Detuning, Not Silencing: Testing Whether MCI Progression Breaks Frequency-Matched Routing

When magnetoencephalography (MEG), a scalp recording of the brain's magnetic fields at millisecond resolution, is taken from people with mild cognitive impairment (MCI), one of the best predictors of who will progress to Alzheimer's dementia is a drop in beta-band power (16–36 Hz) over the back of the head. A drop in power has two possible causes. The beta rhythm could be getting weaker everywhere, as though the volume were turned down. Or the rhythms in different regions could be drifting apart in frequency, so that areas that used to oscillate in step now oscillate at slightly different rates. Theories of frequency-based routing, in which brain areas communicate by tuning to a shared oscillation the way a radio receiver tunes to a transmitter, treat these two cases very differently. A weaker signal still gets through. A mistuned receiver does not. This note proposes to tell the two apart, using a biophysical model to say what each failure should look like and an existing longitudinal MEG cohort to see which one occurs.

The core claim is concrete. In biophysical network models, the strength of chemical synapses sets how large an oscillation is, while the intrinsic preferred frequency of inhibitory neurons, together with their electrical coupling, sets its frequency and whether it is sustained at all. Each kind of damage therefore leaves a different fingerprint in the spectrum. We hypothesize that part of the beta loss in MCI converters is "detuning": the peak frequencies of different cortical regions spread apart. If so, a detuning index should predict conversion beyond what beta power already predicts. If the beta loss is pure amplitude loss at fixed frequency, the index will add nothing, and the routing interpretation of MCI beta loss should be dropped.

What the two directions supply

The routing literature supplies the mechanism. The adaptive-threshold network analysis in Tchumatchenko and Clopath shows that global oscillation frequency is set by the subthreshold resonance of inhibitory interneurons, meaning their tendency to respond most strongly to inputs at one preferred rhythm. Self-sustained oscillation requires both that resonance and electrical gap junctions. Chemical synapses change only the gain. That three-way split, with gain affecting amplitude, resonance affecting frequency, and coupling affecting whether the oscillation is sustained and coherent, is the forward model we reuse. Jahnke et al. supply the routing consequence. In their feed-forward spiking networks, transmission peaks only when a driving oscillation matches the circuit's natural propagation frequency, or a rational ratio of it, so a mismatch in frequency gates signals off. Deger et al. show that coupled adapting populations filter information in a band, and that this filtering sharpens across layers. That result matters here because it means small frequency mismatches can compound along a pathway. Fountas and Shanahan show in a full basal ganglia spiking model that the frequency of the cortical drive decides which pathway carries information. Their pathology was Parkinsonian rather than age-related, so we treat it as evidence that frequency gating can fail in disease, not as a model of MCI.

The cognitive-decline literature supplies the observation and the data. Gaubert et al. analyzed resting MEG from 117 MCI patients in the BioFIND cohort, 64 of whom converted to dementia and 53 of whom stayed stable. They used fine-grained wavelet spectra and Riemannian covariance features, a way of comparing sensor covariance matrices that respects their geometry. Parieto-occipital beta loss was their main predictor of conversion. They explicitly leave the cellular cause unmodeled, and their features are amplitudes and covariances, not peak frequencies or how much those frequencies vary across regions. The other cross-direction work in the file does not test this question. The phase-connectivity study of Cai et al. measures phase in fMRI signals slower than 0.1 Hz, far below beta. The segregation findings in Qian et al. and the small-vessel-disease study Functional Segregation-Integration Preference are also fMRI-based and do not measure frequency. The tACS meta-analysis of Zhang and Cui found that fixed-frequency stimulation does almost nothing overall (Hedges's g = 0.060). That motivates a frequency-matching account, but the meta-analysis contained too few aging or MCI studies to test one. A species caution from Kalmbach et al. constrains the modeling: human and mouse pyramidal neurons differ in h-channel resonance, so parameter priors must not be copied from rodent values.

Research Program

The first stage builds the forward model. We extend the Tchumatchenko–Clopath inhibitory-excitatory population to several regions coupled by delayed long-range connections. This adds axonal delays, which the original analysis omits and lists as a limitation. We then degrade the model in three separate ways: lowering chemical synaptic gain, shifting the interneurons' resonance frequency with different amounts per region, and weakening gap-junction coupling. For each kind of damage we record how the regional spectra change, and how inter-regional transfer entropy changes, transfer entropy being a measure of how much one region's past activity predicts another's future. The expected result is a set of distinguishable signatures. Gain loss should lower peak power while leaving peak frequency unchanged and inter-regional transfer largely intact. Heterogeneous resonance shifts should spread peak frequencies across regions and reduce transfer more than the power drop alone would predict. Loss of electrical coupling should broaden peaks and weaken coherence. If the model does not produce separable signatures, the study stops there and reports that resting spectra cannot distinguish these mechanisms.

The second stage is inference. Using simulation-based inference, we fit the model's gain, resonance, and coupling parameters to each subject's source-reconstructed regional spectra and obtain a posterior distribution for each parameter. Before touching patient data, we check identifiability on synthetic subjects with known damage, asking whether the method recovers the true mechanism at the signal-to-noise levels of BioFIND.

The third stage is the empirical test. Alongside the model parameters, we compute a model-free detuning index. Let $f_r$ be the beta peak frequency estimated in region $r$, $w_r$ the precision of that estimate (one over its bootstrap variance), and $\bar f$ the precision-weighted mean. Then

$$D = \sqrt{\frac{\sum_r w_r (f_r - \bar f)^2}{\sum_r w_r}}.$$

Weighting by precision matters because converters have weaker beta peaks. Weaker peaks yield noisier frequency estimates, and noisier estimates would inflate $D$ even if no real detuning occurred.

Baselines

The naive reference point is the clinical-only model from Gaubert et al., which uses age, education, and Mini-Mental State Examination score and reports an area under the ROC curve (AUC) of 0.71. A second structural reference is that clinical model plus hippocampal volume (AUC 0.75 in the same paper).

The strongest existing method is the Gaubert pipeline itself. Clinical variables plus MEG beta power reach AUC 0.81. Clinical variables plus MRI plus MEG reach AUC 0.83, with 78% sensitivity and 76% specificity. Their Riemannian covariance features add delta-band information, but no reported result found on Lacuna for a separate Riemannian-only AUC. We will re-implement this pipeline on the same BioFIND subjects, with the same cross-validation folds and the same logistic-regression family and regularization search, and report our reproduced numbers alongside the published ones.

The main comparison is nested. Beating the baseline means that adding $D$, or the model-inferred resonance-dispersion parameter, to clinical plus beta power improves out-of-fold log-likelihood under a pre-registered test, and that its coefficient keeps the same sign across folds. With 117 subjects, an AUC gain of a few hundredths cannot be resolved by confidence intervals. We therefore treat the likelihood test, together with the direction of the effect, as the primary criterion, and treat AUC as descriptive.

A phenomenological baseline fits each regional spectrum as a single peak on a smooth background and uses peak amplitude, frequency, and width directly, with no biophysics. No reported result found on Lacuna for this approach on BioFIND.

The key ablation compares the model's gain-only parameters against its resonance-dispersion parameters, which isolates the tuning hypothesis from the amplitude hypothesis.

Why not just…?

Why not just feed band power and covariance into a stronger classifier? That is what Gaubert et al. did, and it predicts conversion well. But power and covariance features mix amplitude and frequency together, so a classifier built on them cannot say whether the beta loss is silencing or detuning. The question here is mechanistic. A detuning account predicts a specific kind of failure (lost communication despite preserved power) and a specific remedy (individually matched frequencies), and a power-only account predicts neither.

Why not just fit peak frequencies directly, without the biophysical model? This is the real threat. If the model-free $D$ carries all of the predictive signal, the biophysical inversion adds interpretation but no predictive power, and we will say so. The model earns its place only if it does something $D$ cannot. One candidate is separating a power drop caused by gain loss from one caused by coupling loss, since both can look like a weaker peak. Another is predicting the transfer-entropy loss a given $D$ should produce. The ablation that compares the phenomenological fit against the model fit settles this directly.

Why not just use structural MRI or end-to-end deep learning? On this cohort, hippocampal volume adds less than beta power does (0.75 versus 0.81 in Gaubert et al.). Deep structural models such as TA-SSM Net classify MCI versus controls well, but they say nothing about the timing of neural communication. Structural features enter our analysis as covariates, not as competitors.

Why not just run individualized-frequency tACS and see if it helps? The meta-analysis in Zhang and Cui shows how noisy and small fixed-frequency effects are. A stimulation trial without a prior measurement of detuning would not know which frequency to match, or which regions to match it in. This study supplies that measurement first.

Decisive Checks

The first check is estimator bias. We will simulate converters whose only change is lower amplitude, then confirm that precision-weighted $D$ does not rise in those simulations. If it does rise, the detuning result is an artifact of noisier peak estimates and is withdrawn.

The second check is field spread. Magnetic fields from one source reach many sensors, which can blur frequency differences between regions. We will compute $D$ in source space with leakage correction and repeat the computation with sensor-level estimates. A real effect should keep the same sign in both, and should be stronger where leakage correction is applied.

The third check is mechanistic consistency. In the forward model, the dispersion inferred for each subject should predict a specific reduction in inter-regional transfer. That predicted reduction should correlate with the reduction actually measured in the subject's resting MEG.

If converters show higher $D$ under all three checks, frequency-matched routing becomes a measurable target in MCI. If they do not, beta loss in MCI is best read as amplitude failure, and frequency-based routing accounts of decline lose their most direct test case.