Recording an ECAP through your own stimulus artifact

What the work claims

The proposal argues that deep brain stimulation (DBS) for Parkinson's disease should be tuned in space and time to a patient's own moment-by-moment state, rather than set once and left open loop.¹ Three aims carry the argument: characterize how spatiotemporal stimulation parameters shape information transmission through basal-ganglia motor subcircuits at single-cell, ensemble and network levels; develop a Bayesian Dual Adaptive Control algorithm that relates stimulation settings to electrically-evoked compound action potentials and maps those onto the four cardinal motor signs (bradykinesia, rigidity, tremor and postural instability); and use ground-truth high-density microelectrode array recordings at the DBS site to validate the pathway-activation models that the controller relies on.¹

This is a competitive renewal in its tenth funded year, so it is a mature program rather than a speculative one. The bold move is not that DBS can be adaptive; several groups and one or two commercial devices already close loops on oscillatory biomarkers. The bold move is the choice of feedback signal. An electrically-evoked compound action potential (ECAP) is the synchronous, stimulus-locked response of the axon population that a pulse recruits, its earliest components appearing on nearby contacts within roughly a millisecond of the pulse. Using it as the control variable places the entire therapy on the back of a measurement that is, by construction, the hardest kind to make.

How it works

Start with the biomarker, because the instrumentation follows from it. When a stimulation pulse depolarizes a bundle of axons, they fire nearly together; the summed extracellular signature of that volley is the ECAP. Its amplitude scales with how many fibers were recruited and its latency with how fast and how far they conduct, so an ECAP is a compact readout of which pathway a given electrode configuration actually activated. That is exactly the quantity the Bayesian controller needs. "Dual" adaptive control means the controller does two jobs at once: it regulates the system toward a therapeutic target while also probing it to reduce its own uncertainty about the patient-specific model. In plain terms, it deliberately spends some settings on learning, not just on treating.

Now the hard part, which the abstract does not dwell on and which is my reading as an instrumentation engineer rather than the grant's stated method. To measure an ECAP you record on contacts next to the one that just delivered the pulse; sensing on the very same contact that stimulated is generally avoided, because that electrode is left maximally polarized, so it is the pathological worst case rather than the normal one. A clinical DBS pulse is large either way: devices run in constant-voltage mode at roughly 1 to 3.5 volts, up to about 5, or in constant-current mode at roughly 1 to 3 milliamps, with the other quantity following from an electrode impedance of order a kilohm. Against a microvolt-to-millivolt evoked response, that stimulus and the electrode polarization it leaves behind as the double layer discharges are, on an adjacent contact, some three to five orders of magnitude larger, reaching the full six only in the same-contact worst case, and it arrives in the same short window. A conventional biopotential front-end saturates, and its output crawls back along a slow recovery tail that sits directly on top of the fibers you most want to see, the fast ones that respond first, whose earliest components arrive within roughly a quarter to one millisecond of the pulse. So the ECAP is recoverable only if the acquisition chain does three things well: it presents an input range wide enough not to saturate; it settles or blanks fast enough to reopen a clean window, which is a genuine trade-off rather than a solved specification, because a blank of a few hundred microseconds can itself swallow the fastest fibers and only aggressive fast-settle front-ends reach the tens of microseconds that leave them intact; and it sits on an electrode whose impedance is stable enough that the artifact is reproducible from pulse to pulse on that same electrode, though not across electrodes, where the delivered current can vary severalfold from site to site even at an identical commanded amplitude.¹

The high-density microelectrode array in Aim 3 plays the reference role. A microelectrode array (MEA) is a grid of many recording sites; high density means pitch fine enough to resolve single units and local ensembles. By recording ground-truth activity at and around the DBS site, it lets the team check whether a given ECAP feature really corresponds to the pathway activation their models assume, rather than to something incidental. In effect the MEA calibrates the cheap, artifact-contaminated clinical measurement against a rich, clean laboratory one.

Where a skeptic should push

The single most load-bearing assumption is that an ECAP feature, recorded beside the stimulating contact, is a faithful and stable proxy for pathway activation that maps onto symptom relief, and further that a local, near-contact recruitment signal can be resolved onto individual motor signs at all, given those signs emerge from distributed basal-ganglia-thalamocortical circuits rather than from one fiber bundle. Push there and two cracks appear. First, an extracellular evoked response is dominated by the largest, most synchronous, nearest axons; smaller or less synchronized contributions are simply invisible, and the artifact-contaminated early window is precisely where the fastest fibers live. The feature is therefore a biased sample of recruitment, not a census. Second, and more corrosive for a controller, the electrode-tissue interface is non-stationary. Encapsulation and glial scarring raise interface impedance over weeks to months. Under constant-current stimulation that leaves the delivered charge fixed by design but drives up the compliance voltage the stimulator must supply and reshapes the artifact tail the amplifier must reject; under constant-voltage stimulation the delivered charge changes as well. Either way the artifact the recorder sees drifts, and a loop that reads an ECAP feature can then chase a slow change in its own electrode rather than a change in the brain, because from the measured response alone it cannot separate the two without independently tracking impedance. That failure mode is not hypothetical; it is the default behavior of any evoked-response loop that does not separately track interface impedance.

Two more cautions. The ground-truth mapping is established in animal recordings, and its transfer to the human basal ganglia is asserted rather than shown; the abstract is explicit that the MEA data validate model parameters, which is a modeling claim, not a clinical one. And "moment by moment" personalization implies a latency budget that is never stated: sense the ECAP, estimate state, update parameters, all inside a control interval, with artifact recovery eating the front of that budget. A proposal can assert closed-loop operation; only measured recovery times and loop latencies can demonstrate it. None are in the abstract, which is appropriate for a grant but leaves the central instrumentation question open.

Recording while the array is stimulating

Here is the non-obvious implication for anyone building array hardware to interface living tissue with silicon. This program only works if a single array stops being a passive listener and becomes a device that records its own stimulus. Strip away the Parkinson's framing and the grant is a live-tissue argument that concurrent stimulate-and-record on one array is worth engineering for, and that the enabling work is analog front-end design and electrode stability, not algorithm design. The controller, however elegant, is downstream of an amplifier that must not saturate and an electrode whose impedance must not wander.

The opportunity for organoid and neural-culture platforms is direct. The field is moving toward closed-loop stimulation, whether to probe circuits or to train tissue, and this grant hands over two reusable pieces: a principled control framework in dual adaptive control, and a concrete feedback signal in the evoked population response. High-density CMOS arrays that can switch some sites to stimulation while neighbors record are, in principle, already the right substrate. What this work makes explicit is the precondition: without an artifact-tolerant record path, a fast-settle or blanking front-end and per-site impedance that is characterized rather than assumed, the elegant loop has nothing clean to read. That is a specification an array vendor can design to.

The threats are just as concrete and worth naming plainly. The impedance-drift confound means a closed-loop array can be fooled by its own aging; per-electrode impedance therefore has to become a first-class, logged, provenance-carrying channel of data, not a calibration you do once and forget, or the therapeutic signal and an instrumentation artifact become inseparable. There is also a quieter obsolescence angle. If a coarse evoked feature carries most of the control-relevant information, the pressure for ever-higher channel counts weakens for the control task even as it strengthens for the ground-truth mapping task; value migrates from raw channel count toward artifact-clean, well-characterized channels. And the dual-use point is unavoidable: a front-end that can treat and probe a living network in the same loop is exactly what closed-loop conditioning of an organoid would require, so the artifact-recovery hardware that enables responsive therapy also lowers the barrier to training tissue, with the governance questions that raises.

The bottom line

The instrumentation claim buried in this neuroscience grant is the one to watch: that you can recover a usable evoked population response through your own stimulus artifact, on the same array, stably over time. That is an analog front-end and electrode-stability problem before it is a control problem, and the abstract, correctly for its genre, asserts the loop without measuring it. What would confirm the claim is concrete and checkable when results appear: reported ECAP signal-to-noise and artifact-recovery times from the high-density array data, and evidence that the ECAP-to-pathway mapping holds up as interface impedance drifts. What would break it is equally concrete: if the feature the controller leans on turns out to track electrode polarization and impedance change as much as it tracks neural recruitment. Until those numbers exist, treat the closed loop as a well-motivated hypothesis resting on an unproven measurement, and treat the front-end as the part that decides the outcome.