Lite-EEGTrans: An Interpretable Real-Time Motor Imagery EEG Denoising and Classification for Embedded Robotic BCI

Abstract

Real-time brain-computer interface (BCI) systems for motor imagery face persistent challenges due to EEG artifacts, limited generalization across datasets, and strict latency constraints in embedded robotic applications. This work presents Lite-EEGTrans, a lightweight and interpretable EEG denoising and classification pipeline designed for real-time motor imagery decoding in assistive robotic control for stroke rehabilitation. We systematically evaluate multiple pretrained deep learning-based EEG denoisers, originally trained on a large benchmark dataset, for their ability to generalize to an independent acute stroke motor imagery dataset. Based on accuracy-latency trade-offs, a dual-pathway 1D-CNN denoiser is identified as the most effective candidate. We further propose a compact transformer classifier with multi-head attention explicitly designed to capture motorrelated beta-band burst activity, providing neurophysiologically meaningful interpretability. The combined denoiser-transformer pipeline achieves a mean validation accuracy of 74.46%, high recall (94.73%), and low inference latency ($7-8 ~\text{ms}$ per chunk) on commodity hardware, outperforming or matching state-of-theart lightweight CNN and statistical baselines. Attention analysis confirms consistent allocation to beta-band oscillatory patterns associated with motor imagery, supporting the physiological validity of the model. These results demonstrate the feasibility of deploying interpretable, low-latency transformer-based EEG decoding pipelines for embedded and real-time BCI-driven robotic rehabilitation systems.