Answers are grounded in this paper and cite their sources.
Background/Objectives: Participant-disjoint evaluation prevents electroencephalography (EEG) windows from crossing training and test partitions, but cannot remove cohort-wide technical variables aligned with outcome. We examined this distinction in the public Mumtaz/HUSM depression EEG archive. Methods: After excluding duplicate groups and performing fixed signal quality controlling, 53 participants were retained. Five whole-head relative-power features were evaluated using nested participant-disjoint class-balanced logistic regression with training-fold local scaling and selection. Results: In this exposed cohort, the model achieved mean balanced accuracy of 0.9158 and ROC-AUC of 0.9363; tested graph-recurrent and domain adaptation models were weaker. A subsequent 1000-shuffle nested permutation audit conditional on the selected five-band pipeline gave p=0.000999. Exploratory controls using file size, raw channel count, and normalized byte rates each achieved balanced accuracy of 0.9046, while duration alone achieved 0.6788. Raw European Data Format (EDF) channel configuration nearly determined archive label: 22/26 controls had 22-channel files, compared to only 1/27 depression-labeled participants (Fisher’s exact p=6.80×10−10). Conclusions: These metadata expose a label-aligned acquisition/export structure, showing that strong participant-disjoint performance and permutation significance do not establish depression-specific EEG information when acquisition format and archive label are closely aligned. Therefore, public EEG studies should combine participant isolation with early metadata audits, and should require outcome variation within acquisition strata prior to physiological or clinical interpretation.
No takes yet. Share an insight, caveat, or question.
Yuen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: