ThinkSTra: a transformer-driven architecture for decoding imagined speech from EEG with spatial-temporal dynamics

Brain-Computer Interfaces (BCIs) enable direct communication between the brain and external devices without requiring physical movement, offering a transformative solution particularly for individuals with impaired or lost motor functions. By providing an alternative communication pathway, BCIs hold considerable promise for both clinical interventions and cognitive neuroscience research. In this study, we introduce ThinkSTra, a novel Transformer-based framework for classifying inner speech commands from electroencephalography (EEG) signals. Unlike conventional models, ThinkSTra jointly captures the temporal dynamics and spatial distributions of neural activity, thereby enabling a more comprehensive representation of the complex structure inherent in EEG signals. We systematically evaluated ThinkSTra on multiple datasets, including the sentence-level TSEEG dataset and the Kumar EEG datasets encompassing character, digit, and visual object classification. To rigorously examine its robustness and generalizability, we additionally performed region- and channel-wise contribution analyses, conducted pretraining and cross-validation experiments, and visualized the learned feature representations using t-SNE. ThinkSTra consistently surpassed existing state-of-the-art approaches, achieving accuracies of 100% on sentence-level, 98.10% on character recognition, 98.34% on digit classification, and 99.5% on visual object tasks. Overall, this study advances inner speech decoding by introducing a robust Transformer-based framework and uncovering how distinct cortical regions contribute to this process, offering both methodological and neuroscientific insights for future brain-computer interfaces.