Slow-Phase Gating of Faster Brain Fluctuations as a Late-Stage Alzheimer's Marker in ADNI Resting-State fMRI

Brain rhythms interact across speeds. In phase-amplitude coupling, the timing (phase) of a slow rhythm controls how large a faster rhythm gets. The fast activity "rides" on particular parts of the slow cycle. In electrical recordings, this gating supports memory retrieval and the routing of information between brain areas. Alzheimer's progression models, however, almost never measure it. The large public cohort these models are built on, the Alzheimer's Disease Neuroimaging Initiative (ADNI), collects structural MRI, PET, spinal fluid and cognitive tests, but not EEG. Its main source of timing information is resting-state functional MRI (fMRI), which measures slow blood-oxygen fluctuations that can be split into several narrow frequency sub-bands.

We propose to measure cross-band phase gating inside the blood-oxygen signal itself. The question is whether the fidelity of that gating tracks tau pathology in the prodromal stage of Alzheimer's (cognitive impairment with confirmed pathology, before dementia). Fidelity here means how consistently, cycle after cycle, the faster band's amplitude peaks at the same phase of the slower band. Prodromal disease is where simpler within-band connectivity measures appear to stop tracking pathology. In two sentences: within-band network measures saturate once neurodegeneration is under way, but the cycle-by-cycle precision with which slow fluctuations gate faster ones may keep degrading with tau. We will test this in ADNI against tau-PET stage and conversion to dementia, with controls designed to rule out the many ways such coupling can be spurious.

Core Claim

The motivating observation comes from high-density EEG with tau-PET. That study measured alpha-band network structure using the phase-lag index, a measure of consistent timing offsets between channels. Tau burden correlated strongly with alpha network disintegration in amyloid-positive but cognitively normal people (clustering coefficient r = −0.82). The correlation vanished in prodromal AD (r = −0.03, p = 0.92). One reading is that within-band synchrony is already maximally disrupted by the prodromal stage, so it no longer varies with further tau.

Our claim is that a cross-frequency quantity, specifically the precision of slow-phase gating, keeps varying with tau through the prodromal stage. We also claim that in ADNI it can be estimated from fMRI sub-bands. The new elements are three. First, the target is gating precision rather than coupling strength. Second, the coupling is computed across hemodynamic sub-bands rather than across EEG bands. Third, it is evaluated against tau stage and conversion in ADNI, where no cross-frequency measure has been used.

The prior work plays different roles. Some of it supplies the substrate. Wee et al.'s multi-spectrum fMRI study split the resting fMRI signal into five sub-bands between 0.025 and 0.100 Hz. It showed that impairment-related information concentrates in specific bands, rising from 59.5% to 86.5% accuracy over full-spectrum analysis in 37 non-ADNI subjects. It analyzed each band in isolation, never interactions between bands. The LFP–BOLD band study found that different electrical bands carry non-redundant information about the blood-oxygen response. That gives some reason to expect structure across blood-oxygen sub-bands too, though it does not establish phase gating there.

Other work supplies the method. Coupling is extracted with band-pass filtering and the Hilbert transform, following the review Untangling cross-frequency coupling in neuroscience. It is then scored with a modulation index that measures how far the amplitude-by-phase histogram departs from flat, as compared in the epileptogenicity methods study. The fidelity idea borrows from the entrainment literature. Auditory afferents in sound-producing fish lock firing to individual stimulus cycles, and the integrate-and-fire coding study shows that phase within a slow cycle carries information beyond rate. Both suggest that cycle-by-cycle consistency, not average strength, is the more informative quantity.

A third group motivates the idea without addressing Alzheimer's. Allen et al.'s comodulogram analysis found disease-related reorganization of cross-frequency modulation in schizophrenia EEG. Lega et al. showed that theta-gamma coupling predicts hippocampal memory retrieval independently of theta power. Neither touches neurodegeneration or molecular pathology.

Measures

For a region, let $\phi_s(t)$ be the Hilbert phase of its slow sub-band signal and $A_f(t)$ the Hilbert amplitude of a faster sub-band. Coupling strength is the standard modulation index: the divergence of the phase-binned mean amplitude from uniform. Gating fidelity works cycle by cycle. For each slow cycle $k$, compute the slow-band phase $\theta_k$ at which $A_f$ peaks within that cycle. Across the $K$ cycles in the scan, the fidelity is

$$ F = \left| \frac{1}{K}\sum_{k=1}^{K} e^{i\theta_k} \right| , $$

which is 1 when the fast amplitude always peaks at the same slow phase and near 0 when the peak phase is random. Coupling strength and fidelity can come apart. A region can show a large average modulation driven by a few cycles while the preferred phase drifts. Our hypothesis is that tau degrades $F$ before it degrades modulation index.

The faster band in each pair is taken from Wee's sub-bands around 0.054–0.068 Hz. The slow phase comes from the lowest band the scan duration can support. Because the slow band must be well below the faster band, only a handful of cycles fit in an ADNI scan. Fidelity estimates will therefore be noisy. For that reason we pool $\theta_k$ across regions within a network: medial temporal, default mode, and a control network such as primary visual. We also treat test-retest reliability as a hard gate rather than an afterthought.

Research Program

The first stage builds the measurement pipeline on ADNI resting-state fMRI and validates it before any clinical analysis. Validation rests on surrogate tests. Time-shifting the amplitude series relative to the phase series gives a null distribution for each subject. Phase-randomized surrogates that preserve each band's spectrum test whether coupling exceeds what filter overlap produces. Regressing out motion and physiological noise, and checking coupling against framewise displacement, tests whether head motion manufactures it. Reliability is assessed between repeated ADNI visits in clinically stable, amyloid-negative controls. If within-subject reliability of network-pooled $F$ is poor, the project stops at a reported negative result about what ADNI's fMRI can support.

The second stage relates the measures to pathology. The outcomes are continuous medial-temporal tau-PET uptake and SuStaIn stage. SuStaIn is the unsupervised staging method of the missing-data-enabled ADNI subtype work, which the inter-cohort validation showed generalizes beyond ADNI. These relationships are fitted separately within amyloid-positive cognitively normal participants and within prodromal participants. Stratifying this way directly tests the saturation pattern seen with alpha phase-lag index. Age is handled with a normal-aging reference fitted on amyloid-negative controls, in the spirit of the aging-disentanglement MRI work. Scanner and site effects are modeled explicitly, following the concern raised in GSplit LBI.

The third stage asks whether fidelity adds prognostic information. Among MCI participants with fMRI and follow-up, conversion to dementia is modeled with Cox and logistic models. Fidelity is added to a reference set of hippocampal volume, amyloid status and baseline cognition.

Baselines

The naive reference is full-spectrum resting connectivity: correlation of the unfiltered 0.01–0.1 Hz signal between regions. It represents doing no frequency analysis at all. Its only reported clinical result is Wee's 59.5% MCI-versus-control accuracy, on a different cohort, so we will recompute it on our ADNI sample.

The strongest within-band competitor is Wee's multi-spectrum approach, reimplemented on ADNI: per-sub-band weighted clustering coefficients plus per-band fluctuation power. It reported 86.5% accuracy and AUC 0.863 on the 37-subject Duke-UNC cohort; no reported result was found on Lacuna for it on ADNI. It is the most dangerous competitor because it uses exactly the same sub-bands and differs from our proposal only by ignoring how the bands interact. To beat it, fidelity must explain tau variance within the prodromal group that this feature set does not, with a confidence interval on incremental R² excluding zero. In a direct head-to-head on the same subjects, fidelity alone should also correlate with tau within the prodromal stratum where the within-band measures do not.

For conversion prediction, the strongest reported ADNI numbers come from methods using richer data than ours. HiMAL reports an AUROC of 0.923 using longitudinal MRI, cognition and clinical variables. The amyloid-PET-plus-atrophy composite reports an AUC of 0.807 and a multivariate hazard ratio of 5.70. SuStaIn with MMSE and spinal-fluid amyloid reaches an AUC of 0.76 on external test data. We do not expect fMRI coupling to beat HiMAL's AUROC outright. The fair test is incremental: the reference model with and without fidelity, on the same MCI subjects with the same nested cross-validation. All models get the same tuning budget, a regularized linear model with a fixed hyperparameter grid, so that differences cannot come from model capacity. The claim succeeds if fidelity improves concordance or AUROC over the reference with a bootstrap interval excluding zero.

Two ablations isolate the key ingredient. Replacing fidelity with modulation index tests whether cycle-by-cycle consistency matters beyond average strength. Replacing phase information with amplitude-amplitude correlation between the same two bands tests whether phase matters at all, or whether any cross-band envelope relationship would do. Both receive the same feature count and pooling.

Why not just…?

Why not just use tau-PET, which already measures the pathology directly? For staging, tau-PET will win, and we do not claim otherwise. The point of the second stage is mechanistic: does a functional signal keep tracking pathology where within-band synchrony saturates? The conversion analysis keeps tau-PET in the reference model wherever it is available. Fidelity therefore has to add information beyond it. If it adds nothing once tau is included, we will report that the measure is a proxy rather than independent signal.

Why not just use per-band power and within-band connectivity, as Wee did? This is the real threat. Wee's sub-band features already capture most of the diagnostic structure their data contained. Any cross-band effect could turn out to be a re-expression of band power, since amplitude estimates scale with power. We address this by using the Wee features as the primary baseline. Fidelity is computed from peak phase, which is invariant to overall amplitude scaling. The decisive check is incremental tau variance within the prodromal stratum. If fidelity carries no information beyond Wee's features, we will conclude that within-band analysis suffices in this regime.

Why not just run the EEG coupling methods that already work elsewhere, such as Riemannian cross-frequency covariances or Phase-Swap pretraining? The SPD cross-frequency covariance approach reached 91.90% on mental workload. Phase-Swap self-supervision improved sleep staging to 72.0% versus 67.9% at 20% labels. Both require dense electrical recordings that ADNI does not have, and both target classification accuracy rather than tracking pathology through a saturation regime. An EEG–tau cohort like van Nifterick's would be the natural place to apply them. That is a complementary study, not a substitute for testing what is measurable at ADNI's scale.

Why not just call any cross-band coupling in fMRI an artifact? It may be. Overlapping filters, non-sinusoidal slow waves, motion bursts and very few slow cycles per scan can all manufacture apparent coupling. This is why surrogate nulls, motion regression and test-retest reliability form a gate the project must pass before any clinical analysis. A primary visual control network should show no tau relationship. A failure at the gate is itself a useful, publishable constraint on the many fMRI cross-frequency analyses this question invites.

Decisive Checks

The project hinges on three results, in order. First, network-pooled fidelity in stable amyloid-negative controls must exceed surrogate nulls and show acceptable visit-to-visit reliability. Second, within prodromal participants, fidelity must correlate with medial-temporal tau or SuStaIn stage after adjusting for Wee's within-band features, age and site. This is the pattern absent for alpha phase-lag index in the EEG study, and finding it would resolve that study's saturation puzzle in favor of cross-frequency measures. Third, fidelity must outperform both ablations, modulation index and amplitude-amplitude correlation, showing that phase consistency rather than generic cross-band covariation carries the signal.

Conversion improvement over the reference model is a secondary result. We will report it whether positive or null, alongside the published HiMAL, composite-feature and SuStaIn numbers, with a clear statement that those used different inputs and subsets.

References

  1. Effect of tau on global graph measures from high‐density EEG in preclinical and prodromal stage of Alzheimer disease
  2. Resting-State Multi-Spectrum Functional Connectivity Networks for Identification of MCI Patients
  3. Different LFP frequency bands convey complementary information about the BOLD signal
  4. Untangling cross-frequency coupling in neuroscience
  5. Assessing Epileptogenicity Using Phase-Locked High Frequency Oscillations: A Systematic Comparison of Methods
  6. Temporal Encoding for Auditory Computation: Physiology of Primary Afferent Neurons in Sound-Producing Fish
  7. Information content and robustness of various types of codes in integrate and fire networks presented with naturalistic stimuli
  8. Components of Cross-Frequency Modulation in Health and Disease
  9. Theta power and theta‐gamma coupling support long‐term spatial memory retrieval
  10. Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference
  11. Inter-Cohort Validation of SuStaIn Model for Alzheimer's Disease
  12. Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images
  13. GSplit LBI: Taming the Procedural Bias in Neuroimaging for Disease Prediction
  14. HiMAL: A Multimodal Hierarchical Multi-task Auxiliary Learning framework for predicting and explaining Alzheimer disease progression
  15. Prediction of Conversion from Mild Cognitive Impairment to Alzheimer’s Disease Using Amyloid PET and Brain MR Imaging Data: A 48-Month Follow-Up Analysis of the Alzheimer’s Disease Neuroimaging Initiative Cohort
  16. Novel SPD Matrix Representations Considering Cross-Frequency Coupling for EEG Classification Using Riemannian Geometry
  17. Boosting Generalization in Bio-signal Classification by Learning the Phase-Amplitude Coupling