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 58-feature pool (“Config J”) drawn from rest- and task-state recordings achieves 89.6% balanced accuracy for HC-vs-AD and 80.5% for HC-vs-FTD (AUC = 0.951 and 0.823, respectively; n = 65 and n = 52). Adding the task-evoked features to the rest-only baseline contributes a substantial AD uplift and a 12-point FTD uplift. A pre-specified sensitivity analysis adjusting for linear sex-based differences within each training fold drops HC-vs-AD by 2.1 points to 87.6% and leaves HC-vs-FTD unchanged at 80.5%, well above the demographics-only floor. Config J was selected post hoc on this cohort; multi-site, sex-balanced replication is required before generalization claims. The pipeline applies an 80%-coverage filter that drops 12 sparse photic-driving features (present in only 1 to 6 of 88 subjects) before classification; demographic columns (age, sex, MMSE) are excluded from the classifier pool to prevent leakage. Both rest (ds004504) and task (ds006036) features are produced by the unified canonical ingest layer of the framework series (extract_subject_features in the shared _framework_core module), with cryptographic pipeline-version stamping in every output JSON to prevent silent stale-cache drift.