Long-term stability of cellular-resolution brain-computer interface recordings after stroke

TL;DR: A person with a large left-hemisphere stroke and chronic aphasia had four Utah arrays implanted in the intact right hemisphere, in regions mirroring her destroyed language network. Over 1,240 days and 111 sessions the recordings stayed put: unit yield, SNR, and the ability to track individual neurons across days were all comparable to what's been reported in structurally normal brains. This is the first evidence that the stroke-injured brain doesn't sabotage chronic single-neuron interfaces — a precondition for ever building a restorative BCI for the tens of millions of stroke survivors.

Why anyone doubted this would work

Every intracortical BCI success you've heard of — Willett's speech neuroprosthesis, Hochberg's robotic reach-and-grasp — has been in people whose brains are structurally normal. The damage is downstream: a severed spinal cord, dying motor neurons. The cortex under the electrode is healthy tissue.

Stroke breaks that assumption, and it does so in three ways that all threaten the electrode–tissue interface, even far from the lesion itself:

1. Mechanical. A large infarct causes the brain to atrophy and cerebrospinal fluid space to expand. More room to slosh means larger pulsatile brain motion with each heartbeat — and a Utah array is a rigid bed of 1.5 mm silicon needles that needs the surrounding neurons to stay in place relative to the tips. Micro-motion is the enemy of stable single-unit recording.

2. Inflammatory. Post-stroke neuroinflammation, glial activation, and vascular remodeling are not local — they spread across the whole brain. Chronic inflammation is exactly the process that encapsulates electrodes in glial scar and kills signal quality over months.

3. Functional. Even undamaged regions connected to the lesion get depressed (diaschisis), while others reorganize to compensate. So the neural activity you record might itself drift as the network rewires — a moving target for any decoder.

Any one of these could mean that even if you implant in pristine-looking tissue, the signal degrades or the neurons you're tracking today are gone next week. Nobody had tested it, because nobody had implanted microelectrode arrays in a stroke patient's brain this way.

The setup

The participant was a 51-year-old woman, six years post-stroke (an M1 middle-cerebral-artery occlusion), with non-fluent aphasia but largely preserved comprehension. Four 64-channel Utah arrays (256 electrodes total, one pedestal) went into right middle frontal, inferior frontal, supramarginal, and angular gyri. Recordings ran in blocks (up to 3/week) across six task datasets — language, working memory, mental calculation — with the participant responding by saccades or speech.

The clinical logic is worth noting because it's genuinely new: this is a cognitive BCI aimed at reconstructing word retrieval from surviving tissue, bypassing the destroyed language network entirely. Motor and speech BCIs decode intended movements from intact motor cortex; here there is no intact language cortex to decode from, so they go to the homolog. Whether that homolog carries decodable language content is a separate paper (ref. 15) — this paper only asks the prior question: can you record from it stably at all?

The measurement problem, and the clever bit

"Stable recording" has two levels. First, gross signal quality — impedance, unit counts, amplitude, SNR. That part is straightforward to report, and it held up (numbers below).

The harder, more interesting question is whether you're recording the same neurons over time. This matters because BCI decoders are trained on specific units' activity; if the population silently turns over, your decoder rots and you have to recalibrate constantly. The gold standard for tracking neurons across days is unitMatch, but it was built for high-density Neuropixels probes, where each neuron shows up on many channels and you get a rich spatial waveform "fingerprint."

A Utah array is the opposite: 400 µm between electrodes, so each unit appears on essentially one channel. No spatial fingerprint. So how do you know that a spike waveform on electrode 37 today is the same neuron as a similar waveform on electrode 37 three weeks ago, versus a different neuron that happens to look alike?

The authors' trick is to exploit the array geometry to build training labels for free. Because electrodes are 400 µm apart, two units recorded on different electrodes are almost certainly different neurons. And a single unit's waveform split across the first and second half of one session is, by construction, the same neuron. So:

• "Same" pairs = a unit's waveform in the first half vs. second half of a session.
• "Different" pairs = units on different electrodes.

They compute four similarity features between waveforms — shape (correlation of the averaged waveforms), slope (correlation of first derivatives), amplitude similarity, and width similarity — and train a per-electrode SVM to separate same from different. Decision scores become posterior match probabilities P(\text{match}). Within a session this classifier hits 96.6% cross-validated accuracy. They then raise the merge threshold from the default 0.5 to 0.85, buying precision (0.96 → 0.99, i.e. ~1% false matches) at a small cost in sensitivity (0.98 → 0.96) — a sensible choice when a false "this is the same neuron" is more damaging than a miss.

One more nice detail: they compare all 111×111 session pairs simultaneously, so a unit can vanish for several sessions and reappear later and still be linked ("intermittent tracking"). They contrast this with a stricter "continuous tracking" that forbids reappearance.

What they found

Signal quality held for nearly four years. Impedances dropped then plateaued (classic electrode-tissue equilibration). Unit yields were stable and regionally heterogeneous: MFG ~83, IFG ~65, SMG ~49 units/session; the angular gyrus array was low (~10) because it partially dislodged from cable torque right after surgery — yet even that array recovered, climbing from near-zero to consistent activity over hundreds of days. SNR was flat everywhere (~14–16 dB).

Neuron tracking matched non-lesioned brains. The decay of tracking probability with time was fit to a biexponential P(\text{track})_k = a e^{-bk} + c e^{-dk}, where k is days between sessions, a,c the fast/slow amplitudes and b,d the decay rates. The fitted parameters land right on top of the best prior human data (Downey et al. 2018) from structurally normal BCI participants:

Individual units were trackable across spans of 82.8 ± 1.8 days on average with intermittent tracking (3.2 days for the strict continuous variant). Mean consecutive-session tracking probability was 16.6%. Strikingly, short-term stability improved over the recording period (a significant positive effect of time-since-implant, t=2.81, p<0.001) — consistent with the electrode–tissue interface progressively stabilizing rather than degrading. The authors speculate a mature glial scar may actually anchor the electrodes and reduce micro-motion. That's the opposite of the usual "glial scar ruins recordings" story.

Independent functional validation. Because the classifier only sees waveforms, they checked whether tracked units also share features the classifier never used: inter-spike-interval distributions and spike-triggered population responses (how a neuron's firing co-varies with the overall population — a stable "signature"). Using a mixture decomposition, 51% (intermittent) / 63% (continuous) of tracked ISI correlations and 78% / 89% of stPR correlations matched the same-unit reference distribution. Stimulus responses matched far less (17% / 24%) — but crucially, similarity declined with inter-session interval (t=-15.8, p<0.001), which false matches wouldn't produce (they're interval-independent). So the lower stimulus-response stability reflects genuine changes in neural dynamics (or differing tasks), not misidentified neurons.

What changes if this holds, and what to doubt

If the finding generalizes, it removes a foundational objection to post-stroke intracortical BCIs: the worry that a globally inflamed, atrophying, rewiring brain simply can't host stable cellular-resolution recordings. It can, at least in intact contralesional cortex, at a stability level indistinguishable from the healthy-brain benchmark. Given that stroke dwarfs spinal-cord injury and ALS in prevalence, that's a meaningful expansion of who intracortical BCIs might one day serve. The unit-tracking code (adapted unitMatch for single-channel MEAs) is public and directly reusable by anyone running Utah arrays.

Now the caveats, which the paper is mostly honest about:

• n = 1. One participant, one lesion type, one implant. Stability could depend on this specific anatomy, atrophy pattern, or surgical outcome. The authors explicitly call for larger cohorts.
• Intact tissue, not damaged tissue. The arrays are in the contralesional hemisphere. This bounds feasibility for the homotopic-target strategy; it says nothing about recording from peri-lesional cortex, where local damage and vascular pathology would be far more aggressive. The authors flag this as the key open question.
• Stability ≠ usefulness. This paper demonstrates you can record stably — not that the right-hemisphere signal decodes language well enough to restore word retrieval. That's the companion claim (ref. 15) and the actual clinical payoff, not shown here.
• The stimulus-response result is a small yellow flag. Only ~17–24% of tracked units had stimulus responses consistent with same-unit references. The authors attribute this to genuine dynamics and differing tasks, and the interval-dependence argument is a fair defense — but it's a reminder that "same neuron by waveform" and "same functional response" aren't identical, which matters for decoder stability in practice.

The section most worth your time is the Unit tracking pipeline in Methods plus Figure 2 — the geometry-as-labels trick is the transferable methodological contribution — and the biexponential comparison in Results, where the head-to-head numbers against non-lesioned humans carry the paper's central claim.