Effects of EEG-data normalization on EEG-Transformer-based motor classification

Abstract The application of Transformer-based architectures to electroencephalography (EEG) motor classification has shown promising results, yet the critical role of data normalization in these models remains underexplored. This study systematically investigates the impact of Z-score normalization on Transformer performance for Motor Imagery (MI) and Motor Execution (ME) classification tasks using the Upper Limb Movement EEG Dataset. Transformer architectures, EEGDeformer, ContraNet, and Conformer, were evaluated across binary and multiclass classification scenarios under different preprocessing and normalization conditions. The results demonstrate that Z-score normalization dramatically improves classification accuracy, with relative performance gains when applied to preprocessed data. Without normalization, models frequently failed to exceed chance-level performance. Critically, Z-score normalization provided minimal benefit when applied to non-preprocessed, artifact-contaminated signals, highlighting a synergistic relationship between artifact removal and normalization. ME tasks benefited more than MI, achieving accuracies of up to 0.89±0.01 in binary classification and 0.66±0.04 on a three-class problem. subject-independent validation confirmed that normalization facilitates learning of generalizable features across individuals. These findings establish that Z-score normalization is not an optional preprocessing step but rather a critical requirement for successful application of Transformer architectures to EEG motor classification, with implications for both computational neuroscience research and practical implementation considerations.