Parkinson's EEG microstates fail to travel: cohort origin is decodable at AUC 0.966 in the same feature space
Synopsis
Expanding resting-state EEG A/B/C/D microstate sequences into static, directed-transition, flow and complexity features separated Parkinson's disease from controls within a development cohort at ROC-AUC 0.823, but locked testing on an independent OpenNeuro cohort fell to 0.513, near chance. In that same feature space, a classifier predicting dataset origin reached ROC-AUC 0.966, showing the two cohorts' feature distributions differ sharply. The study therefore evaluates within-cohort discrimination and cross-cohort transportability separately, and argues for explicit distribution-shift checks before microstate features are treated as reproducible electrophysiological markers.
Interpretation
Resting-state EEG microstate sequences are decomposed into four complementary feature families describing Parkinson's-related brain-state dynamics, rather than counting only how much time each state occupies. Conventional descriptors are static quantities such as duration, occurrence and coverage, which capture how strongly an individual state is expressed. On a development cohort of 100 Parkinson's patients and 49 healthy controls, random forest reached repeated cross-validation ROC-AUC 0.823 and balanced accuracy 0.734, the highest discrimination among the evaluated classifiers.
Discrimination weakens as validation becomes stricter and ends near chance on the independent cohort. Within-cohort cross-validation had served as the estimate of signal strength, although template estimation and preprocessing can make that estimate optimistic. Leakage-aware evaluation with templates re-estimated inside each training fold gave ROC-AUC 0.650, and locked testing that re-estimated no template or model on external labels gave random forest ROC-AUC 0.513 and balanced accuracy 0.520 on the 25 patients and 25 controls of ds003490.
The performance drop coincides with cohort structure inside the feature space itself, placing the issue in feature distributions rather than model selection alone. If external failure only reflected weak disease signal, the same feature space should not also encode cohort membership so strongly. A dataset-origin classifier trained on the same multi-family microstate features separated development from external participants at ROC-AUC 0.966, and the first two principal components explained 42.7% and 18.9% of variance.
Perspective
The findings rest on two public OpenNeuro datasets: development cohort ds004584 with 100 Parkinson's patients and 49 healthy controls, and locked external cohort ds003490 with 25 patients and 25 controls, all standardized to a 120-second analysis window. Classifier performance is framed as a marker-evaluation metric, not clinical diagnostic accuracy.
Acquisition protocol, channel configuration, medication state, cohort composition and preprocessing compatibility vary together between the two datasets, so the shift cannot be attributed to a single source. On cognition, the Parkinson's group had lower MoCA scores, within-group Spearman correlations between MoCA and leading features were weak with intervals crossing zero, and restricting patients to MoCA ≥21 or ≥23 did not restore discrimination, leaving the residual role of cognitive heterogeneity uncertain.