Band-Resolved EEG Brain Age as a Disease-State Marker

Doctors would like one number from a resting EEG, a scalp recording of the brain's electrical rhythms, that says whether a person's brain looks older than it should. The usual version of that number is the "brain-age gap": train a model to predict chronological age from healthy recordings, then measure how far a patient's prediction deviates from their true age. For EEG, the current best evidence says this broadband gap mostly measures who someone is rather than whether they are ill. When patients in a large clinical archive moved between normal and abnormal recordings, their gap did not move, and classifying those transitions from the gap was at or below chance (Brain age revisited).

We propose that this conclusion comes from collapsing the EEG into one age estimate. It does not show that EEG age markers are uninformative. Pathology and normal aging act on different rhythms, and they sometimes act in opposite directions. The same kind of cancellation shows up in motor tasks, where low-frequency power falls while high-frequency power rises (Spectral Changes in Cortical Surface Potentials during Motor Movement). If one age gap is estimated separately inside each frequency band, the result is a short, readable profile: "your alpha rhythm looks eight years old, your theta rhythm looks normal." Such a profile could carry disease-state information that a single number averages away. It would also be interpretable by construction, because each entry of the profile is defined as a statement about one rhythm.

Core Claim

A vector of per-band brain-age gaps behaves differently from the broadband gap. The broadband gap acts as a stable personal trait. We claim the band profile has at least one band whose gap shifts when a person transitions between health and pathology, and that the size of this shift depends on the person's age.

Two design choices carry the claim. The first is anchoring each band to the person's own alpha peak rather than to fixed textbook boundaries. With fixed bands, normal age-related slowing of the alpha rhythm would leak into the "theta" and "alpha" estimates and look like disease. The second is choosing how much to regularize each band model by its disease effect size rather than by its age-prediction accuracy.

The second choice follows from the "brain-age paradox" found in structural MRI. There, more accurate age predictors were worse disease detectors, and heavily regularized linear models gave the largest patient–control effect sizes (Brain-age models with lower age prediction accuracy have higher sensitivity for disease detection). That paper, and the research file behind this note, state that no one has tested the paradox for EEG.

Concretely, for band $b$ and subject $i$ with chronological age $a_i$, we define the gap

$$g_{i,b} = \hat a_b(x_{i,b}) - a_i - c_b(a_i),$$

where $x_{i,b}$ are subject $i$'s features restricted to band $b$ and $\hat a_b$ is an age regressor trained only on healthy recordings. The term $c_b(a)$ is the usual age-dependent bias correction, fit on held-out healthy data so that $g_{i,b}$ has zero mean at every age in health. The five-dimensional vector $g_i$, one entry per band, is the object we study.

Research Program

The normative models are trained on the non-pathological recordings of the Temple University Hospital Abnormal EEG corpus (TUAB). We add the LEMON and CHBP EEG cohorts, using the harmonized preprocessing of A reusable benchmark of brain-age prediction from M/EEG resting-state signals. Cam-CAN, which is MEG rather than EEG, is used only as a replication cohort.

Within each band, features are the sensor covariance matrices mapped into a flat space with Riemannian geometry, a way of comparing covariance matrices that respects their curved structure. That benchmark reports these features among its strongest, and they partially undo the spatial blurring of scalp signals. We also extract log band power and coherence. The age regressor is ridge regression. A per-band small convolutional network serves as a higher-capacity comparison.

Band edges come from each recording's individual alpha peak. We locate the peak after removing the 1/f background trend that Let EEG Models Learn EEG identifies as a defining property of EEG spectra, then place band edges at fixed ratios of that peak. For recordings with no detectable peak, we fall back to canonical bands and flag them.

The program then asks three linked questions.

The first is whether the paradox holds in EEG. We sweep ridge strength and network capacity within each band. For each setting we record age mean absolute error on healthy held-out data alongside the pathology effect size (Cohen's d) on TUAB.

The second is state versus trait. We reuse the four longitudinal groupings of Brain age revisited: repeated non-pathological, repeated pathological, normal-to-pathological, and pathological-to-normal. For each band we test whether the within-person change in $g_{i,b}$ follows the direction of the change in pathology label.

The third is age dependence. We estimate each band's pathology effect size within age deciles to check whether diagnostic weight passes from one rhythm to another across the lifespan. The lifespan brain-age work in Rate of brain aging associates with future executive function in Asian children and older adults shows the direction of this kind of question: the same accelerated-aging signal meant maturation in children and decline in older adults.

Finally, we carry the frozen band-gap vectors, without retraining, to the AHEPA dementia cohort of 88 subjects (Alzheimer's disease, frontotemporal dementia, healthy controls). We ask whether a five-number, label-free profile separates these groups. We also compare its band ordering with the alpha/theta emphasis reported by Frequency bands EEG Biomarkers for Dementia using Graph Neural Networks and the alpha-band connectivity finding of Flexible and Explainable Graph Analysis for EEG-based Alzheimer's Disease Classification.

This design is kept deliberately simple. The per-band ridge gap is little more than a normative deviation score. The contribution is not a new architecture. It is showing, or refuting, that band resolution, individual-peak alignment and effect-size-based regularization together turn EEG brain age from a trait into a state marker.

Baselines

Naive reference. Chronological age alone, plus the TUAB pathology base rate. Any gap measure that does not beat a logistic model on age for pathology prediction carries no information.

Broadband brain-age gap. The TCN model of Brain age revisited is the direct competitor on the state-versus-trait question. It reports age MAE of 6.6 years and R² of 0.73 on non-pathological TUAB. Its gap-based pathology classification reaches balanced accuracy of 61.12% on TUAB, 44.74% on normal-to-pathological transitions and 47.79% on pathological-to-normal transitions. We reproduce it on our exact splits. We also build a broadband Riemannian ridge gap from the same pipeline as our band models, so that the only difference is band resolution. Beating it means at least one band's within-person gap change predicts transition direction above 60% balanced accuracy, with a confidence interval excluding 50%, on the transition groups. Our band models must also be tuned with the same hyperparameter budget and the same healthy-only training data.

Strongest age predictor. The spatially aware Riemannian and ConvNet models of A reusable benchmark of brain-age prediction from M/EEG resting-state signals report R² between 0.60 and 0.71 on TUAB and Cam-CAN. We do not expect to beat them on age accuracy, and the paradox hypothesis predicts we should not need to. The test is whether our effect-size-tuned models reach larger pathology effect sizes at worse age accuracy.

Direct supervised classifier. This is an upper reference for pathology detection, not a competitor on label-free state tracking. On TUAB, Brain age revisited reports at least 86% balanced accuracy for direct EEG pathology classification.

On AHEPA, the strongest reported numbers are 91.0% three-way accuracy on 2-second segments from EEG-SSM and 76.06% under leave-one-subject-out from the graph transformer of Frequency bands EEG Biomarkers for Dementia using Graph Neural Networks. The segment-level figure may mix segments from the same subject across splits, so we compare only under leave-one-subject-out.

Foundation-model reference. LEAD reports 100% subject-level accuracy on three of its five downstream AD datasets. It gives no single AHEPA leave-one-subject-out figure in the research file, so we fine-tune it ourselves under our splits.

Ablations. Two ablations isolate the proposal's ingredients. Canonical fixed bands versus alpha-peak-anchored bands isolates the aging confound. Accuracy-tuned versus effect-size-tuned regularization, both chosen on a disjoint validation split, isolates the paradox.

Why not just…?

Why not just regress band power on age and use the z-scores? This is the most serious alternative. It is close to our ridge gap: a normative deviation per band without the "brain age" wrapper. We include it as a variant rather than argue against it. The scientific claim, that band-resolved normative deviation is state-like where the broadband gap is trait-like, stands whichever variant wins. If raw band-power z-scores match the Riemannian band gaps, we will report that the simpler measure suffices. The only thing we expect it to lose is spatial information that covariance features carry.

Why not just fine-tune a foundation model on disease labels and read saliency maps? For raw detection this will win. Direct classification on TUAB already exceeds 86% balanced accuracy. But such a model is trained against labels, so it cannot tell a benign older-than-average rhythm from pathology. Its saliency maps are post-hoc. Brain age revisited used exactly such gradients to attribute young predictions to occipital beta, without being able to test that attribution. Our per-band models make the band attribution a property of the model itself. We test this by checking that deleting a band from the input changes only that band's gap. We also compare how stable the band attribution of a fine-tuned CodeBrain or LEAD model is across random seeds.

Why not just build a band-decomposed unrolled signed-graph network? Adding frequency bands to the interpretable graph-denoising model of Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling is the obvious move at this intersection. But it addresses classifier transparency, not whether an age-normed marker tracks disease state, and it adds capacity that the paradox predicts could hurt sensitivity. If our ridge models show the effect, such a network becomes a candidate backbone to test later, not a prerequisite.

Proposed Experiments

The first decisive check replicates the trait finding and then breaks it. On the four longitudinal TUAB groupings, the broadband ridge gap should reproduce near-chance transition classification. At least one alpha-anchored band gap should shift with pathology direction. If no band shifts, the trait conclusion extends to band-resolved EEG, and that is itself worth reporting.

The second check traces, per band, the curve of age MAE against pathology Cohen's d across regularization strengths. If the paradox holds in EEG, the peak effect size will fall at weaker age accuracy than the accuracy-optimal setting, as in the MRI result.

The third check reports per-decile effect sizes with bootstrap intervals. Under fixed bands, we expect an apparent rise in low-frequency pathology signal in older deciles that shrinks under peak-anchored bands. That would show age-related slowing was being counted as disease.

The fourth check transfers the frozen models to AHEPA under leave-one-subject-out. We report how well the five-number profile, fed to a logistic model, performs against the 76.06% graph-transformer figure and our re-run foundation model, along with which bands carry the separation between Alzheimer's disease and frontotemporal dementia.

A last robustness check reruns the effect-size analyses with synthetic Gaussian and DC-offset noise at graded levels, mirroring EEG2Rep. It asks whether heavy regularization's advantage grows as signal-to-noise falls, which is the mechanism we propose for the paradox in low-signal EEG.