Work

Cross-release transfer of Riemannian EEG pipelines: nearly lossless, but no robust gain from domain adjustment

Year: 2026
Type: preprint
Abstract: Whether resting-state electroencephalography (EEG) pipelines developed within one dataset transfer across independently collected data batches is rarely tested, although developmental cohorts are assembled from multiple releases. We evaluated a filter-band tangent-space pipeline on n = 2,525 participants (ages 5–21) from ten releases of the Healthy Brain Network (HBN)–EEG dataset, using leave-one-release-out (LORO) generalization with ridge regression on log-Euclidean and Riemannian tangent-space features (225 and 11,545) under five domain-handling conditions. Brain-age prediction generalized with almost no loss: LORO R² was 0.596 (log-Euclidean) and 0.397 (Riemannian) versus pooled upper bounds of 0.602 and 0.400, with positive R² in all 10/10 held-out releases. Domain adjustment gave no robust gain: paired δR² depended on representation, with rotation-based alignment gaining + 0.111 under Riemannian features but losing − 0.062 under log-Euclidean features; re-centering left Pearson r exactly unchanged (δr ≡ 0) while producing small nonzero δR². Four phenotype tasks showed no detectable signal (all |r| ≤ 0.194 versus an 80%-power minimum detectable effect of 0.248), and 9/10 representation × condition combinations were equivalent within ± 0.02. Cross-release generalization is thus nearly lossless, whereas harmonization is not a reliable improvement; the phenotype nulls are reported as undetectable rather than absent.
Language: English
Open Access status: green
OpenAlex ID: W7226034385
Cites: 13
Cited by:
Related to: 10