Binita S. Ashar

See attached file(s)

The Unexplainable Case
Evidence Readiness and a Surgical Pilot for GenAI-Enabled Devices

Comment to FDA Docket No. FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback
October 17, 2026
Binita S. Ashar, MD, MBA, FACS, MAMSE
Founder and Principal, Clarity Surgical Advisors
Research Professor, Western Michigan University Homer Stryker M.D. School of Medicine
Clinical Assistant Professor, Carle Illinois College of Medicine, University of Illinois
Regulatory Venture Partner, Puma Ventures

Disclosure. I consult for medical device manufacturers and am a regulatory venture partner at Puma
Ventures, a medical device investor. I serve on the boards of the Society of Robotic Surgery and the
Institute for Surgical Excellence and take part in MDMA and American College of Surgeons AI groups.
These views are my own.

Ashar, FDA-2026-N-7874, page 1
Executive Summary
Every device in a connected surgical case can work as specified, and the case can still fail. I call this the
unexplainable case. It will happen more often as GenAI-enabled functions are added to connected
systems, and today no one can reliably reconstruct what happened (Figure 1).

Figure 1. The unexplainable case. The record sources in a connected surgical episode today, and the same sources
organized around a proposed minimum record, which a pilot would test.

I support FDA’s proposal to pair a risk-based framework with competency-based premarket evaluation,
and I offer two steps that build on it (Figure 2). First, before FDA accepts more premarket uncertainty in
exchange for postmarket evidence, it should confirm that the postmarket study and monitoring plan can
answer the open question. In parallel, FDA should support a voluntary surgical pilot to test whether a
minimum record helps explain how connected systems performed. Any broader minimum record
requirement should wait until the pilot shows value. Both steps fit within existing authorities.

Ashar, FDA-2026-N-7874, page 2
Figure 2. How the two proposals build on FDA’s proposed GenAI framework. The near-term step relies on devicelevel evidence where that evidence can answer the open question, and the pilot tests what is needed where it
cannot.

Recommendations
1. Before relying on postmarket evidence, confirm that the monitoring plan can answer a defined
question, with GenAI-specific criteria and a defined endpoint and exit.
2. Keep greater premarket evidence as the expectation where a resulting action cannot be reversed
and no downstream safeguard is available.
3. Define roles so that manufacturers, health systems, professional societies, and standards bodies
each supply the evidence they are best placed to provide, including clinician readiness.
4. Test a minimum record in a voluntary surgical pilot, and consider broader implementation only if it
shows value.
5. Prioritize MDDT qualification of GenAI evaluation tools, and encourage their development through
national challenges, to set common metrics.

Ashar, FDA-2026-N-7874, page 3
Response
Summary by FDA question
Section FDA question Recommendation
1 1 Consider whether an episode can be reconstructed when assessing risk
2 18, 19, 20 Confirm that postmarket monitoring can answer the defined question and
expect more premarket evidence where actions cannot be reversed
3 21 Define complementary roles, including clinician readiness
4 14, 15, 26 Evaluate human-AI teams against appropriate comparators and plan to
observe system-level performance
5 16 Prioritize qualified tools for GenAI evaluation
6 22 to 25 Use qualified tools within PCCPs and treat foundation model changes as rebenchmarking triggers
7 Cross-cutting Test a minimum record in a voluntary surgical pilot

Questions not addressed
This comment does not respond to Questions 2 through 13 or Question 17. Those cover informational
and patient-facing functions, benchmark design, clinical confirmation, synthetic data, and other model
architectures. They are important, and others are better positioned to answer them. I have focused on
what I know best as a surgeon and former CDRH office director, which is what happens when a GenAIenabled device is used in a connected procedure.

1. Reconstructability and a minimum record (Question 1)
FDA Question 1. Does the two-axis risk framework, organized around device activity and the consequence
of relying on an incorrect output, appropriately capture the dimensions most relevant to the risk of a GenAIenabled software function? If there are additional dimensions—such as the reversibility of a resulting action,
the availability of downstream safeguards, the time pressure of the deployment setting, or the traceability
of the output (i.e., to primary source materials)—that should be represented in a risk framework, please
describe and provide examples of how they should be represented.
Question 1 already lists reversibility and traceability as possible added dimensions. I recommend
applying traceability to the whole episode by asking whether what happened can be reconstructed
afterward. The minimum record mentioned in Figure 1 and further described below offers a practical
way to do this.
A function whose errors cannot be detected or traced to a cause carries more risk. This is especially true
for an agentic function that works across several devices before a clinician sees the result. In the
framework, a function whose output cannot be reconstructed should be treated as one or more steps
higher on the consequence axis. Reconstruction does not establish what caused an injury, but it may
show where the fix belongs. If a device fails in a situation it was never tested for, the gap is in premarket

Ashar, FDA-2026-N-7874, page 4
evidence. If the output was correct but misunderstood, the gap is in design or readiness. In a complaint
file, these two failures look the same.

A minimum record
As a starting point for a voluntary consensus standard, I propose a minimum record with six elements. It
is limited to what reconstruction needs and would be protected in line with FDA’s cybersecurity
expectations.
1. Time-aligned event timestamps across connected devices
2. Device, software, and model versions in use, including the GenAI configuration
3. AI inputs as received and outputs as displayed, including outputs not acted on
4. Clinician actions in response to outputs, and whether a person or an AI agent took each action
5. Key procedural events and device alerts
6. Linkage to outcomes and complications, where feasible
Fuller reconstruction would be triggered by serious complications, near misses, device faults, unusually
high override rates, and model changes, and by a sample of routine cases.
Example. A surgeon operates a robot from another site, with a GenAI overlay on the video. A
vessel is injured. The robot recorded the patient’s side, but not what the surgeon saw at the
remote console after network delay, with the overlay applied. That record is split among the
robotic device manufacturer, the network provider, the AI vendor, and two care teams across
two hospitals. Each can show its own part worked. The minimum record would put their
records on one timeline, so reviewers could see what the surgeon and bedside team saw and
when.

Most of the pieces already exist.[2][3][4] The question is no longer whether a surgical episode can be
recorded, but what minimum record is needed across technologies and who provides it. No one has
agreed on a minimum record.

2. Evidence readiness now (Questions 18, 19, and 20)
FDA Question 18. CDRH is considering whether it may be appropriate to accept greater premarket
uncertainty regarding a GenAI-enabled device’s benefit-risk profile through greater reliance on postmarket
monitoring. Under what conditions might such an approach be appropriate, and what characteristics of a
monitoring program would need to be in place to justify reduced premarket evidence? Are there device
types or risk profiles for which this approach would not be appropriate?
FDA Question 19. Please comment on the potential approaches to postmarket performance evaluation,
including periodic re-benchmarking, sample-based clinician review, and performance degradation
monitoring. What additional approaches should CDRH consider, and how should the cadence and triggering
events for reassessment be determined?
FDA Question 20. Please comment on whether the proposed approaches to postmarket monitoring can be
facilitated by machine-based supervisory agents. What considerations, including the evaluation and
reliability of the supervisory agent itself, should CDRH take into account for such an approach?

Ashar, FDA-2026-N-7874, page 5
Accepting more premarket uncertainty makes sense only when postmarket monitoring can resolve it. As
an office director, I saw postmarket commitments that were slow to start, hard to enroll, and sometimes
never finished. Even when they finished, the results often described a version of the device that was no
longer on the market or an outdated standard of care. GenAI models will change faster still. Monitoring
that starts after a problem appears cannot recover data that was never recorded.
CDRH should start now, within existing reviews, by confirming that the sponsor’s postmarket plan does
five things (Figure 2). It should identify the model and configuration behind each output, sample routine
cases as well as adverse events, use independent clinical adjudication, set triggers for reassessment, and
answer a defined question with a defined endpoint and exit.
This works when the open question is specific to one device, the sponsor’s records can answer it, and
errors can be caught downstream. A plan that meets these checks under those conditions gives FDA a
sound basis to authorize earlier, and a sponsor’s record of meeting them could inform FDA’s confidence
in that sponsor’s later plans. When these conditions are not met, or whenever an action cannot be
reversed and no downstream safeguard is available, FDA should continue to expect greater premarket
evidence.
Complaints capture only the errors someone notices. When a clinician accepts a plausible but wrong AI
output, there is often no complaint. Postmarket monitoring therefore has to include review of routine
cases, and not only reported events. Higher-risk devices need these checks more often, and any warning
sign should lead to a new benchmark test.
Example. An AI tool tells a surgeon that a structure is safe to divide. It is wrong, but the output
looks confident, and the surgeon proceeds. Three days later the patient has a complication, and the
team attributes it to the patient’s anatomy or to technique. No one connects it to the AI, so no one
reports it.

If this happens at hundreds of hospitals, monitoring still shows nothing, and a device with zero
complaints looks safe.

Supervisory agents could help monitoring scale, but only if they can be audited themselves. They should
log their own version, inputs, and outputs, and their sensitivity should be regularly checked against
clinician review.

3. Shared responsibility requires shared evidence (Question 21)
FDA Question 21. What roles might clinicians, healthcare institutions, professional societies, standardssetting bodies, and other stakeholders appropriately play in postmarket monitoring, and how can these
roles be structured without diffusing manufacturer accountability?
FDA, manufacturers, health systems, and professional societies each answer different questions about
AI in surgery, and each needs evidence to do so. In a recent paper I describe System Performance
Readiness, the ability to reconstruct a connected episode and learn from it across products and sites.[5]
A qualified reviewer should be able to answer the five questions in Figure 1 for any episode, whatever
the manufacturer, without access to model internals or other proprietary content, which can be
recorded as a version reference.

Ashar, FDA-2026-N-7874, page 6
Manufacturers remain accountable for device performance and for truthful and not misleading labeling.
For a device’s own records, FDA can use existing tools, such as special controls or conditions of approval,
to expect that evidence be reconstructable. Manufacturers should provide device records in
interoperable formats, with version identifiers and advance notice of model changes. Health systems
govern conditions of use and grant competency-based, vendor-neutral privileges.[6] Professional
societies and certifying boards set competency expectations, and standards bodies define common
record specifications. Even so, no single manufacturer can assemble a complete record across other
companies’ devices. One model the pilot could test is for the hospital to flag the case and preserve the
records, for each manufacturer to supply its part, and for a qualified reviewer, such as a Patient Safety
Organization or a professional society program, to assemble them. This would spare small hospitals from
doing this alone and could extend the confidentiality protections given to patient safety work.[7] It
would not change anyone’s reporting obligations, including a manufacturer’s duty to report adverse
events to FDA.
Competency applies to the surgeon as well as the device. No one has defined when a surgeon is ready to
use an AI feature, including when to trust it, when to override it, and how to tell when it is wrong. The
profession can define this through workplace standards[8] and competency-based training,[9] and
hospitals can treat a GenAI-enabled function as a new privilege that begins with supervised use.

4. Human-AI teams and connected systems (Questions 14, 15, and 26)
FDA Question 14. For open-ended device outputs, how should performance comparators and acceptance
criteria be selected? When a panel of qualified clinicians serves as the comparator, how should the
applicable standard (for example, the standard of care versus the performance of a median clinician in
practice) be defined and justified? Should generalist or specialist physicians be used as a performance
standard? When should human-AI team performance, rather than the device operating alone, serve as the
basis for evaluation?
FDA Question 15. Are there ways in which performance might be assessed relative to the care, technology,
or course of action likely to occur in the absence of the device, rather than to the comparators described in
this section? What approaches might be used to identify and justify a comparator such as unaided clinical
judgment, delayed specialist review, or no intervention?
FDA Question 26. Are there additional considerations that inform the premarket and postmarket evaluation
of agentic GenAI-enabled devices, beyond those applicable to non-agentic GenAI-enabled devices? How
should the elevated risk associated with autonomous multi-step action, tool use, and reduced opportunity
for human review be reflected in acceptance criteria and oversight?
For many surgical functions, the best comparator is a proficient surgeon doing the same task without
the device, because that shows whether care is better with the device. FDA should ask sponsors to
define proficiency, with input from professional societies, since learning curves for robotic and AIassisted procedures are long. If the device will be used by specialists, the comparator should be a
specialist.
The surgeon and the device should be tested together, with surgeons at different levels of experience. A
tool that helps a junior surgeon may mislead a senior one. Simulation lets sponsors test this before
authorization. Simulated cases, which could themselves be qualified as MDDTs, should deliberately
include a range of wrong AI outputs, so the study can test to failure and measure how often surgeons
catch them.

Ashar, FDA-2026-N-7874, page 7
For agentic functions, acceptance criteria should be stricter when there are fewer points where a person
can stop an action. As FDA’s proposed agentic competencies already contemplate, an agent should
generally pause for the surgeon’s confirmation before an action that cannot be undone, such as firing a
stapler or dividing a structure. Because problems after authorization will often show up at the level of
the whole system, reconstructability should be part of postmarket oversight.

5. Qualified tools (Question 16)
FDA Question 16. Should independent third parties be involved in some or all aspects of a competencybased approach, including device benchmarking and clinical confirmation? If so, in what ways might
qualified, independent third-party participation contribute to this approach, and what qualifications and
independence criteria should apply? Are there aspects of the assessment for which third-party involvement
would be impractical or inadvisable? What safeguards or program design features would be critical to
prevent third-party participation from limiting competition or preventing innovation?
Today each sponsor builds or picks its own evaluation tools, so results are hard to compare. The Medical
Device Development Tool program can fix that by qualifying benchmarks developed by third parties.[10]
CDRH should make GenAI benchmarking and adjudication tools a qualification priority and, with NIH,
NSF, and ARPA-H, encourage their development through national challenges. Because qualification is
public and any sponsor may use a qualified tool, it widens access to evaluation rather than limiting
competition. A tool one sponsor builds for its own device usually stays proprietary, so each competitor
builds another and no one can compare results across devices. Qualification breaks down those silos, so
every sponsor is measured the same way. FDA has already qualified a cleared device this way, Apple’s
AFib History Feature.[11]

6. Change control (Questions 22, 23, 24, and 25)
FDA Question 22. CDRH envisions that a manufacturer’s premarket competency-based assessment might
serve as a baseline against which post-deployment modifications could be re-evaluated. How might the
extent of re-benchmarking or other evidence be scaled to the nature and expected impact of a given
modification? For example, are there categories of change that might not significantly affect safety or
effectiveness of a GenAI-enabled device, or might be appropriately managed within a sponsor’s quality
management system versus requiring FDA premarket review and authorization before implementation? Are
there categories of changes that might be appropriate for inclusion in a PCCP?
FDA Question 23. How might PCCP concepts or other change-control approaches be adapted for GenAIenabled devices when the nature or scope of future modifications cannot be fully prespecified?
FDA Question 24. For devices built on third-party foundation models, changes to the underlying model may
be initiated by the third-party model developer rather than the device manufacturer. How can a
manufacturer detect, evaluate, and respond to such changes in a timely manner? What mechanisms—for
example, contractual, technical, or through a PCCP—could provide reasonable assurance that third-party
developer-initiated changes do not compromise the safety or effectiveness of the device?
FDA Question 25. Would voluntary Foundation Model MAFs be practical for foundation model developers to
provide and to keep current, and what content might they include? Given that participation would be
voluntary and that model developers may have limited incentive to disclose safety-relevant information,
what would make such a program sufficiently useful for premarket review? Are there alternative
mechanisms CDRH should consider for obtaining information about underlying foundation models?

Ashar, FDA-2026-N-7874, page 8
GenAI devices will change often, and not every change can be planned in advance. An MDDT helps here
too. Used as the test method in a PCCP, it keeps the measuring stick the same while the device changes.
[12] The size of the re-test should match the size of the change. A configuration change needs a targeted
check. A new underlying model needs the full baseline again.
Some GenAI devices run on an AI model made by another company, which can change that model
without the device maker’s involvement.[1] The minimum record should capture the model version on
every case, and the device maker should watch its own records for changes in output. A model change
or a shift in output should trigger re-testing. FDA’s proposed Foundation Model Master File could record
the model company’s promise to warn device makers first.[1]

7. Start with surgery, a voluntary pilot
Surgery is a practical place to begin. Procedures are discrete episodes, and operating rooms already
generate video and device logs. I encourage CDRH, working with ASTP/ONC, to support a small voluntary
pilot with professional societies, building on related feasibility work such as that proposed by the
Congress of Neurological Surgeons (FDA-2026-N-7874-0116), and using the minimum record described
in Section 1 as the record to test. Taking part should count toward the evidence-readiness check
described above, which gives manufacturers a reason to share their records. A first phase would use
existing records at willing sites, including community hospitals, to see whether a modest cross-system
record improves case review and whether performance differs across patient populations. Success could
be measured by the share of episodes reconstructed, completeness of the record, time to reconstruct,
agreement among reviewers, and whether product issues can be told apart from workflow or
connectivity issues. Aggregate findings should be published, and if the pilot does not show value, it
should be narrowed or stopped.
CDRH could begin by discussing evidence readiness and reconstruction at its December 2–3, 2026 public
workshop on robotic medical devices, naming reconstruction tools as an MDDT priority, and inviting
NEST or MDIC to scope a surgical pilot for 2027.

Conclusion
Clinicians need to trust that GenAI tools work, that mistakes will be caught early, and that we will learn
from them. Patients and families need to trust that they are getting the benefit of current tools, and that
when a case fails, it will be understood so the next patient is spared. A learning health system everyone
can trust is the goal, and these are first steps toward it.
FDA can close today’s evidence gaps and create system-level learning at the same time. It can confirm
now that postmarket evidence will answer the outstanding questions, and test in parallel whether a
minimum record, shared across manufacturers, health systems, and clinicians, helps explain how
connected systems perform.
Thank you for considering these comments. I would welcome the opportunity to work with CDRH,
professional societies, health systems, and manufacturers on a pilot.

Ashar, FDA-2026-N-7874, page 9
References
1. US FDA. Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback. August 2026. https://www.fda.gov/media/194242/download
2. Jung JJ, Jüni P, Lebovic G, Grantcharov T. First-year analysis of the operating room black box study. Ann Surg.
2020;271(1):122-127. https://doi.org/10.1097/SLA.0000000000002863
3. Garcia Vazquez A, Verde J, Hernandez Lara A, Mutter D, Swanstrom L. Consensus for operating room
multimodal data management, identifying research priorities for data-driven surgery. Ann Surg Open.
2024;5(3):e459. https://doi.org/10.1097/AS9.0000000000000459
4. IHE International. IT Infrastructure Technical Framework, Volume 1, Chapter 7: Consistent Time (CT)
Integration Profile. https://profiles.ihe.net/ITI/TF/Volume1/ch-7.html
5. Ashar BS. When every product works but the case fails, the need for system performance readiness in
connected surgery. J Robot Surg. Published online October 2026. doi:10.1007/s11701-026-03993-3
https://doi.org/10.1007/s11701-026-03993-3
6. Ashar BS, Selber JC. A competency-based privileging model for robotic platforms. JAMA Surg.
2026;161(7):659-660. https://doi.org/10.1001/jamasurg.2026.1350
7. Agency for Healthcare Research and Quality. Patient Safety Organization (PSO) Program and the Patient
Safety and Quality Improvement Act of 2005. https://pso.ahrq.gov/resources/psqia
8. Wood DE, Wolinsky PR, Dodgion CM, et al. Developing specialty-specific workplace standards for surgeons. J
Am Coll Surg. Published online March 3, 2026. https://doi.org/10.1097/XCS.0000000000001880
9. Stain SC, Ellison EC, Farmer DL, et al. The Blue Ribbon Committee II report and recommendations on surgical
education and training in the United States, 2024. Ann Surg. 2024;280(4):535-546.
https://doi.org/10.1097/SLA.0000000000006360
10. US FDA. Medical Device Development Tools (MDDT). https://www.fda.gov/medical-devices/medical-devicedevelopment-tools-mddt
11. US FDA. MDDT Summary of Evidence and Basis of Qualification: Apple Atrial Fibrillation History Feature.
May 2024. https://www.fda.gov/media/178230/download
12. US FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial
Intelligence-Enabled Device Software Functions. Final guidance. December 2024.
https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submissionrecommendations-predetermined-change-control-plan-artificial-intelligence

Ashar, FDA-2026-N-7874, page 10