Generalizable EEG diagnosis via microstate analysis on the TUAB corpus

Automated EEG interpretation is essential for clinical neurology but is currently hindered by the absence of robust, generalizable biomarkers. While EEG microstates offer insights into brain dynamics, their large-scale clinical utility remains largely unvalidated. In this study, we employed a hierarchical two-stage clustering approach to extract robust microstate features from 2,994 clinical recordings within the TUAB Corpus. A suite of machine learning classifiers was trained for automated abnormality detection, with their decisions elucidated via SHAP analysis. The Support Vector Classifier (SVC) yielded the superior performance with an AUROC of 0.877 [95% CI: 0.833–0.917], consistently outperforming other architectures like MLP and Logistic Regression. Statistical analysis identified 16 discriminative microstate features (p $$\:<\:$$ 0.0022, Bonferroni-corrected), with six exhibiting large effect sizes (|Cohen’s d| $$\:>$$ 0.8), most notably in microstate duration variability and A-D state transition probabilities. Quantitative template comparisons confirmed that discriminative power arises from both topographical reconfiguration (spatial) and sequence fragmentation (temporal). Specifically, the abnormal group displayed a breakdown in temporal stability, characterized by increased duration variability and disrupted bidirectional flow between canonical states. These findings validate microstate dynamics as powerful, interpretable biomarkers, providing a scalable, data-driven framework for clinical EEG diagnosis.