Exploration on the improvement of EEG dataset collection and preprocessing methods based on neuroscience theory

This paper analyzes current methods of EEG dataset collection and preprocessing within the framework of neuroscientific theories and proposes potential improvements. Existing datasets are limited because they primarily capture specific stages of neural activity during continuous recitation, missing other crucial stages of language generation. This limitation impacts the performance of deep learning models used for EEG data classification. By examining the multi-stage process of language formation, the paper highlights the need for data collection methods that encompass all stagesconceptualization, grammatical encoding, lexicalization, morphological and phonological encoding, and articulation. The review also discusses the benefits of using nonlinear dynamic system analysis methods over traditional covariance or cross-correlation matrices to better capture the complexities of EEG data. Finally, it suggests treating EEG data at various stages of language formation as multimodal data, as a strategy to improve the comprehensiveness and practicality of EEG datasets, thereby enhancing the accuracy and practical application of EEG-based brain-computer interfaces.