Brain Metastability Analyzer Tool

2026 reality-audit reset: this repository contains a real EEG state-space idea, but the historical Alzheimer's claims are exploratory and are not a validated diagnostic tool.

Live audit page: https://anttiluode.github.io/BrainMetastabilityAnalyzerTool/

EEG phase at sensors
 ↓
spatial graph-Laplacian basis built from electrode geometry
 ↓
dominant sensor-phase mode per frequency band
 ↓
dwell times / transitions / discrete multi-band states

The safest name is spatial sensor-phase modes. These are modes of the electrode-layout graph, not structural-connectome eigenmodes and not a holographic reconstruction of the brain.

The older repository used terms such as holographic brain, criticality, grammar, and brain viscosity. Those can be metaphors, but they are not evidence. See AUDIT_2026.md.

Keep:

• graph-Laplacian projection of multichannel EEG phase onto spatial sensor-layout modes;
• per-band dominant-mode dwell times and transition statistics;
• multi-band state words as an exploratory discretization;
• the already-discovered dwell gradient as a frozen candidate for independent replication.

Repair / quarantine:

• Pearson correlation of integer mode IDs is not a sound categorical coupling measure;
• CV > 1 is dwell variability, not proof of criticality;
• Alzheimer bigram perplexity is trained and scored on the same sequence;
• PhysioNet "task doubles vocabulary" used unequal observation time and whole-run task labels despite alternating T0/T1/T2 epochs;
• PSI / Gerchberg–Saxton remains an unvalidated exploratory transform;
• the legacy age parser reads lowercase age even though ds004504 uses Age.

The later brain_viscosity.py branch contains a simple feature that does not need the viscosity story:

[ g = \operatorname{slope}\left[\log(1+D_\delta),\log(1+D_\theta),\log(1+D_\alpha),\log(1+D_\beta),\log(1+D_\gamma)\right]. ]

Historical discovery on OpenNeuro ds004504 reported strong-looking AD/CN separation and a pooled MMSE association. Those were discovery statistics from the same cohort on which the feature was developed.

The preregistered-style internal audit processed all 88 subjects with zero failures and compared the frozen dwell gradient with ordinary spectral slowing.

Internal subject-wise CV:

A = age + spectral AUC = 0.729 ± 0.142
B = age + dwell_gradient AUC = 0.756 ± 0.132
C = age + spectral + dwell_gradient AUC = 0.768 ± 0.128
C - A = +0.039 AUC

That looked interesting, but it still reused the discovery cohort and therefore could not validate the feature.

Full raw-style receipt: INTERNAL_SPECTRAL_AUDIT_RESULT_2026.md.

The same frozen dwell transform was then recomputed on the dataset's derivative / cleaned EEG. Again, all 88 subjects completed.

So the univariate dwell difference survives cleaning, but its apparent incremental value does not:

A = age + spectral AUC = 0.778 ± 0.121
B = age + dwell_gradient AUC = 0.694 ± 0.128
C = age + spectral + dwell_gradient AUC = 0.777 ± 0.118
C - A = -0.001 AUC

Internal classification:

INTERNAL_DERIVATIVE_DWELL_SEPARATION_NO_INCREMENT

This is the main current result. Φ-Dwell still measures a disease-associated difference, but on cleaned EEG it does not improve held-out AD/CN discrimination beyond age + ordinary spectral slowing in this cohort.

Full cleaned-data receipt: INTERNAL_DERIVATIVE_AUDIT_RESULT_2026.md.

The pooled historical MMSE association does not survive as a within-disease severity result.

Raw-style:

within AD: rho = +0.215, p = 0.207
within FTD: rho = +0.121, p = 0.582

Cleaned:

within AD: rho = +0.269, p = 0.113
within FTD: rho = -0.009, p = 0.969

The defensible claim is group association in a discovery cohort, not cognitive-severity tracking.

The raw-style audit found a strong dwell/age association in controls (rho = -0.599, p = 0.000597). After derivative preprocessing it vanished (rho = -0.100, p = 0.607). That makes the earlier age signal a robustness warning rather than a stable biological result.

Alzheimer's EEG slowing is plainly visible in this dataset. In the cleaned analysis, age + alpha relative power + theta/alpha ratio + peak alpha frequency reached mean AUC 0.778, while adding dwell gradient changed that to 0.777.

So the current interpretation is:

Φ-Dwell may be an interesting spatial-dynamical representation of disease-related EEG change, but ds004504 does not show added diagnostic value beyond simple spectral slowing after cleaning.

That is a useful result. It removes the strongest easy explanation for calling this a new cheap Alzheimer detector.

The only decisive next test is on completely independent AD/CN subjects with the definition unchanged.

Model A: age + spectral baselines
Model B: age + dwell_gradient
Model C: age + spectral baselines + dwell_gradient

Primary comparison: C versus A on independent subjects.

No changing bands, graph modes, graph sigma, word step, dwell definition, log transform, gradient direction, age handling, or primary endpoint after external labels are inspected.

External verdicts remain:

• EXTERNAL_DWELL_GRADIENT_NULL
• REPLICATES_BUT_NO_INCREMENT_OVER_SPECTRAL_SLOWING
• EXTERNAL_INCREMENTAL_SIGNAL

Even EXTERNAL_INCREMENTAL_SIGNAL would establish a research signal, not clinical diagnostic utility.

• eigenmode_metastability.py — foundational dwell analysis.
• phidwell_alzheimers.py — historical discovery analyzer.
• brain_viscosity.py — origin of the dwell-gradient candidate.
• phidwell_spectral_audit.py — 2026 spectral-slowing audit.
• phidwell_dwell_recompute.py — frozen dwell recomputation on raw/derivative EEG.
• AUDIT_2026.md — methodological audit and frozen external gate.
• INTERNAL_SPECTRAL_AUDIT_RESULT_2026.md — raw-style internal receipt.
• INTERNAL_DERIVATIVE_AUDIT_RESULT_2026.md — cleaned EEG robustness receipt.
• Results/phidwell_spectral_audit.json — raw-style machine-readable receipt.
• Results/phidwell_spectral_audit_derivatives.json — cleaned machine-readable receipt.
• Alzheimers Phase Stability Index Test/ — quarantined exploratory PSI work.

This repository is research software. It is not a medical device, diagnostic test, or clinical decision tool. No prospective diagnostic accuracy, acquisition-system generalization, or clinical utility has been established.

MIT