Enhancing cross-dataset EEG microstate decoding: A frequency-band dependent framework with bidirectional long short-term memory network

Electroencephalogram (EEG) microstate analysis serves as a critical tool for investigating dynamic brain networks. However, limited robustness in cross-dataset applications significantly constrains model generalizability and clinical utility. This issue arises because microstates in different frequency bands exhibit distinct responses to data heterogeneity, yet this critical factor lacks systematic investigation and quantification, resulting in an absence of theoretical guidance for optimal frequency band selection in fields such as brain-computer interfaces. To address this gap, this study proposes a novel deep learning evaluation framework, utilizing standardized data preprocessing and multi-model comparisons to systematically investigate the topological stability and classification performance of microstates in the theta, alpha, and beta frequency bands across datasets for the first time. We integrated three independent, publicly available EEG datasets and employed a sequence-to-sequence model based on bidirectional long short-term memory networks to achieve efficient cross-dataset classification of four canonical microstates. Our key findings demonstrate that alpha-band microstates exhibit exceptional cross-dataset robustness, characterized by highly stable topological structures (within-class correlation ranging from 0.894 to 0.933) and cross-subject classification accuracy exceeding 91%. In the most challenging cross-dataset validation scenario, the classification accuracy of alpha-band microstates reached a peak of 85.12%, significantly outperforming theta (62.85%) and beta (79.04%) bands. Compared to state-of-the-art methods, our approach improved cross-dataset classification performance by 16.93%. In sum, our study quantifies the stability advantage of alpha-band microstates for the first time, providing a robust theoretical and methodological foundation for frequency band selection strategies, with direct implications for brain-computer interfaces and closed-loop neuromodulation systems.