Hierarchical attention enhanced deep learning achieves high precision motor imagery classification in brain computer interfaces.

Motor imagery-based Brain-Computer Interfaces (BCIs) hold transformative potential for individuals with severe motor impairments, yet their clinical deployment remains constrained by the inherent complexity of electroencephalographic (EEG) signal decoding. This study presents a systematic investigation of hierarchical deep learning architectures for motor imagery classification, introducing a novel attention-enhanced convolutional-recurrent framework that achieves state-of-the-art accuracy of 97.2477% on a custom four-class motor imagery dataset comprising 4,320 trials from 15 participants. By synergistically integrating spatial feature extraction through convolutional layers, temporal dynamics modeling via long short-term memory networks, and selective attention mechanisms for adaptive feature weighting, our approach significantly outperforms conventional methods while providing interpretable insights into the spatiotemporal signatures of motor imagery. Beyond demonstrating competitive performance, this work elucidates the critical role of attention mechanisms in capturing task-relevant neural patterns amidst the high-dimensional, non-stationary nature of EEG signals. Our findings demonstrate that biomimetic computational architectures that mirror the brain's own selective processing strategies can substantially enhance BCI reliability, offering immediate implications for neurorehabilitation technologies and broader applications in restorative neuroscience. Our code is available at https://github.com/Laboratory-EverythingAI/-EEG_Classification .