Implantable Brain-Computer Interface Collaborative Community (iBCI-CC)

The Implantable Brain Computer Interface Collaborative Community (iBCI-CC) appreciates the opportunity to comment on FDA’s discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. Our community brings together researchers, clinicians, medical device manufacturers, patient advocacy groups, and people with lived experience to advance the development, safety and effectiveness evaluation, and access to implantable brain-computer interfaces that support greater independence for people with neurological conditions.
Implantable brain-computer interfaces (iBCIs) have demonstrated restoration of communication, movement, and sensation for people with paralysis, including amyotrophic lateral sclerosis (ALS) (Hochberg et al., 2012; Flesher et al., 2021; Card et al., 2024). Generative AI (GenAI) may enter these systems through language-assisted decoding, speech synthesis, reply suggestion, clinician-facing tools, and potential future uses involving agentic control. For example, an agentic system could translate a user’s request to pick up an object into a planned sequence of robotic-arm actions. Evaluation should also consider pretrained neural foundation models used within the decoder itself, as discussed in our response to Question 17.
We support the function-specific approach discussed by FDA’s Center for Devices and Radiological Health (CDRH), with benchmarking elements selected according to intended use and risk. The recommendations below explain how that approach can accommodate iBCIs, particularly functions that convey a user’s intent. These principles should apply across recording modalities, including intracortical, surface electrocorticography (ECoG), and endovascular systems, with evidence addressing the signal characteristics and stability of each modality and device.
We offer six recommendations, organized by topic rather than ranked priority:
1.Distinguish conveying user intent from exercising judgment (Questions 2, 5, 6, 8, 26).
2.Recognize fidelity to intent within effectiveness evaluation, using measures suited to intended use (Questions 1, 7, 9).
3.Recognize the user as the primary authority on intent, with clinical support where appropriate (Questions 10, 11, 14, 15).
4.Adapt evidence standards to small populations and individually calibrated systems (Questions 11, 12, 13, 17, 18).
5.Clarify when individual calibration is ordinary operation and when it is modification (Questions 17, 22, 23).
6.Protect the privacy of neural data and user communications and the continuity of function (Questions 19, 20, 24, 25).

We ask CDRH to clarify how GenAI evaluation would integrate with FDA’s May 2021 guidance on implanted BCI devices (FDA, 2021). Existing evidence should be accepted where it adequately addresses the relevant safety and effectiveness questions, with additional evidence directed to unresolved risks and interactions among the neural decoder, generative components, user interface, and connected effectors or stimulators, as applicable.

Detailed comments on each recommendation are in the attached file.

iBCI-CC Response to FDA Request for Feedback on the
Regulation of Generative AI-Enabled Medical Devices
October 19, 2026
Docket FDA-2026-N-7874

Summary
The Implantable Brain Computer Interface Collaborative Community (iBCI-CC) appreciates the
opportunity to comment on FDA’s discussion paper, Considerations for the Regulation of Generative AIEnabled Medical Devices. Our community brings together researchers, clinicians, medical device
manufacturers, patient advocacy groups, and people with lived experience to advance the
development, safety and effectiveness evaluation, and access to implantable brain-computer
interfaces that support greater independence for people with neurological conditions.
Implantable brain-computer interfaces (iBCIs) have demonstrated restoration of communication,
movement, and sensation for people with paralysis, including amyotrophic lateral sclerosis (ALS)
(Hochberg et al., 2012; Flesher et al., 2021; Card et al., 2024). Generative AI (GenAI) may enter these
systems through language-assisted decoding, speech synthesis, reply suggestion, clinician-facing
tools, and potential future uses involving agentic control. For example, an agentic system could
translate a user’s request to pick up an object into a planned sequence of robotic-arm actions.
Evaluation should also consider pretrained neural foundation models used within the decoder itself, as
discussed in our response to Question 17.
We support the function-specific approach discussed by FDA’s Center for Devices and Radiological
Health (CDRH), with benchmarking elements selected according to intended use and risk. The
recommendations below explain how that approach can accommodate iBCIs, particularly functions
that convey a user’s intent. These principles should apply across recording modalities, including
intracortical, surface electrocorticography (ECoG), and endovascular systems, with evidence
addressing the signal characteristics and stability of each modality and device.
We offer six recommendations, organized by topic rather than ranked priority:
1. Distinguish conveying user intent from exercising judgment (Questions 2, 5, 6, 8, 26).
2. Recognize fidelity to intent within effectiveness evaluation, using measures suited to intended use
(Questions 1, 7, 9).
3. Recognize the user as the primary authority on intent, with clinical support where appropriate
(Questions 10, 11, 14, 15).
4. Adapt evidence standards to small populations and individually calibrated systems (Questions 11,
12, 13, 17, 18).
5. Clarify when individual calibration is ordinary operation and when it is modification (Questions 17,
22, 23).
6. Protect the privacy of neural data and user communications and the continuity of function
(Questions 19, 20, 24, 25).

We ask CDRH to clarify how GenAI evaluation would integrate with FDA’s May 2021 guidance on
implanted BCI devices (FDA, 2021). Existing evidence should be accepted where it adequately
addresses the relevant safety and effectiveness questions, with additional evidence directed to
unresolved risks and interactions among the neural decoder, generative components, user interface,
and connected effectors or stimulators, as applicable.
1. Distinguish conveying intent from exercising judgment
Many of CDRH’s examples concern functions that recommend treatment, assess symptoms, or decide
whether to escalate care. An iBCI that transmits “I need help” performs a different function from an AI
that decides a user needs emergency care.
We recommend that CDRH distinguish functions that convey a user’s intended communication or
action from functions that independently generate clinical recommendations or make care decisions.
This distinction also applies to tools that transmit a person’s own words, such as dictation and
translation. When a user reports a symptom, asks for help, or gives consent, evaluation should assess
whether the device delivered that message accurately and reliably, including the reasonably
foreseeable consequences of getting it wrong. Features that generate and convey replies without user
acceptance, interpret clinical information, or autonomously act on the user’s behalf carry distinct risks.
Conveying a user’s intent can still carry substantial risk; the distinction concerns the function
performed and the degree of independent device judgment.
For multi-turn communication (Question 5), evaluation should assess errors that change meaning in
contexts relevant to intended use, such as a dropped “not” in a report of pain. For help requests
(Question 6), evaluation should distinguish conveying a user’s request from independently identifying
an emergency or deciding to escalate care. A user’s ability to communicate a request for help should
not, by itself, imply that the device provides a dedicated emergency communication function. Where
such a function is claimed, testing should address failure to transmit a genuine request, transmission
of an unintended request, and relevant failure conditions such as degraded signals or failure of
generative components.
Users should retain control over whether and how suggested communications or actions are adopted.
For communication iBCIs, evaluation should examine how reply suggestions and sentence completion
influence what a user communicates (Question 2). Relevant factors include how specific or
personalized the suggestions are, whether assistance is enabled by default, whether suggestions are
transmitted automatically or require user acceptance, and how readily the user can reject or correct
them. Such features can influence meaning even when they do not provide clinical advice. Assistance
that changes output without the user recognizing the change may affect the user’s sense of agency
(Haselager, 2013). Similar concerns arise when shared control shapes an intended movement, such as
a robotic-arm reach or a stimulator-driven step.
These considerations also apply to other augmentative and alternative communication (AAC) systems,
including those accessed through eye tracking, when they incorporate similar generative assistance.
For agentic functions, evaluation should address both whether the system correctly interprets the
user’s goal and whether it safely carries out the planned actions, as discussed in our response to
Question 26.

2. Recognize fidelity to intent within effectiveness evaluation
For functions that relay intent, effectiveness depends on whether output matches what the user meant
to say or do, and whether any assistance stays within what the user has chosen to allow. The
competency-based approach can accommodate these functions (Questions 7 and 8). The cliniciancredentialing analogy is useful for functions that exercise clinical judgment. For functions that relay a
user’s words or actions, a closer analogy is an interpreter, who is judged on how faithfully they convey
the user’s meaning. We refer to this as fidelity to intent.
Fidelity can be assessed within existing effectiveness evaluation (Question 9). Word error rate captures
only part of fidelity, because a single incorrect word can reverse a sentence’s meaning. We recommend
that the regulatory framework allow sponsors, working with intended users and relevant clinicians, to
select measures suited to intended use, such as:
• How often errors change meaning, such as a dropped negation or a wrong number
• Unintended output, produced when the user was not trying to communicate or initiate an action
• How much effort corrections take, and whether the user’s message gets through after correction
• For motor functions, unintended movements or selections, and accuracy in reaching targets
• For sensory functions, whether the user correctly perceives what stimulation is meant to convey,
such as the location or pressure of touch

We recommend that CDRH apply element E.4 according to the device’s intended use and performance
claims, distinguishing user-authored communication from device-authored warnings and prompts. For
user-authored communication, evaluation should prioritize whether the device conveys the intended
message accurately enough for its intended use, including the consequences of meaning-changing
errors. Measures such as preservation of tone may provide useful additional information without
necessarily serving as minimum acceptance criteria for every device.
Device-authored warnings, confirmation prompts, and uncertainty displays should be assessed
separately for clarity and accessibility to intended users.
Where technically feasible and informative, comparing performance with and without generative
assistance can help characterize its benefits and limitations (Question 1). For communication
functions, both outputs should be assessed against the intended message, where known, rather than
treating the unassisted decoder output as the reference standard. Such comparisons can identify
improvements in communication as well as errors introduced by generative assistance.

3. Recognize the user as the primary authority on intent with clinical support where
appropriate
The user is best placed to confirm what they intended (Question 11). We recommend that CDRH accept
structured user confirmation, with clinician support where appropriate, when the confirmation method
is sufficiently reliable for the outcome and risk being evaluated. If the user confirms through the same
iBCI, a decoding error could affect the confirmation too. Sponsors should develop and validate
confirmation procedures that address this possibility without requiring a second means of movement
or speech that some users lack. For example, sponsors could evaluate a simple yes/no selection that
does not use generative text completion, where feasible. Validation should establish whether this
method reduces confirmation errors, including errors shared with the original decoding process.
Relevant clinical support may include speech-language pathology or physical therapy, depending on
the function being evaluated.
For points of comparison (Questions 14 and 15), useful options include:
• The user’s intended message or action, where it is known in advance (as in prompted tasks)
• The user’s current assistive method, such as eye-tracking or a letter board, reflecting what happens
without the device
• The same decoder without generative assistance, which shows what the generative component
adds
Prompted tasks can provide a known reference for assessing decoding accuracy. Evaluation during
self-directed use may provide complementary evidence about practical benefit, correction burden, and
performance in everyday settings, while recognizing that the intended message may not be
independently verifiable. The balance between these approaches and their acceptance criteria should
reflect the device’s intended use and risk; fidelity to intent should not imply error-free communication.
For some users, the alternative to an iBCI is slower or less reliable assistive technology, or no reliable
means of communication or control. Benefit-risk assessment should compare iBCI performance with
the user’s available alternative, including differences in speed, reliability, and control. For user-facing
functions, the relevant human-AI team is the user and the device working together.
Individual calibration and specialized intended uses may warrant sponsor-developed benchmarks.
Independent review can assess study design, scoring, and analysis without displacing the user’s
authority over their intended meaning (Question 16). We recommend that CDRH allow sponsors to
establish, through co-design with intended users and relevant clinical specialists, how benchmark
performance relates to everyday, self-directed use (Question 10). Card et al. (2024) report both
prompted-task and conversational performance, illustrating settings in which this relationship can be
examined.

4. Adapt evidence standards to small populations and individual calibration
iBCI studies typically enroll few participants, but each participant can engage in repeated device
assessments across many sessions (Question 12). Repeated sessions can characterize reliability
within participants, but do not by themselves establish generalizability to new users. Statistical
analyses should account for repeated observations from the same people, for example through mixedeffects models where appropriate, and should report per-user and across-user results separately.
When benchmarking and clinical data are combined, each observation should be counted once, and
results should be pooled only when endpoints, relevant functional characteristics, and conditions are
compatible. Different diagnoses should not, by themselves, preclude pooling when participants meet
common functional criteria relevant to intended use.
For pretrained neural models, CDRH should consider evidence concerning transfer from adjacent
populations, including intraoperative and epilepsy-monitoring recordings, when sponsors justify its
relevance to the intended users, recording modality, tasks, and deployment conditions. Pretraining
data may support model development, while generalization should be demonstrated through
independent evaluation with appropriate separation between training and testing. Evidence from
adjacent populations should complement, and should not automatically substitute for, evidence in the
intended population (Questions 12 and 17).
Offline testing with previously recorded neural data allows models to be compared using the same
inputs (Question 11). Such testing cannot fully capture how users respond during live use to changes in
feedback, speed, correction behavior, or control. The need for additional evidence following a
modification should depend on its expected effect on safety and effectiveness and whether existing
data and appropriate non-clinical testing adequately address that effect. A targeted prospective study
may be appropriate when material effects cannot be adequately assessed by other suitable methods.
Neural signals are highly individual and methods for generating realistic synthetic neural data are still
developing (Question 13). Sponsors should justify that synthetic data reflects real users, with evidence
proportionate to how much the evaluation relies on it. For example, test sentences written by a
language model may favor the same common phrasing as the model being tested, obscuring errors.
Benchmark sets should reflect the device’s intended use. For communication functions, this includes
rare words, names, and, where appropriate, user-specific vocabulary.
For some devices, it may be reasonable to address certain remaining questions about performance
through postmarket monitoring rather than requiring all evidence before authorization (Question 18).
FDA’s existing guidance on uncertainty in benefit-risk determinations provides relevant context for
applicable submission pathways (FDA, 2019). The approach should depend on the questions that
remain, the risks involved, and whether monitoring can identify problems in time to protect users.
Monitoring could include comparison with each user’s established performance, predefined criteria for
investigating deterioration, accessible reporting, and a timely plan to correct problems or restore
function. Greater reliance on postmarket evidence would be inappropriate when a serious failure could
cause harm before it could reasonably be detected and addressed, particularly when the user lacks an
adequate fallback.

5. Clarify individual calibration versus modification
Decoder calibration is routine in iBCI use. We ask CDRH to clarify three categories (Question 22),
consistent with the applicable change-assessment framework and FDA’s guidance on predetermined
change control plans (PCCPs) (FDA, 2025):

Category Clarification requested

Routine calibration within an authorized When this constitutes ordinary device operation rather than
operating process a modification

Planned updates to adaptive algorithms or
When an authorized PCCP can cover them
generative components

Other changes not covered by an authorized How change assessment and submission requirements
operating process or PCCP apply
For a pretrained neural foundation model, we recommend that CDRH distinguish per-user fine-tuning
within a validated, authorized operating process from retraining the shared model backbone on new
pooled data (Questions 17, 22, and 23). The former may be treated as routine calibration when it stays
within that process’s defined limits and acceptance criteria. Retraining the shared backbone is a
modification that may be addressed through an authorized PCCP when the change and its validation
are covered by that plan. The characterization should depend on the scope and effects of the change,
rather than the label “fine-tuning” alone.
Changes to language or voice synthesis models affect how a user’s words are formed, and a PCCP
could define them as a distinct modification category. Interface changes that affect confirmation,
cancellation, or uncertainty displays may warrant assessment, even when decoded content is
unchanged.
For changes whose details cannot be fully prespecified, CDRH could explore whether defined limits on
permissible changes, fidelity metrics, per-user thresholds, and escalation procedures provide
adequate control (Question 23). For instance, speech iBCIs have used synthesized voices modeled on
users’ prior speech (Card et al., 2024), so model changes can affect how a user sounds. Changecontrol plans should address advance notification or confirmation of changes to a user’s synthesized
voice.

6. Protect neural data, user communications, and maintain continuity of function
Privacy protections should extend to raw neural recordings and derived data, as well as users’
communications. Collection, access, and retention should be proportionate to their stated purposes,
and secondary use should be consistent with informed consent and applicable law, with safeguards
addressing re-identification and potential inferences beyond the output the user chose to
communicate. Appropriately consented and protected pooled neural data may help reduce calibration
burden and improve performance for future users; those benefits should be evaluated rather than
assumed. We recommend that CDRH address data practices insofar as they affect device safety and
effectiveness, coordinate with applicable privacy frameworks, and avoid creating duplicative general
privacy requirements through this GenAI evaluation framework.
For communication iBCIs, real-world data include private conversations (Question 19). Routine
monitoring should rely on less intrusive means wherever they adequately assess performance and
should carry no presumption that manufacturers collect or review the substance of users’
communications. Suitable approaches include minimized telemetry, user-controlled settings that limit
recording or sharing of communication content while preserving necessary safety monitoring,
accessible user reports, periodic structured tasks, and governed investigation of selected events.
Safety investigations and required reporting should use the minimum information necessary, with
access to communication content justified for the specific purpose and governed by relevant
permissions and safeguards. Automated supervisory agents used for monitoring should operate under
the same content-minimization limits as human review (Question 20). Their ability to detect relevant
failures should also be evaluated, including failures they may share with the model being monitored.
Changes in neural recordings over time may affect the performance of GenAI-assisted iBCIs. Evaluation
should consider whether acceptable aggregate output performance could mask reduced fidelity to user
intent. Where appropriate, sponsors should compare performance with and without generative
assistance, test relevant degraded-input conditions, and assess interactions among decoding,
generative, and stimulation components. Postmarket monitoring and change assessment should
address these risks using methods suited to the device’s architecture and intended use (Questions 19,
20, 22, and 24).
Reassessment should be triggered by relevant performance changes or modifications, with periodic
intervals justified by risk and the adequacy of ongoing monitoring. Routine calibration within an
authorized operating process should not, by itself, trigger reassessment.
For an implanted device, loss of a software dependency can mean loss of function (Question 24).
Neurotechnology has a documented history of device abandonment (Drew, 2022; Kouvas et al., 2025).
Good practice for continuity planning may include version pinning, testing before adoption of new
versions, and a validated fallback or recovery strategy where feasible. Manufacturers could seek
agreements with model providers for advance notice of model changes, access to prior versions for a
defined period, and reasonable deprecation timelines. These are recommended practices, not
proposed mandatory contract terms. They involve trade-offs among continuity, security, ongoing
support, and available computing resources.
Benchmarking and continuity planning should address the actual deployment configuration, whether
inference occurs on-device, in the cloud, or through a combination of the two. For real-time
communication and motor control, relevant measures include end-to-end latency, variation in latency,
behavior during loss or degradation of connectivity, and recovery after interruption. Testing should
establish whether delayed outputs remain valid and whether fallback or safe-stop behavior is adequate
for the intended use. Cybersecurity assessment should address the device and its connected services
using applicable FDA recommendations and, for devices that meet the statutory definition of a cyber
device, section 524B of the Federal Food, Drug, and Cosmetic Act (FDA, 2026).
We recommend that CDRH explore voluntary Foundation Model Device Master Files (MAFs) that
include information relevant to device safety and effectiveness (Question 25). For communication
iBCIs, a model’s refusal behavior and content policies may apply to a user’s own speech. Filtering or
rewriting could alter what a user means to say, which bears directly on effectiveness. MAFs could
provide FDA with information on these behaviors. Sponsors also need sufficient information to evaluate
these effects and manage subsequent changes. Where model providers supply limited information,
sponsors should be able to rely on adequate alternative evidence, including characterization of a fixed
model version.
Brief Responses to Additional Questions
Question 1. Neural decoding is a signal processing function. Risk assessment should consider whether
users can recognize when an output differs from their intended communication or action, whether they
can correct it in a timely manner, and the consequences for anyone who relies on or is affected by that
output. The activity axis could also reflect users who supervise device actions themselves, often with
limited capacity to intervene. Reversibility and time pressure matter: a communication error may
prompt someone to act before the user can correct it, while an unintended movement may cause harm
before the user can stop it. Recognizing an error should therefore be assessed separately from the
ability to interrupt or correct it before harm occurs. The benefits of safeguards such as confirmation
steps should be weighed against the time, effort, and independence they cost the user.
Question 3. Heightened caution for patient-facing functions should rest on the function’s intended use,
reasonably foreseeable harms, and the user’s ability to assess and act on its outputs. Evaluation
should distinguish functions that convey a user’s own intent from those that provide clinical
information or recommendations to the user. Confidence indicators should be validated for the outputs
and conditions in which they are used, and all safeguards should be usable by people with severe motor
impairment.
Question 16. Independent third parties could maintain datasets held separately from sponsor
development data. Governance should address consent scope, re-identification risk, and cross-border
transfer. Particular care is needed in single-participant and small studies, where participants may be
more readily identifiable. Consistent informed-consent practices are relevant to this governance (Young
et al., 2026). Shared iBCI benchmarks may be candidates for Medical Device Development Tool (MDDT)
qualification, and the iBCI-CC could provide a forum for developing them. Third parties could also
support individual calibration or independently assess calibration methods and validation procedures,
where feasible, provided they have appropriate expertise and access to necessary device and userspecific information. Their practical role may be limited by the individualized implant, signals, and
calibration workflow.
Question 17. CDRH should clarify how the competency-based approach applies when a pretrained
neural foundation model serves as the decoder itself or is integrated into a multimodal system. Selfsupervised pretraining or use of a foundation-model architecture should not, by itself, determine
whether a function is treated as generative AI; evaluation should address the deployed function and its
capabilities. Relevant considerations include transfer across people, recording modalities, and tasks;
per-user calibration; and performance under changes in signal quality. Clinical-knowledge or
conversational benchmarks may be inapplicable to a decoder whose intended function is
communication or movement control. Such functions should instead be evaluated using appropriate
measures of decoding performance, fidelity, and control. Where generative assistance cannot be
separated from the decoder, sponsors should justify alternative comparisons or component analyses
rather than being required to test an unavailable unassisted configuration. Recommendations 4 and 5
address supporting evidence and changes to these models.
Question 26. An iBCI function may qualify as agentic when the user communicates a goal and the
system autonomously plans and executes multiple steps to achieve it. That planning is a form of
judgment and should be evaluated as such, consistent with Recommendation 1. Two risks apply to iBCI
users: the user’s task goal may be incorrectly decoded, and the system may act faster than the user can
intervene. Acceptance criteria (under element A.1) should therefore measure intent decoding and task
execution separately. Oversight checkpoints should reflect each action’s consequence, speed, and
reversibility, and the user’s ability and time required to recognize, interrupt, or correct an action. For
movement functions, evaluation should demonstrate how unsafe actions are prevented or interrupted
when decoding degrades or timely user intervention is not feasible.
Conclusion
The iBCI-CC shares CDRH’s goal of timely access to safe and effective devices. Function-specific
evaluation grounded in fidelity to user intent serves that goal. The iBCI-CC welcomes continued
collaboration with CDRH.
References
Card NS, Wairagkar M, Iacobacci C, et al. An Accurate and Rapidly Calibrating Speech Neuroprosthesis. N Engl J
Med. 2024;391(7):609-618. doi:10.1056/NEJMoa2314132.
Drew L. Abandoned: the human cost of neurotechnology failure. Nature. Published December 6, 2022.
doi:10.1038/d41586-022-03810-5.
Flesher SN, Downey JE, Weiss JM, et al. A brain-computer interface that evokes tactile sensations improves
robotic arm control. Science. 2021;372(6544):831-836. doi:10.1126/science.abd0380.
Haselager P. Did I do that? Brain-computer interfacing and the sense of agency. Minds Mach. 2013;23(3):405-418.
doi:10.1007/s11023-012-9298-7.
Hochberg LR, Bacher D, Jarosiewicz B, et al. Reach and grasp by people with tetraplegia using a neurally
controlled robotic arm. Nature. 2012;485(7398):372-375. doi:10.1038/nature11076.
Kouvas G, Coenen C, Hommrich D, Stieglitz T, Elger BS, Jotterand F. Building resilience when neural implants are
abandoned. Nat Med. 2025;31(9):2850-2853. doi:10.1038/s41591-025-03772-x.
U.S. Food and Drug Administration. Consideration of Uncertainty in Making Benefit-Risk Determinations
in Medical Device Premarket Approvals, De Novo Classifications, and Humanitarian Device
Exemptions. Guidance for Industry and Food and Drug Administration Staff. August 2019.
U.S. Food and Drug Administration. Implanted Brain-Computer Interface (BCI) Devices for Patients with Paralysis
or Amputation: Non-clinical Testing and Clinical Considerations. Guidance for Industry and Food and Drug
Administration Staff. May 20, 2021.
U.S. Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change
Control Plan for Artificial Intelligence-Enabled Device Software Functions. Guidance for Industry and Food and
Drug Administration Staff. August 18, 2025.
U.S. Food and Drug Administration. Cybersecurity in Medical Devices: Quality Management System
Considerations and Content of Premarket Submissions. Guidance for Industry and Food and Drug
Administration Staff. February 2026.
Young MJ, Sandbrink JD, Cabrera LY, et al. Unifying consent standards for implantable brain-computer
interfaces. Device. Published online August 11, 2026. doi:10.1016/j.device.2026.101271.