EEG Signal Classification Reinforced by LDConv and Transformer

EEG-based brain-computer interfaces (BCIs) have garnered significant global research interest in recent years. Commonly adopted paradigms include motor imagery (MI), mental arithmetic (MA), and P300 event-related potentials. To address EEG's limited spatial resolution, this work introduces a novel deep learning architecture termed LDCT. The LDCT framework integrates Linear Deformable Convolution (LDConv) with Transformer modules. LDConv adapts its spatial sampling dynamically, while the Transformer component models global temporal relationships efficiently. The model was validated using a publicly available dataset comprising 29 individuals. It achieved classification accuracies of 92.13% for MI tasks and 91.63% for MA tasks. Compared to existing methods, the proposed model demonstrates superior classification performance.