Cross-modal coupling and the source prior

What the work claims

This is a methods study evaluated entirely in simulation: no scalp recordings, no implanted arrays, ground truth supplied by coupled Epileptor neural mass models generating seizure dynamics.¹ The starting point is geometric eigenmode source localisation, which represents cortical activity in a basis of eigenmodes of the cortical surface geometry, a compact spatial prior that had previously been shown to mitigate the underdetermined nature of EEG and MEG source localisation. This paper asks whether neural field theory can add a temporal constraint on top: each eigenmode is assigned an analytically derived transfer function, so modes are weighted not just by how observable their spatial scale is but by how they should evolve in time.¹

The claim that survives contact with the data is negative and precise. Applying the fixed analytical transfer function did not reliably improve reconstruction across metrics and noise levels: with no added noise it degraded all three metrics, with effect sizes of Cohen's dz equals -0.947 for region localisation error, -4.566 for normalised mean-squared error and -2.502 for cosine similarity; at 3 dB SNR only cosine similarity improved, at dz equals 1.548.¹ In contrast, empirically estimated transfer functions, which retain the off-diagonal terms describing how activity in one eigenmode drives another, substantially improved reconstruction, especially in noise, with the strongest non-oracle results from a hybrid combining empirical cross-modal structure with the analytical mode-specific scaling. Time-varying oracle transfer functions, estimated from the known simulated sources, set the upper bound, with effects from dz equals 2.819 to 8.683.¹

How it works

The forward model is linear and quasi-static: measured EEG equals the lead-field matrix times source activity plus noise, with 88 electrodes in a 10 to 10 montage and 20,484 cortical dipole sources at roughly 3.1 mm spacing, mapped through a three-shell boundary-element head model built with OpenMEEG and MNE-Python.¹ Because 88 sensors cannot determine 20,484 sources, the source activity is expanded in a few geometric eigenmodes and the pseudoinverse is truncated, with the regularisation threshold tied to the assumed SNR and an inverse-wavenumber weighting that favours the spatial scales EEG can actually see.¹

The analytical temporal constraint comes from the Robinson neural field model: in the linearised theory, each eigenmode evolves independently according to its own transfer function, which yields a diagonal weighting matrix in mode space.¹ That independence assumption is the failure point. Real cortical dynamics involve recurrent coupling, thalamocortical feedback and nonlinear population effects that redistribute activity across modes, so the true relationship between modes is a full matrix, not a diagonal one. The empirical transfer function is exactly that full matrix, estimated from short-time Fourier transforms of the source activity, and the paper's hierarchy tests how much specificity it needs: one matrix averaged over all simulations, a matrix estimated per simulation, and the time-varying oracle. Performance climbs the hierarchy, and the averaged matrix only beats the spatial-only baseline once noise is added, which tells you the population-level dynamical prior carries real but limited information.¹

Where a skeptic should push

Everything is simulated, and the simulation stack is itself a chain of assumptions. The Epileptor model is a low-dimensional neural mass caricature chosen to produce seizure-like dynamics; the forward model is a three-shell boundary element approximation; and the ground truth used to build even the non-oracle empirical transfer functions is the known simulated source activity. The paper is honest about the resulting circularity: with real scalp EEG the sources are unknown, so the transfer functions you need are the transfer functions you are trying to estimate, and the practical estimation problem, from intracranial recordings, from fMRI, or from a coupled analytical model, is explicitly left unsolved.¹ A simulation-specific matrix estimated from the same simulation whose sources you are reconstructing is closer to the oracle end of the hierarchy than to any estimator a laboratory could deploy.

Second, the headline failure of the analytical prior is partly a statement about seizure dynamics, which are strongly nonlinear and non-stationary, exactly the regime where a fixed linear transfer function should fail by construction. The authors reproduce the main pattern in three coupled cortico-thalamic and hippocampal-septal neural field simulations fitted to real patient seizures, which broadens the evidence, but there is still no demonstration on measured EEG of any kind.¹ Third, the apparent improvement of oracle methods after noise is added is partly an artefact of the SNR-dependent truncation of the pseudoinverse rather than pure temporal information, which the authors flag themselves.¹ None of this overturns the central mechanistic finding, that independent-mode analytical dynamics misdescribe coupled cortical activity. It does bound the paper's practical payload: what is proven is that measured cross-modal dynamics would help if you could estimate them, not that anyone yet can.

Array geometry and the observable subspace

The array connection is not metaphorical, because the array version of this problem is worse, not better. Here the underdetermination is 88 sensors against 20,484 sources, a ratio of about 1 to 233. A state-of-the-art CMOS MEA with tens of thousands of electrodes recording an organoid containing millions of neurons faces the same ill-posed inversion through a near-field rather than volume-conduction operator, and adding electrodes never closes the gap; it only enlarges the observable subspace.¹ This paper quantifies what actually buys back reconstruction quality, and the answer is not more forward-model sophistication. An analytically elegant prior built from idealised independent dynamics made the estimate worse, with an effect size of -4.566 on mean-squared error, while measured cross-modal coupling helped most exactly when the data were noisiest. For everyone running current-source-density estimates, laminar analyses or effective-connectivity inference on array data, that is a direct caution: a spatial or dynamical prior that has not been validated against the preparation's measured coupling can degrade the very quantity it was meant to stabilise, silently, because regularised inverses do not report when their prior is wrong.

The opportunity runs the other way. The paper's own proposed remedy for the chicken-and-egg problem is to estimate cross-modal transfer functions from complementary measurements with greater spatial specificity, naming simultaneous or previously acquired intracranial recordings, with cited precedent for fitting neural field coupling parameters from invasive data.¹ That reframes dense array hardware as calibration instrumentation: a high-density MEA or depth electrode array is the instrument that can measure the modal coupling a macro-scale inverse model needs, turning arrays from endpoints of research into ground-truth estimators for the wider source-imaging pipeline. A vendor that sells not just electrodes but characterised dynamical priors, measured per preparation class, would be selling the thing this paper shows is actually load-bearing.

There is also a procurement-grade lesson in the runtime numbers. Reconstructing a 60 s seizure over the 20,484-dipole source space took about 0.4 ms per time step for the spatial-only method and 2.1 ms per time step for the temporal method on an Apple M1 Pro, while the time-varying oracle was slower than the recording itself.¹ Inverse-problem processing is therefore part of the acquisition chain in the same way the amplifier is: it has a throughput budget, a latency and a memory footprint that scale with retained modes, frequency bins and time windows, and the best-performing priors are the ones that blow the real-time budget. When array vendors quote channel counts and sampling rates, the companion specification that matters for anyone doing source-level work is the reconstruction cost per unit time at a stated mode count, because that number decides whether the analysis lives beside the rig or after the experiment.

The bottom line

Established, in silico across 2038 simulated seizures: fixed analytical temporal priors from linear neural field theory do not improve geometric-eigenmode EEG source localisation and frequently degrade it, because independent modal evolution misdescribes recurrently coupled dynamics; empirically estimated full transfer-function matrices, retaining cross-modal coupling, do improve reconstruction, with the benefit growing as the prior becomes more specific and as noise rises. Not established: that any of these empirical matrices can be estimated from real data without knowing the sources, or that the effect survives measured EEG, real volume conduction and non-Epileptor dynamics. For MEA instrumentation the transferable conclusions stand on the mechanism, not the platform: electrode density selects which spatial modes are observable and can never make the inverse well-posed by itself; priors applied to array data must be calibrated against measured coupling from the same preparation class; and the inverse solver's compute cost belongs in the instrument's specification sheet next to bandwidth and channel count. The experiment that would move this from warning to workflow is the one the paper proposes: estimate modal coupling from dense invasive or array recordings, apply it as a prior to a sparse modality, and show the reconstruction beats both unregularised baselines and theory-only priors on held-out real data.