The inverse problem does not stop at the scalp

What the work claims

This is a methods review and tutorial, not a primary experimental result, and it should be weighted accordingly. Leahy and Medani, both at USC and closely associated with the Brainstorm and BrainSuite software ecosystem, survey the full pipeline of noninvasive electrophysiology: the complementary physics of EEG and MEG, the electromagnetic forward problem, the ill-posed inverse problem and its regularization-based solutions, and the estimation of functional and effective connectivity on reconstructed source time series.¹ The unifying claim is disciplinary rather than novel: meaningful network inference from field measurements requires an explicit forward model and source-space reasoning, because sensor-level analysis is dominated by volume conduction artifacts that no amount of post-processing at the sensor level can fully remove.

How it works

EEG and MEG both measure synchronous postsynaptic currents in cortical pyramidal ensembles with millisecond resolution, but through different physics. EEG records scalp potentials, which are smeared by the skull's low conductivity and depend on the full head as a volume conductor. MEG records magnetic fields that pass through skull and scalp almost undistorted, making sources easier to localize, though MEG favors tangential sources in sulci while EEG sees both tangential and radial orientations. Either way, what a sensor captures is a field, not a source.

The forward problem formalizes this: given a current distribution, compute the fields at the sensors, using a volume-conductor model of the head. Early work used concentric spheres; modern practice builds MRI-derived models with boundary element or finite element methods, the latter accommodating anisotropic conductivity in white matter inferred from diffusion imaging, at substantially higher computational cost. The inverse problem then asks the question an experimenter actually cares about, and it is ill-posed: infinitely many source configurations produce the same sensor pattern. The review catalogs the constraint families used to pick one, from equivalent current dipole fitting for focal events like interictal spikes, through minimum-norm estimation and its noise-normalized variants dSPM and sLORETA, to sparse Bayesian methods such as maximum entropy on the mean. Each prior buys a solution at the price of a bias, and the choice of bias is rarely discussed in papers that consume the resulting maps.

Connectivity estimation inherits all of this. Functional measures like coherence, phase-locking value, and weighted phase-lag index are computed preferably on source time series rather than sensors, because volume conduction, acting nearly instantaneously across the head, makes two electrodes share a source and report a strong coupling that reflects one region, not two. Effective connectivity goes further and claims direction: Granger causality tests whether the past of one signal improves prediction of another, dynamic causal modeling inverts a biophysical generative model by Bayesian methods, and transfer entropy generalizes the idea to nonlinear dependencies. The review is appropriately blunt about failure modes. Granger causality assumes linear, stationary interactions and can report spurious directionality under filtering and field spread, which is why it is only defensible on source reconstructions. Dynamic causal modeling is hypothesis-driven and scales poorly, so it applies to a handful of regions, not whole-brain maps. Transfer entropy is data-hungry and still vulnerable to spatial leakage at the sensor level.

Where a skeptic should push

The most load-bearing assumption is that a well-chosen forward model plus a reasonable prior yields source estimates trustworthy enough to support connectivity claims. The review presents the methods as a mature menu, but the literature it synthesizes keeps cataloguing how fragile the inference is: ghost interactions persist even in source space, directionality estimates flip with preprocessing choices, and template head models introduce localization errors that vary across subjects. A tutorial written by the developers of two of the main toolchains is authoritative about practice but not neutral about which tools matter most. Treat the paper as a map of where the traps are, not as evidence that any given connectivity result is correct. Its own conservatism is the tell: it repeatedly advises interpreting connectivity conservatively and validating against invasive recordings where available.

What scalp mixing teaches the organoid array

For microelectrode array work, the temptation is to read this review as someone else's problem. An MEA sits inside the volume conductor, at micrometer range, so the array is supposedly the ground truth modality that scalp EEG and MEG calibrate against. That is true and it is not the whole story. A high-density array over a three-dimensional organoid has its own mixing geometry: extracellular fields spread through tissue and culture medium with finite spatial decay, every electrode records a weighted sum of many cellular sources, and the choice of reference electrode redefines what every other channel means. The review's central caution, that sensor-level correlation is not connectivity, transfers almost verbatim to electrode-level correlation on an array. An electrode pair on an organoid that reports strong coherence may be sharing one patch of synchronous cells, not witnessing an interaction between two networks.

The sharper point is that the organoid field already runs its own inverse problem without naming it. Spike sorting, decomposing each channel's voltage trace into the contributing neurons, is an ill-posed decomposition exactly in the sense this review formalizes: more sources than observations, priors doing the deciding. The scalp community's discipline, an explicit forward model of the conductor, stated priors, and connectivity claims made in source space, is the template. Vendors and pipelines that ship organoid arrays with one-click connectivity heatmaps from raw electrode correlations are re-selling the ghost-interaction literature under a new label.

The opportunity runs the other direction too. Because an array's conductor geometry is known, fabricated, and stable, the forward problem for an organoid-plus-electrode system is far more tractable than for a head: no skull, no subject-to-subject anatomy, no MRI pipeline. A platform that models its own field spread and reports connectivity in estimated cell-source space, with the model versioned alongside the data, would be offering the thing scalp neuroimaging cannot: reproducible, physics-checked network claims. That is a differentiating capability an acquisition vendor could actually build, and this review is the specification for what not to skip.

The bottom line

Established for decades and restated clearly here: field sensors measure mixtures, the inverse problem needs priors, and sensor-level connectivity is confounded by volume conduction. Open for organoid arrays: whether the field adopts the same rigor when the volume conductor is a dish rather than a head. The review supplies the vocabulary and the failure modes; it does not, and cannot, certify any individual connectivity result. Confirming evidence would be array studies that report connectivity on source estimates with an explicit forward model; disconfirming evidence would be demonstrations that electrode-level metrics on dense 3D organoid recordings agree with optically verified cell-resolved connectivity. Neither exists in quantity yet.