Cross-Cohort Transportability of Multi-Family Resting-State EEG Microstate Dynamics in Parkinson's Disease

Resting-state electroencephalography (EEG) microstates provide a millisecond-scale representation of large-scale brain-state dynamics, but the cross-cohort stability of Parkinson's disease (PD)-related microstate signatures remains uncertain. We tested whether a multi-family representation of canonical A/B/C/D microstate sequences---integrating static descriptors, directed transitions, flow-related measures, and sequence complexity---retains disease-related information across increasingly stringent validation levels. The development analysis used repeated stratified cross-validation, followed by leakage-aware fold-wise template estimation and locked external testing on an independent OpenNeuro cohort. Random forest achieved a development-cohort ROC-AUC of 0.823, whereas performance fell to near chance during external testing (ROC-AUC = 0.513). In the same feature space, a dataset-origin classifier achieved ROC-AUC = 0.966, demonstrating substantial cross-cohort distribution shift. These findings show that resting-state microstate dynamics contain PD-related information within a cohort, but that the present representation is not sufficiently stable across cohorts. Cross-cohort validation and explicit assessment of dataset shift are therefore essential when evaluating EEG microstate features as reproducible electrophysiological markers.