Title: ERP-XTTN: Leveraging Interpretable Prototype-Guided Cross-Attention for Cross-Subject ERP Classification

Abstract: Developing interpretable brain-computer interface (BCI) classifiers that generalize across different subjects without requiring recalibration remains a significant unresolved challenge. This study investigates whether prototype-based cross-attention mechanisms can deliver competitive and interpretable event-related potential (ERP) classification within deployment-ready constraints. We introduce ERP-XTTN, a novel cross-attention architecture that directs input EEG patches toward fixed difference-wave prototypes. This process utilizes query-key-only cross-attention, omitting value projection, which ensures that classification relies exclusively on attention routing and guarantees structural rather than post-hoc attention faithfulness. The prototypes are generated automatically from the extrema identified in the training-fold difference wave.

Our evaluation spans three public datasets—BNCI Horizon 2020, HRI Cursor, and ERP CORE—covering eight distinct ERP components: ERN, LRP, ErrP, N170, P300, N2pc, MMN, and N400. The assessment employed a leave-one-subject-out (LOSO) protocol with causal filtering, testing performance at both 3-channel and full-montage configurations. We compared ERP-XTTN against EEGNet and xDAWN combined with Riemannian geometry (xDAWN+RG). The performance gap between the top-performing baseline and ERP-XTTN was minimal, measuring .018 AUROC at 3 channels and .034 at full montage. This slight disparity stems from two primary factors: a temporal-flexibility penalty relative to EEGNet and a spatial-exploitation penalty relative to xDAWN+RG, with the latter influenced by signal-to-noise ratios in full-montage settings.

In addition to accuracy metrics, the transparent routing mechanism of ERP-XTTN exposes cross-subject signal structures that opaque, black-box models fail to reveal. Notably, false positives exhibited greater similarity to true positives than to true negatives, suggesting that classification errors are grounded in neurophysiological realities. ERP-XTTN demonstrates robust generalization across diverse ERPs under causal, calibration-free conditions, incurring only a minor interpretability trade-off at minimal channel counts. To our knowledge, this work presents the first epoch-level LOSO benchmark on the ERP CORE dataset.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC