EEG-Based Speech Imagery Classification via Neural Architecture Search

Abstract

Speech imagery (SI) has emerged as a promising paradigm for brain–computer interfaces (BCIs) that enable hands-free communication without external stimuli. Classification of electroencephalography-based SI electroencephalography signals (SI-EEG) remains challenging due to limited data and high inter- and intra-subject variability. Most existing approaches rely on manually designed architectures, which is a time-consuming and labor-intensive process that requires significant domain expertise. To address these challenges, we propose a novel neural architecture search (NAS) framework for SI-EEG signal classification with bilevel optimization and path-level binarization to automatically and adaptively discover optimal model architectures. Furthermore, we propose Deformable ConvNeXt block, a novel convolutional neural network-based operation that integrates transformer-inspired design principles and EEG-specific priors. This block enables each neuron to adaptively control the size of the EEG signal’s receptive field. It allows the model to more effectively capture temporal variations in mental activity in inter- and intra-subject classification settings. Evaluated across three SI-EEG datasets, our approach achieves classification accuracies of 68.30%, 68.49%, and 72.32%, surpassing the state of the art. These results demonstrate the effectiveness and scalability of our NAS-based framework for next-generation real-time BCI systems.