PRSEPTransformer-EEG: source-space sLORETA and residual transformers for robust motor imagery and execution EEG decoding

Abstract

Decoding motor intentions from non-invasive scalp EEG remains a core challenge in brain–computer interface research, due to spatial blurring, low signal-to-noise ratio, and inter-subject variability. This work introduces PRSEPTransformer-EEG, a unified framework that integrates standardized preprocessing, source localization using sLORETA, and a deep learning model that combines squeeze–excitation residual blocks with a positional-encoding Transformer encoder. The framework is evaluated across three public datasets, encompassing both motor imagery and motor execution paradigms recorded via gel-based, water-based, and dry electrodes. On the BCI Competition IV 2a dataset, the model achieves 99.53 % accuracy in the Beta band, surpassing the best previously reported results. On the BCI2000 four-class imagery dataset, it reaches 98.29 % in the Beta band and exceeds 92 % across all typical frequency ranges, outperforming existing source and sensor space models. Results from the reach-and-grasp dataset further confirm that source-space decoding offers substantial gains when high-density recordings are available, while sensor-level features remain effective in sparse configurations. These findings demonstrate that combining source-space projection with temporally attentive spatial encoding can substantially advance the reliability of non-invasive EEG decoding for both clinical and assistive applications. Our code is available at (https://github.com/sinamakhdoomi/PRSEPTransfromerEEG.git).