There is a newer version of the record available.

Published April 22, 2026 | Version v1

Dynamic Signatures of Dementia: A Task-Evoked Five-Dimensional EEG Classifier for Alzheimer's Disease

Authors/Creators

Description

Electroencephalographic (EEG) biomarkers for Alzheimer’s disease (AD) that perform well under rigorous subject-level validation remain elusive; much of the published accuracy literature inflates apparent performance through validation protocols that leak information between training and test subjects. We evaluate an individual-subject EEG classifier on 88 participants (29 healthy controls, 36 AD, 23 frontotemporal dementia; OpenNeuro ds004504 / ds006036), combining resting-state and photic-stimulation task recordings. We use a fully deterministic doubly-nested leave-one-subject-out (LOO-LOO) protocol. Within each fold, standardization, feature ranking, and 𝑘-selection are performed on training data only.

On this cohort, a post-hoc-selected 70-feature pool (“Config J”) drawn from rest- and task-state recordings achieves 91.3% balanced accuracy for HC-vs-AD and 86.1% for HC-vs-FTD. Under a 10,000-permutation LOO-LOO test these yield 𝑝 < 0.0001 (HC-vs-AD; 𝑛 = 65; AUC = 0.952) and 𝑝 = 0.0003 (HC-vs-FTD; 𝑛 = 52; AUC = 0.865). Adding the task-evoked features to the rest-only baseline contributes a 14.9-point AD uplift and a 16.0-point FTD uplift. A pre-specified sensitivity analysis adjusting for linear sex-based differences within each training fold reduces HC-vs-AD to 84.4% and leaves HC-vs-FTD essentially unchanged at 85.7%. The 91.3% / 86.1% numbers should be read as a cohort-specific ceiling: the best performance this pipeline achieves on this cohort under this protocol. They are not a clinical-grade generalization estimate. Config J was selected post hoc on the same cohort, the sample is single-site and cross-sectional, and only linear sex-coupled variance has been characterized; multi-site, sex-balanced replication is the outstanding requirement.

Files

AD_Classification_v1.pdf

Files (132.6 kB)

Name Size
md5:75755bd16cc85332226e1dd42954ca7a
132.6 kB Preview Download

Additional details