BrainMetastabilityAnalyzerTool

The group effect survived cleaning. The diagnostic advantage did not.

Φ-Dwell remains an interesting spatial sensor-phase representation of EEG dynamics. But after recomputing the frozen feature on cleaned derivative EEG, ordinary spectral slowing explains the useful AD-versus-control discrimination just as well.

88/88 derivative subjects processeddwell AD/CN p = 0.0062spectral AUC = 0.778spectral + dwell AUC = 0.777increment = −0.001not a clinical diagnostic tool
Cleaned EEG receiptFull auditRepository
What survives

Keep the measurement. Drop the metaphysics.

The graph Laplacian is built from electrode geometry. Its eigenvectors are therefore modes of the sensor layout, not structural-connectome modes and not a holographic reconstruction.

EEG phase at sensors
↓
project onto graph-Laplacian sensor modes
↓
dominant spatial mode in δ / θ / α / β / γ
↓
dwell time · transitions · multi-band states
Keep

Spatial phase modes

A deterministic transform of multichannel phase organization.

Keep

Dwell dynamics

How long each band remains in the same dominant sensor mode.

Frozen candidate

Dwell gradient

Slope of log dwell across delta → theta → alpha → beta → gamma.

Raw versus cleaned

The crucial result changed after preprocessing.

In the raw-style representation, adding dwell gradient to age + three simple spectral features increased mean internal CV AUC by about +0.039. After recomputing the same frozen dwell feature on derivative / cleaned EEG, that increment became essentially zero.

QuantityRaw-styleDerivative / cleaned
dwell AD-CN p0.0002770.006175
Model A · age + spectral0.729 AUC0.778 AUC
Model B · age + dwell0.756 AUC0.694 AUC
Model C · age + spectral + dwell0.768 AUC0.777 AUC
C − A+0.039−0.001
The univariate disease association survives cleaning. The claim that Φ-Dwell adds useful Alzheimer discrimination beyond ordinary slowing does not survive this internal robustness check.
0.0062 p

Dwell still differs

AD mean −0.3551 versus CN −0.3275.

0.778 AUC

Spectral baseline

Age + alpha power + theta/alpha + peak alpha frequency.

−0.001 AUC

Dwell increment

No improvement after cleaning.

The boring explanation

Alzheimer's spectral slowing is not background theory here. It is in the data.

Cleaned featureAD meanCN meanp
alpha relative power0.04940.07940.002351
theta / alpha ratio2.53181.60450.0000732
peak alpha frequency7.493 Hz8.681 Hz0.000155

These simple measures already capture strong disease-related slowing. The current evidence therefore favors Φ-Dwell as an alternative spatial-dynamical view of the same disease physiology rather than a new diagnostic signal.

Do not retune bands, graph modes, dwell definition or gradient direction on ds004504 to recover the lost +0.039. That would turn a clean robustness result into post-hoc rescue.
Severity claim

The historical pooled MMSE story remains dead.

Within AD
ρ 0.269

p = 0.113, n = 36 after cleaning.

Within FTD
ρ −0.009

p = 0.969, n = 23.

The old pooled correlation mixed diagnosis with MMSE. The current data do not support saying that dwell gradient tracks cognitive severity.

A useful negative

The scary control-age correlation was also preprocessing-sensitive.

Raw-style controls showed dwell versus age at rho = −0.599, p = 0.000597. In the cleaned derivative representation it became rho = −0.100, p = 0.607. That argues against treating the earlier age effect as a stable biological property.

Frozen next gate

New people, unchanged feature.

The only decisive Alzheimer test now is an independent cohort. The cleaned internal result makes the prediction harder and therefore more useful.

Primary feature
dwell_gradient unchanged.
Model A
Age + alpha relative power + theta/alpha ratio + peak alpha frequency.
Model B
Age + dwell gradient.
Model C
Age + spectral features + dwell gradient.
Primary comparison
C versus A on completely independent AD/CN subjects.
No rescue
No changing bands, graph sigma, number of modes, word step, log transform, gradient direction or age handling after labels are seen.
External null

Retire biomarker claim

Replicates only

Interesting representation, no new biomarker

Incremental signal

Worth serious follow-up

Context

Why the baseline matters

Recent dementia EEG literature repeatedly reports spectral slowing, lower dominant/alpha frequency and altered theta/alpha balance. Any proposed low-cost EEG biomarker should therefore demonstrate added value beyond those simple measures.

2025 systematic review of electrophysiological dementia markers · spectral and connectivity features in AD · 2025 EEG connectivity review