Sensing & Adaptive Neuromodulation

Evidence status. Device indications and MRI labeling change. The examples below reflect public FDA and manufacturer information available in 2026; always verify the exact model, leads, software, indication, and current labeling before use.

1.Signal, feature, state, and policy

The sensor records a raw signal. Processing extracts a feature such as band power, line length, or ECAP amplitude. A detector classifies a state or event. The control policy changes amplitude, pulse delivery, or another permitted parameter inside clinician-set limits. Latency, blanking, artifact rejection, sampling schedule, and fallback behavior are part of the therapy.

2.Biomarkers are contextual

Medication, movement, sleep, posture, impedance, lead location, stimulation artifact, and time after implantation can alter recordings. A beta peak associated with bradykinesia and rigidity in Parkinson disease may be less informative for tremor, gait freezing, dyskinesia, or mood. Patient-specific calibration is not a one-time act; drift and state transitions require surveillance.

3.BrainSense and adaptive DBS

Medtronic Percept systems can sense local field potentials from implanted DBS leads. BrainSense tools support survey, streaming, event-linked recordings, and chronic trends depending on device and software. The approved implementation is Medtronic BrainSense Adaptive on the Percept platform, CE marked in January 2025 and approved by FDA on February 20, 2025 (announced February 24) as PMA supplement P960009/S478, a design-change supplement that added adaptive stimulation as an optional programming feature to the deep brain stimulation system already on the market, under the bilateral STN or GPi indication for levodopa-responsive Parkinson disease of at least four years' duration with symptoms inadequately controlled by medication. It uses an LFP-based control signal to move amplitude between a clinician-set lower and upper limit according to a selected patient-specific LFP feature in the alpha-beta range. At the time of writing it is the only closed-loop DBS system with US approval; sensing capability on another platform is not the same thing as an approved adaptive therapy.

The therapeutic idea is to reduce stimulation when the biomarker suggests less need and increase it when pathologic activity returns. The clinician still chooses contacts, frequency, pulse width, amplitude limits, thresholds, and safety settings. The loop is constrained control, not unsupervised programming. Timescale depends on the algorithm and programmed settings: single-threshold control can respond rapidly, whereas dual-threshold control commonly uses slower ramps to follow changes in clinical state. Neither mode should be represented as a validated detector or abortive treatment for individual freezing or tremor episodes.

The published chronic programming workflows start from an optimized conventional program and preserve a clinically effective fallback. Candidacy assumes four things: a stable, well-titrated conventional program already in place; a sensing-compatible contact and frequency configuration; a clinically interpretable recording with cardiac and movement artifact excluded or adequately controlled; and an identifiable patient-specific beta peak. Each of these fails often enough to matter. In the published ADAPT-START series, 20 patients were evaluated and 9 started adaptive stimulation (5 remained on chronic adaptive DBS, 3 were still in optimization, and 1 reverted to conventional DBS). Among those not started, three were excluded for cardiac artifact in the LFP and three because their therapeutic settings were incompatible with sensing: two required the most dorsal contacts bilaterally, because sensing and stimulation contacts must meet the system's montage constraints, and one required stimulation above the sensing frequency ceiling (180 Hz with SenSight leads, 185 Hz with 3389 leads). Not every Percept patient can use adaptive stimulation. Say so before promising the feature, and record which prerequisite failed, because a later revision or reprogramming can change the answer.

4.RNS: detect and respond

The NeuroPace RNS System records electrocorticography from one or two leads at seizure foci, applies programmed detection tools, and delivers brief responsive stimulation. Stored events support longitudinal review and iterative programming. It is approved for adults with partial-onset seizures localized to no more than two foci, refractory to at least two antiseizure medications, with frequent disabling seizures; it is not a general-purpose psychiatric or pain implant. The two-focus limit is a labeled clinical-selection criterion; the limited lead coverage also constrains what the device can observe and stimulate.

Detection is a therapeutic proxy, not a seizure classifier, and the gap is wider than that phrase suggests. Most patients receive hundreds to thousands of brief stimulations a day, far more than any plausible seizure count, so the great majority of therapy is delivered in the interictal state. Read the stored data accordingly: detection counts depend on the programmed features and may include interictal activity or artifact; long episodes become a useful electrographic-seizure proxy only after patient-specific validation against stored ECoGs; and the published outcome figures still rest on patient diaries rather than on device counts. Clinicians tune sensitivity, specificity, therapies, and storage within finite device constraints.

Investigational uses of responsive intracranial stimulation should remain clearly separated from approved epilepsy practice, and the current example is instructive. The randomized NAUTILUS trial in idiopathic generalized epilepsy did not meet its prespecified primary effectiveness endpoint. Its 18-month follow-up reported a 76.8% median reduction in generalized tonic-clonic seizures from baseline; that longer-term comparison does not establish blinded superiority. NeuroPace reported on July 28, 2026 that FDA considered the indication-expansion supplement not approvable in its current form and requested additional clinical evidence. This indication remains investigational in the United States at the time of writing.

5.ECAP-based closed-loop SCS

An evoked compound action potential is the neural response to stimulation. Commercial ECAP feedback is no longer limited to one platform: Saluda Evoke (P190002) and Medtronic Inceptiv (PMA supplement P840001/S512, approved April 2024) both provide FDA-approved closed-loop implementations, with system-specific leads, sensing schemes, control policies, and neural-response targets. The controller adjusts stimulation toward a prescribed neural-response target and can compensate for posture- and movement-related changes in the electrode-cord relationship.

ECAP amplitude estimates recruited neural activity; it is not a direct pain measurement. The controlled variable is spinal-cord activation, with clinical benefit tested separately. Do not assume that evidence for one device, lead geometry, or control policy transfers to another: the double-blind randomized comparison of closed-loop against open-loop stimulation, carried to 36 months, is Evoke's, while Inceptiv was approved on its own pivotal program. ECAP control also presupposes a measurable evoked dorsal-column response under the system's own sensing and stimulation scheme, which is a reason closed-loop control is a property of a therapy architecture rather than a feature that can be added to any waveform.

6.Build a biomarker dossier

Document lead/contact anatomy, signal montage, recording state, medication, symptoms, artifacts, feature extraction, threshold logic, stimulation limits, and the clinical outcome expected to change. Preserve baseline conventional programming so the adaptive strategy has a fair comparator.

Evaluate concordance across clinic ratings, patient events, wearable or diary data, and chronic recordings. A visible peak that does not track the intended symptom should not drive therapy merely because the device can display it.

7.Failure modes

Stimulation artifact can mimic or obscure the feature. Thresholds can chatter near a boundary. A loop optimized for one symptom can worsen another. Stored events can be biased by limited memory or patient-trigger behavior. Firmware or lead changes can invalidate earlier calibration. MRI conditionality belongs in the same list: the conditions are specific to generator model, leads, software, and sometimes to the sensing configuration itself, so check the manufacturer's current MRI guidance against the exact implanted model rather than against the platform name.

8.Consent and data stewardship

Explain what is recorded, when recordings leave the device, who can review them, and the difference between research access and routine care. Patients should understand that sensing does not mean continuous human monitoring and that a recorded physiologic event may not map neatly to lived symptoms. Where recordings go is now a live question rather than a hypothetical: FDA approved migration of the RNS patient data management system from an on-site server to a cloud service in June 2026 (P100026/S100). FDA also approved the optional ECoG Assistant feature in May 2026 (P100026/S101) to highlight recordings for clinician review. The actual deployed software, access controls, data storage, retention, and automated review aids belong in the consent conversation and the program's data-governance record; automated aids do not replace clinical review.

At a program level, version the algorithm settings, archive programming reports, track battery impact, and link changes to outcomes. Adaptive therapy is a longitudinal experiment within a safety envelope.