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.
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
A deterministic transform of multichannel phase organization.
How long each band remains in the same dominant sensor mode.
Slope of log dwell across delta → theta → alpha → beta → gamma.
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.
| Quantity | Raw-style | Derivative / cleaned |
|---|---|---|
| dwell AD-CN p | 0.000277 | 0.006175 |
| Model A · age + spectral | 0.729 AUC | 0.778 AUC |
| Model B · age + dwell | 0.756 AUC | 0.694 AUC |
| Model C · age + spectral + dwell | 0.768 AUC | 0.777 AUC |
| C − A | +0.039 | −0.001 |
AD mean −0.3551 versus CN −0.3275.
Age + alpha power + theta/alpha + peak alpha frequency.
No improvement after cleaning.
| Cleaned feature | AD mean | CN mean | p |
|---|---|---|---|
| alpha relative power | 0.0494 | 0.0794 | 0.002351 |
| theta / alpha ratio | 2.5318 | 1.6045 | 0.0000732 |
| peak alpha frequency | 7.493 Hz | 8.681 Hz | 0.000155 |
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.
p = 0.113, n = 36 after cleaning.
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.
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.
The only decisive Alzheimer test now is an independent cohort. The cleaned internal result makes the prediction harder and therefore more useful.
dwell_gradient unchanged.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