87,342 electrodes, 103,598 sites, and what a human atlas asks of array makers

What the work claims

Ho, Barth, Dister, Srivastava, and a large multi-institution team, with corresponding author Benjamin Rapoport, report a functional atlas of the human central nervous system built from high-resolution surface electrophysiology recordings in the first 70 patients implanted with an anatomically conformal, thin-film neural interface.¹ Rather than pursuing whole-brain recording with a single device in a single patient, they aggregate data across 86 array deployments, more than 100 anatomical target locations, and a range of behavioral tasks, then align everything to a common cortical surface framework. The atlas comprises 87,342 microelectrode channels that recorded from 103,598 unique microanatomic locations.

The claimed findings are structural and dynamic. Aggregated recordings reveal somatotopic organization in motor cortex, including digit representations with lateral-to-medial ordering consistent with classical homuncular maps, with transitions between adjacent digit representations occurring over spatial scales of roughly 0.8 to 2.0 mm. They show canonical frequency-band signatures, including movement-related beta-band suppression and focal high-gamma increases in the 70 to 200 Hz range during motor and speech behaviors. And they report task-modulated population dynamics across cortical, brainstem, and spinal cord regions, including structures that have historically been difficult to access at high electrophysiologic resolution. Real-time decoding demonstrations include upper-extremity motor cortical control and decoding of speech, facial movements, and facial expressions of emotion from ventrolateral cortex.

This is a primary results preprint, posted as new results and not yet peer reviewed. It is also, transparently, a paper about a commercial device: many authors are current or former employees or advisors of Precision Neuroscience Corporation and hold equity, and the corresponding author is the company co-founder. That does not invalidate the data, but it sets the framing.

How it works

The hardware is a thin-film surface microelectrode array embedding up to 1,024 microelectrodes in a polymer substrate, designed to be conformal to exposed neural surfaces. The nominal electrode pitch is 400 µm, which gives a spatial sampling density of 625 electrodes per square centimetre, more than 600 times denser than standard 10 mm-pitch clinical electrocorticography arrays. The recording platform uses head-stage amplification, multiplexing, and high-bandwidth acquisition to capture broadband signals from all contacts simultaneously.¹

The manufacturing and quality-control numbers matter as much as the geometry. Across the cohort, electrode impedances were screened intraoperatively at 1 kHz, with contacts above 4 MΩ excluded from analysis. Of 69,254 included electrodes across 77 arrays, the mean impedance was 1.17 ± 0.50 MΩ and the average functional channel yield was 89.1% ± 12.2%. In plain terms: about one channel in eleven failed or was discarded, and the spread between arrays was wide.

The aggregation method is the other half of the instrument. Array positions were localized with combinations of CT-imaged radiopaque markers, intraoperative photography, and neuronavigation, then projected onto individual pial surfaces and transformed to the fsaverage template space via surface-based spherical registration. In cases where only intraoperative photography and neuronavigation were available, electrode positions were registered to an example patient surface rather than an individual reconstruction. Behavioral tasks were adapted to clinical constraints, so task batteries were not uniform across subjects. Audio was captured at 44.1 kHz and synchronized to the neural data; for analysis, signals were downsampled to 2 kHz, line noise was removed with a 60 Hz notch and harmonics, and high-gamma was defined as 70 to 200 Hz, with the 70 to 150 Hz band included in decoding architectures.

Patients were enrolled between 19 April 2023 and 9 January 2026 under IRB-approved protocols with written informed consent. Recordings were mostly transient, intraoperative or short-term postoperative; in a subset of 12 cases arrays remained in place for limited postoperative recording sessions.

Where a skeptic should push

The most load-bearing assumption is that recordings from different patients, different surgeries, and different task sets can be pooled into a single atlas without the pooling itself creating the structure being claimed. The authors align anatomy carefully, but they cannot align state. Patients differ in pathology, medication, anesthesia exposure, and arousal, and the tasks were explicitly not uniform across subjects. Somatotopy and beta suppression are robust enough that this concern is unlikely to erase them, but finer claims about conserved population dynamics across the entire neuraxis deserve more skepticism.

Second, this is a surface method. It resolves the pial surface at sub-millimeter pitch but cannot see laminar structure or most subcortical volume, and it only reaches brainstem and spinal cord structures when a surgical exposure makes them transiently accessible. The phrase whole-brain interface should be read as a functional aggregation of accessible surfaces, not whole-volume recording.

Third, the data-generating hardware is a single vendor's device, and the authorship carries direct financial interest in that vendor. The impedance and yield distributions are therefore best read as one company's current manufacturing reality, not as a field benchmark. An 89.1% mean yield with 12.2% standard deviation across 77 arrays is respectable for conformal thin-film devices placed on a living, moving surface, but it also means roughly 7,600 of the screened channels were excluded by the authors' own criterion before any science happened.

Finally, the decoding results are feasibility demonstrations in small numbers of participants, performed under the time pressure of clinical care. They show the signal quality is sufficient, not that a clinical product exists.

What a human atlas asks of MEA hardware

The non-obvious implication is that the unit of value in large-scale electrophysiology has shifted. For forty years the MEA field's headline metric was electrodes per device. This paper's method works because its arrays were interchangeable enough, and their metadata clean enough, that 86 separate deployments could be fused into one coordinate system. The enabling specification was not the 1,024-electrode count. It was the impedance distribution held near 1.2 MΩ, the documented 1 kHz screening rule with a 4 MΩ cutoff, the published yield statistics, and the disciplined registration pipeline. Consistency is the feature; density is the advertisement.

For MEA and acquisition-chain designers, that inverts several priorities. Channel yield variance across arrays is no longer just a manufacturing yield problem; it is a sampling-bias problem for any atlas-scale consumer of the data, because lost channels are not randomly distributed across the surface. Per-array quality metadata, impedance screening at a fixed frequency, and calibration records become part of the scientific instrument, on the same footing as the amplifier noise floor. A lab that discards its per-array QC log is discarding data that downstream aggregation needs.

The spatial scale finding carries a direct design constraint. Adjacent digit representations transition over 0.8 to 2.0 mm, and the array resolves them at 400 µm pitch. Functional mesoscale structure in human cortex lives at the millimeter scale and below. Any electrode array, including organoid-scale devices, whose pitch or electrode size approaches that scale will blur exactly the features this atlas resolves. Conversely, resolution without coverage produces beautiful postage stamps; the atlas needed broad, conformable coverage of brainstem and spinal cord surfaces that rigid shank arrays cannot reach. Flexibility and registration, not just density, are what opened those frontier structures to high-resolution recording.

The opportunity for the instrumentation industry is a new product category: not the array, but the comparable array. Vendors that publish screening criteria, yield distributions, and stable geometry make their devices aggregatable, and aggregatability is what converts hardware sales into atlas-scale science. There is also a near-term practical opening in organoid and slice work, where the same conformal thin-film geometry and the same QC discipline apply directly, and where most labs still treat per-array impedance logs as an afterthought.

The threat cuts both ways. A field that aggregates across sites hands enormous structural power to whoever owns the coordinate frame and the metadata schema; an atlas built on one vendor's geometry and one company's registration pipeline is a quiet form of lock-in. And the data challenge the paper names is real: 87,342 channels of broadband signal per cohort is an acquisition, storage, and curation burden that most MEA facilities are not provisioned for. The instrument's price tag increasingly includes its data chain.

The bottom line

Established: conformal thin-film surface arrays with 400 µm pitch can be deployed across more than 100 neuroanatomical targets in 70 patients, survive intraoperative handling with about 89% functional yield, and resolve mesoscale functional organization consistent with classical maps. Asserted but less secure: that pooled cross-subject aggregation yields a stable, generalizable functional atlas of the whole human CNS, and that surface electrophysiology is a credible route toward a whole-brain interface rather than a complement to depth recording. What would confirm the approach is replication with a second, independent device and harmonized task batteries; what would break it is evidence that cross-subject pooling artifacts, rather than conserved organization, account for the reported population dynamics.

For MEA hardware, the durable lesson is that scale arrived through consistency. The next competitive spec sheet should lead with yield distribution, screening criteria, and registration metadata, because those are what turned 86 deployments into an atlas.