibration, it reaches the highest accuracy on every dataset, and a frozen MI-pretrained model transfers across paradigms for mental workload and SSVEP above chance while staying at chance on stimulus-locked ERP.

AI summary of the abstract · may contain errors · not medical advice · read the paper.