Cross-Dataset Generalization of Quantitative EEG Spectral Biomarkers for Alzheimer’s Disease: A Pre-registered Two-Cohort External Validation and What Does Not Transfer

<title>Abstract</title> <p>Most machine-learning work on EEG biomarkers for Alzheimer’s disease (AD) is trained and evaluated on a single acquisition setup, which leaves a basic question unanswered: when the equipment or the montage changes, do the markers still hold? We put that question to the test. The features are three, all spectral. Leading the set is the theta/alpha power ratio (TAR), with relative band power and alpha peak frequency behind it, and the trio is validated across two independent public cohorts recorded on different devices for separating AD from cognitively normal (CN) participants. The design was pre-registered before any analysis. AHEPA contributed 36 AD and 29 CN, Meghdadi 26 AD and 55 CN. One cohort at a time these features separate the groups cleanly under leave-one-subject-out validation, reaching roughly 0.78 accuracy and 0.80 AUC, and the permutation null never matches the observed value. Cross the acquisition boundary and the picture turns fragile, and uneven. Trained on Meghdadi and asked about AHEPA, the classifier reaches a reasonable 0.73 AUC. Go the other way, train on AHEPA and test on Meghdadi, and the same approach lands at 0.33 AUC, below chance. Class imbalance does not explain that, and neither does a single scaling slip. The data point instead to a domain shift in the absolute scale of the features, and inside that shift the relative ratio, the TAR, keeps working, 0.52 in the weak direction and 0.72 in the strong one. The honest takeaway is a measured statement of how far qEEG markers generalize from one acquisition to another, plus a concrete suggestion: favor relative measures for work that aims to stand independent of the montage.</p>