Skip to content
Open access

DC-MASTAE: A Coordinate-Aware Spatio-Temporal Transformer for Multi-Channel EEG Denoising in BCI Applications

2026 · IEEE Access · Vol 14, pp. 158817-158835 · 0 citations · 34 references

Abstract

Electroencephalography (EEG) artifact removal remains a significant challenge in the reliable deployment of brain-computer interfaces (BCIs). BCI frameworks, such as motor imagery, SSVEP, and P300, rely on the accurate preservation of cortical rhythms for consistent decoding. Traditional deep learning based EEG denoisers are usually trained on within dataset linear mixtures and are restricted to fixed electrode configurations. These models require retraining for new montages because they treat EEG channels as unordered sets and ignore the spatial geometry of the scalp. Thus, when moving from a single-channel to a multi-channel input, most methods suffer a substantial drop in correlation coefficient (CC) value. To address these challenges, this study presents the DC-MASTAE, a coordinate-aware spatio-temporal transformer for multi-channel EEG denoising. The proposed work introduces an attention mechanism that is permutation-invariant across electrodes while being conditioned on their 3D-Cartesian coordinates via coordinate positional encoding and per-head DistanceBias. This design allows a single trained model to be deployed across different channel montages by simply substituting the coordinate table. Furthermore, a cross subject self-supervised injection process and domain decoupled spectral consistency loss enable training on unlabelled clinical EEG data without requiring paired ground truth. Experimental results show that the proposed DC-MASTAE reduces the single to multi-channel performance drop from ~0.14-0.35 CC to 0.036 CC while preserving cortical task features across motor-imagery, SSVEP, and P300 paradigms at GPU real-time inference.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.