Where to stimulate is not where the recorded signal deviates most

What the work claims

This is a computational causal study on human data, and it should be weighted as such: no tissue, no electrodes, and a stimulation that exists only inside a fitted model. The authors take resting-state functional MRI parcellated into 121 cortical and subcortical regions and fit, per subject, a reservoir-computing surrogate, a fixed random recurrent network whose simple linear read-out is tuned to reconstruct that individual's lagged functional connectivity.¹ Two read-outs of those fitted weights classify Alzheimer's disease from controls at a modest area under the curve of roughly 0.66 to 0.70, below what structural atrophy achieves, a point the authors concede and then set aside because atrophy is not something stimulation can act on.

The bold claim is causal and counterfactual. Using each personalised model as an in-silico testbed, the authors show that the disease signature is functionally distributed: correcting all 121 sites at once reverts most patients, but correcting the single most abnormal site, or the top five, fails even at supra-physiological amplitude. The sharper claim is a dissociation. The site whose recorded read-out deviates most from the control template is subcortical and limbic, most often pallidum, nucleus accumbens, brainstem and amygdala, whereas the site whose stimulation most efficiently reverts the classifier is cortical, patient-specific, and led by left default-network prefrontal cortex. The locus of maximal pathology and the locus of maximal therapeutic response are not the same place.

How the target is chosen

The model is teacher-forced on the principal components of the recorded signal, then run closed-loop to free-run a synthetic reconstruction, and it is individually identifiable: a subject's simulated connectivity resembles their own empirical connectivity far more than anyone else's. Pathology is scored per site as the column norm of the difference between the control-mean read-out and the patient's, so the most-affected sites are, by construction, where the recording departs most from the template. When that difference is applied as a correction to only those focal sites, it is inert; a single-column edit stays flat at the roughly 18 percent baseline even when its strength is pushed fifty-fold. Only the full distributed correction, spread across all sites, reshapes the dynamics enough to reclassify the majority.

The physical intervention the authors then model is a focal oscillatory drive, and here the criterion for choosing the site is decisive. A drive tuned to the network's dominant eigenmode reshapes connectivity network-wide, but only where the network is dynamically responsive to it. Selecting the site by its own deviation, or by modal excitability, reclassifies only about half of patients. Selecting each patient's site by its effect on the disease discriminant, the criterion the authors call LDA-resonant, reclassifies essentially all of them from a single site. Tellingly, the discriminant-aligned drive and a plain eigenmode drive perturb the simulation by almost the same amount, each landing about 0.32 to 0.34 in correlation with the unstimulated run, that is, both push roughly equally far from baseline, yet only the discriminant-aligned one converts that perturbation into full reclassification. The extra efficacy comes from where the drive lands, not from how hard it pushes.

Two quantitative details carry the instrumentation argument. The effective drive is frequency-tuned: efficacy peaks at the reservoir's dominant eigenmode, about 0.026 Hz in physical time given the three-second sampling interval, an infraslow rhythm rather than anything at spiking rates. And the targets are heterogeneous: 33 distinct therapy sites across 40 patients, with no shared anatomical hub, overlapping the 35 distinct pathology sites in only 11. A closed-loop controller that titrates amplitude from a sliding-window estimate of the biomarker, computed only from the model's own past output with no access to the future, then reaches comparable efficacy, 39 of 40 versus 40 of 40, at a mean dose cut from 6.0 to 3.3, roughly 44 percent lower.

Where a skeptic should push

The single most load-bearing assumption is that a per-subject generative surrogate, fit to describe resting activity, faithfully predicts the counterfactual effect of a perturbation that was never physically applied. Identifiability is not causal validity. The authors demonstrate that each model reconstructs its own subject's connectivity and reproduces its lagged structure toward the test-retest ceiling, which is a real and careful result, but a descriptive model that is individually identifiable can still mispredict what a real drive would do, and no real drive was applied. The reversions, moreover, require supra-physiological amplitudes; the authors are explicit that the significance is methodological, that an identifiable model turns where and how to stimulate into a computable prescription, not that it demonstrates a therapy.

Several of the load-bearing results are also partly properties of the parameterisation, and the authors flag the most important one. The most-affected subcortical and limbic sites are also the most weakly interconnected nodes of the parcellation, exactly where a shared reservoir basis constrains the per-subject read-out least, so the deviation ranking should be read as where the correction kernel carries its weight, not as a validated map of pathology. The classifier is weak, roughly 0.70 and below atrophy, so reclassification is the movement of a modest discriminant across a boundary rather than a clinical change: a surrogate of a surrogate. And the whole apparatus lives in functional MRI at a three-second sampling interval, so the resonance is infraslow and the read-outs are hemodynamic. Nothing here operates at the timescale or the spatial grain of a microelectrode. The demonstrated facts are the in-model distribution of the signature, the dissociation of the two site sets, and the closed-loop dose reduction; the asserted step is that a physical stimulator at the model-identified site would revert anything in a real patient.

What personalised targets demand of arrays

State the boundary first: this is human functional MRI and simulated stimulation, and nothing in it touches an electrode. What survives translation is a design principle about the acquisition-and-stimulation loop, namely that the site you should drive is defined by its effect on a decodable biomarker, not by where the recorded signal deviates most from a template. That principle is agnostic to modality, and it is precisely the intuition a deviation-mapping array invites you to get wrong. A high-density array that renders a beautiful map of which channels look most abnormal is pointing at the pallidum-and-amygdala equivalent, the decoy target, while the electrode that would actually move the readout sits somewhere the map does not flag. Two honesties bound this. The result lives inside the fitted model, and the authors note the most-deviant sites are also the most weakly interconnected nodes of their parcellation, so a single-site correction there may be inert partly by construction; what the finding licenses is scepticism of deviation-ranked targeting, not a law that single-site biological stimulation is futile. Against that, the focal drive stayed inert even when pushed to supra-physiological strength, which does argue the failure is structural rather than a matter of turning up the amplitude.

The threat this poses to hardware is structural and specific. If the responsive target is per-subject and heterogeneous, 33 distinct sites for 40 individuals with no shared anatomical hub, then a sparse implant placed at the anatomically obvious deviation sites will miss it for most people, because you cannot pre-place an electrode at a target you can only discover by fitting a generative model to that individual's own recordings. The heterogeneity is not total; the therapy sites were biased toward left default-network prefrontal cortex, so an array anchored there would catch a plurality, but anchored is not sufficient when the exact site moves from person to person. That is an argument for dense, individually addressable coverage over sparse targeted electrodes: coverage is what lets an array reach a site defined only after the fact. The transfer is by analogy and should be read as qualitative, because the scales do not match, 121 macroscale parcels of one to two centimetres here against microelectrodes tens of microns wide, so the point is the principle of coverage, not a channel count. For an organoid MEA the principle is direct: the electrode worth stimulating through is the one whose drive most moves a decodable population biomarker, found by search over the array, not the one with the largest baseline deviation.

The rest of the loop follows only in part, and I need to retract one reading from my first pass. The controller titrates dose from a biomarker computed only from causally available recordings, which means the same array must estimate the biomarker in real time and deliver the drive, a record-and-stimulate co-design in which the target-selection logic lives close to the front end. This is a different concurrency problem from rejecting a stimulus artifact; here the acquisition chain's job is target identification and dose control from its own data, and it transfers. What does not transfer is the biophysics of the drive. The optimal frequency here, about 0.026 hertz, is a resonance of a reservoir fitted to a hemodynamic signal at a three-second sampling interval, an infraslow mode of the surrogate rather than a circuit resonance of tissue, and it sits orders of magnitude below any spike-band rhythm an MEA records. I first read that frequency tuning as an argument for tunable per-channel stimulation hardware, and I withdraw it: the specific resonance, the resonant-drive mechanism and the abstract dose units do not carry to microelectrodes. The most one can keep is a conditional, that if a real network mode exists in tissue then matching a drive to it would need frequency-agile stimulation, but nothing here establishes that such a mode exists. The closed-loop result carries as a direction rather than a number. The controller cut mean dose by roughly 44 percent at the cost of one responder, 39 of 40 against 40 of 40 open-loop, a favourable but real trade, and less injected charge per unit of correction is what matters at the electrode-tissue interface, where charge-density limits, electrode corrosion and tissue damage bound what can be delivered. The percentage is model-abstract; the direction, more effect per unit charge, is the transferable safety and longevity argument.

The bottom line

Inside the fitted models, several things are established: the disease connectivity signature is distributed rather than focal, a single most-deviant-site drive fails even at supra-physiological amplitude while a per-patient discriminant-aligned resonant drive reclassifies essentially all subjects, the pathology and therapy site sets are largely disjoint, and a causal closed-loop controller nearly matches open-loop efficacy, 39 of 40 versus 40 of 40, at about 44 percent lower dose. What is unproven is everything about contact with reality: that the model's counterfactuals are causally valid, that a physical stimulator would revert anything in tissue, and that a principle established in infraslow hemodynamics transfers to a microelectrode array reading spikes. What would confirm the transferable claim is a real closed-loop, biomarker-titrated stimulation on a cortical slice or organoid on a high-density array, showing that the electrode chosen for its effect on a decoded biomarker beats the electrode chosen for its deviation. What would break it is evidence that the discriminant-aligned advantage is an artifact of the reservoir's weakly-connected nodes, or that it depends on the hemodynamic timescale that spiking tissue does not share. For the array engineer, the durable and modality-independent takeaway is to build coverage and closed-loop biomarker control so the system can address a target defined by effect rather than by deviation, while stating plainly that the evidence so far is in-silico.