Northwell Health (Filippo Filicori)

Please see attached file

Comments on FDA Docket FDA-2026-N-7874
Intraoperative Performance Analytics Laboratory (IPAL), Northwell Health, and Columbia
University
Considerations for the Regulation of Generative AI-Enabled Medical Devices: responses to selected
discussion questions, as they apply to surgery and procedural care

About the commenters
These comments are submitted by the Intraoperative Performance Analytics Laboratory (IPAL) in the
Department of Surgery at Lenox Hill Hospital, Northwell Health, New York, together with its engineering
collaborators in the Department of Electrical Engineering at Columbia University. We appreciate the
opportunity to comment on the discussion paper on generative AI-enabled medical devices issued by
the Center for Devices and Radiological Health [1].

IPAL is led by Filippo Filicori, MD, a minimally invasive and bariatric surgeon and program director of the
minimally invasive and bariatric surgery fellowship, Professor of Surgery and System Chief for Surgical
Innovation at Northwell Health, the largest healthcare provider in the northeast US. It partners with
Zoran Kostić, PhD, Professor of Engineering at Columbia University, and his group. The laboratory brings
together practicing surgeons, surgical research fellows and residents, electrical engineers and data
scientists. It studies how the data generated during an operation (the video, the robot's record of
instrument and camera motion, and the outcome) can be used to measure surgical performance and to
make operations safer.

IPAL has worked on these questions for eight years and deployed the first GPU in the operating room for
real time inference in the US [6]. Its first studies of operations performed from 2018 onward, rated
surgical skill from operative video. The laboratory then turned to the motion data that surgical robots
record and, with Columbia, to models that analyze operative video automatically and in real time. Along
the way it has helped shape the field's shared standards and resources. Its members led the SAGES
Delphi consensus on a standard taxonomy of surgical gestures [2], co-authored the report of the SAGES
Critical View of Safety Challenge [3] and is the clinical leader of Nvidia’s Open-H-Embodiment, described
as the largest open dataset of medical robotic video with synchronized kinematics to date, drawn from
more than 50 institutions [4]. Dr. Filicori chairs the SAGES Informatics Committee and co-chairs the
Surgical Disruptive Technology Summit, an annual international meeting on innovation in surgery and
the Society of Robotic Surgery (SRS) AI committee.

With Columbia, the laboratory built a dataset of 209 robotic inguinal hernia repairs from eight surgeons
and models that recognize the phases of the operation, the best of them about 85% accurate on
validation videos [5], and then ran phase recognition in real time on an edge computing device, with one
prediction per second and a processing latency of about 250 ms [6]. Later work has assessed robotic
suturing automatically from video in dry-lab simulation [7], tracked instrument tips in three dimensions
from a single endoscopic camera [8], trained video-generating models on 1,836 clips of robotic suturing
as a first step toward surgical world models [9] and contributed to real-time three-dimensional

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 1
reconstruction of deforming tissue [10]. In the robot's own record of instrument and camera motion, 31
of 41 measures taken during dissection of the gallbladder from the liver correlated with the surgeon's
experience [11], similar measures could benchmark trainees [12], the angle of the stapler at its first
firing was associated with obstructive complications after sleeve gastrectomy [13], and the smoothness
of motion varied with the difficulty of the case in ways that may signal a surgeon who is struggling [14].
In video-based assessment, blinded crowd-sourced raters found no difference in skill between women
and men [15], several components of rated skill tracked surgeon experience [16], rated skill predicted
weight loss after sleeve gastrectomy [17] but did not explain a novice's learning curve [18], and
experienced surgeons watching operative video were no better than chance at predicting which
operations would be followed by a complication [19].

That work is the basis of the comments that follow. It has shown us what can and cannot be measured
from operative video and from the robot's own record, how hard it is in practice to record, synchronize
and annotate operative data, and where the models now being built, from phase recognition to models
that generate surgical video, are heading. This comment is submitted by the laboratory, their
institutions and its collaborators.

Summary of responses
FDA's framework is a sound starting point, but it was written for decisions made outside the operating
room. During a procedure the device's output is visual or mechanical, the clinician who receives it is
already operating, and acting on a wrong output can cause harm within seconds that cannot be undone.
Our main recommendations, question by question, are these.

• Question 1. Name the reversibility of the enabled action and the time available to act as explicit
dimensions of risk, and judge consequence at the highest-consequence moment of the procedure
rather than on average. Ask sponsors to show how the design handles irreversible steps and limited
time.
• Question 2. Characterize directiveness by the action an output makes available at the moment it is
shown. Treat timing, display location and audience as modifiers, do not treat disclaimers as
reducing directiveness, and ask where the interface makes the surgeon verify an output. That
deliberate friction is most needed in real-time navigation.
• Question 4. In surgery the distinction that matters is experience. Define intended users by
experience as well as specialty, test with the least experienced of them and with deliberately
incorrect outputs, and support research on the clinician's ability to appraise AI output critically, for
which experience is only a proxy.
• Question 5. The procedure is the surgical counterpart of a conversation. Evaluate devices over
complete procedures, including the decision points that matter most, and classify each by the most
directive function it can reach.
• Question 6. Report missed and false escalations per procedure, weigh them by reversibility, and
treat alert burden and display clutter as design risks to be tested with users.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 2
 • Question 9. Add perceptual and real-time competencies, including latency, robustness and behavior
when the view is inadequate. Because benchmarking has no users in it, add a step in which
intended users meet deliberately wrong outputs in simulation before clinical confirmation.
• Question 10. Public surgical benchmarks are small and heavily reused. Use sequestered, multiinstitutional test sets divided by patient and institution, include difficult cases deliberately, and ask
for evidence that benchmark performance predicts performance in use.
• Question 11. Recognize simulation as a confirmation approach. Publish clear cutoffs, with worked
examples, for when studies of intraoperative AI need an investigational device exemption, and clear
guidance on which data from outside the United States are acceptable.
• Question 14. Make human–AI team performance the primary endpoint for intraoperative
inference, establish model accuracy, team performance and patient outcome separately, and do
not accept explanation features in place of testing with users.
• Question 15. The unaided comparator moves as skills erode. Plan periodic assessment with the
assistance off, for which the robot's kinematic record offers a practical measure in robotic
procedures, and test device-failure scenarios before authorization.
• Question 16. Use professional societies and independent auditors for sequestered datasets,
adjudication and user testing in simulation, but not to certify whole devices.
• Question 17. Treat any device with a generative model anywhere in its pipeline, including
annotation and synthetic training data, as within scope. Add physical competencies, and validate
generative simulators such as world models against physical experiments.
• Question 18. Accept more uncertainty before authorization only where something exists afterward
to detect what was left uncertain. Make a registry of adverse events and near misses the
cornerstone, and require time-stamped, exportable device logs.
• Questions 19 and 21. Monitor what clinicians do with outputs as well as the outputs themselves,
set cadence by volume of use, and keep the manufacturer accountable while societies supply
standards, adjudication and registries.
• Questions 22 to 24. Scale re-evaluation by what changed: the model, its imaging input or what the
user sees. Repeat user testing when the interface changes, and label the imaging configurations
with which a device was validated.
• Question 26. Develop requirements for autonomous surgical functions case by case for now, with
checkpoints before named irreversible actions, testing of takeover by the surgeon and a log of every
transfer of control.

General comments
These comments concern artificial intelligence used during operations and other procedures. That
includes computer-vision overlays on the laparoscopic or robotic image, intraoperative navigation and
augmented reality, computer-aided detection and diagnosis during endoscopy, and robotic systems that
take over part of a task. We have answered the discussion questions on which surgeons and engineers
working on procedural AI have something specific to contribute and have left the others, such as those

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 3
on statistical methods and on master files for foundation models, to respondents better placed to
address them.

We think FDA is right to distinguish technologies built with traditional machine-learning techniques from
those built on generative AI. The former produce bounded outputs that can be judged with conventional
accuracy metrics, and they behave fairly predictably. The latter reach their outputs through a chain of
reasoning that can at times be impenetrable, and they therefore pose a different category of risk.

Most of the discussion paper, however, is aimed at decisions made outside the operating room. Its
examples concern risk scores, medication, triage and orders, and none involves a procedure. We believe
the framework can be applied to procedural care, but it has to be translated first, and the Table at the
end of this document offers one translation. In a procedure the output of the device is usually visual or
mechanical. The clinician who receives it is already operating, often has only seconds in which to
respond, and is working on anatomy that moves. Acting on a wrong output has a physical consequence
that frequently cannot be undone. The same is true of therapeutic endoscopy and of catheter-based and
image-guided intervention.

We also expect generative AI to become central to the technologies that guide what physicians do
during procedures. Vision-language models are already being trained on surgical video and its narration
[20], foundation models have been proposed as a route to greater autonomy for surgical robots [21],
and world models are being built to simulate surgical scenes and to train robotic policies (Question 17).
Many of the intraoperative systems in use today rely on conventional computer vision, but those now in
development increasingly do not, and it is for them that this framework will matter most.

A brief account of where the field stands may be useful. Real-time guidance from the endoscopic image
is the most mature area. Deep-learning models can mark the zones in which dissection is or is not safe
during cholecystectomy [22,23], assess whether the critical view of safety has been achieved [24] and
help surgeons recognize autonomic nerves, an application that has been tested in a randomized pilot
trial [25]. Software of this kind has received marketing approval in Japan [26]. Intraoperative navigation,
in which preoperative imaging is registered to the patient and displayed over the operative view, is
established where the anatomy is rigid, but in abdominal soft-tissue surgery it remains limited by the
accuracy of registration [27,28]. A systematic review of image-guided laparoscopic liver surgery found
reported accuracy in the range of 8 to 15 mm and only one study that had examined clinical outcomes
[28].

Autonomy is at an earlier stage. The Smart Tissue Autonomous Robot, developed by a team that
included Dr. Axel Krieger, performed supervised autonomous suturing of bowel in pigs in open surgery
[29], and Dr. Krieger's group later did the same laparoscopically [30]. More recently his group has used
imitation learning to teach a surgical robot basic manipulation tasks [31] and a language-conditioned
model to carry out the clipping and dividing step of cholecystectomy on ex vivo tissue [32]. Dr. Michael
Yip's group at the University of California San Diego has contributed methods for the perception and
control on which such systems depend, among them a framework that maps a deforming surgical field
while tracking the instruments [33], autonomous suction to clear blood from the field [34] and robotic
assistance that maintains exposure during dissection [35]. Dr. Yip, Dr. Krieger and others have surveyed

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 4
the use of AI across medical robotics [36]. Federal research funding has followed. In 2026 ARPA-H
funded several teams, UC San Diego among them, to develop autonomous endovascular robots for
stroke [37].

In this context we see two use cases for which risk mitigation needs particular thought. The first is
augmented dexterity, a term proposed for systems in which a robot controls a surgical subtask under
the close supervision of a surgeon who is ready to take over [38]. Automated retraction, in which the
system holds, tensions or repositions tissue while the surgeon operates, is a good example [35,39]. The
second is intraoperative navigation, in which the system shows the surgeon where a hidden structure
lies or where to cut. We return to both in the responses below.

Our responses are consistent with the SAGES consensus recommendations on surgical video data
[40,41] and with three SAGES white papers, on defining digital surgery, on its regulation and on the
liability questions it raises [42–44].

Responses to the discussion questions
Each response begins with a short paraphrase of the FDA question, in italics.

Section IV. Assessment of risk

Question 1
Does the two-axis framework capture the dimensions most relevant to risk? Should others, such as
reversibility, downstream safeguards, time pressure or traceability, be represented?

We think the two axes are the right starting point. For AI used during a procedure, however, two of the
additional dimensions FDA mentions, the reversibility of the resulting action and the time available to
act, carry so much of the risk that they should be stated explicitly rather than treated as modifiers.

The paper's examples concern medication, triage and orders, where there is often time, and often
another clinician, between an incorrect output and harm. In an operating room the interval between an
output and an irreversible act is often seconds, and usually no one stands in between. A clip across a
misidentified bile duct, a divided vessel or energy applied to the bowel wall cannot be undone, and bile
duct injury remains the central safety concern of one of the most commonly performed abdominal
operations [45]. Unlike diagnostic tools, the technologies discussed here “guide real-time, irreversible
actions where errors can cause immediate and permanent patient harm” [43].

It follows that the same output carries different risk at different moments. An anatomy overlay shown
during the initial exposure simply informs the surgeon, whereas the identical overlay at the moment of
clipping bears on a step that cannot be taken back. Downstream safeguards are also thinner than
elsewhere in medicine, because the only safeguard between the output and the tissue is usually the
operator who received it. Traceability is limited as well, since a surgeon cannot audit a segmentation in
the middle of a dissection in the way a physician can check a cited source.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 5
Both dimensions can be addressed by design, and we would encourage FDA to look for evidence that a
sponsor has done so. For reversibility, the sponsor should identify the irreversible steps within the
intended use and show how the device behaves at each of them. Reasonable measures include
withholding the output when the model's confidence or the quality of the image is poor, so that the
surgeon is not shown a best guess, and asking the surgeon for explicit confirmation before an output is
used at such a step. The function should support, and not replace, the independent means of
confirmation that surgeons already use, such as cholangiography, fluorescence imaging and
intraoperative ultrasound. Systems that move instruments need limits on force and travel and an
immediate means of stopping the system and taking back control. Automated retraction is a useful
illustration: holding tissue is easily reversed, tearing it is not, and so the control that matters is a limit on
tension.

For time pressure, the aim is to move the output away from the last second. Guidance should be
available before the decision point and should not appear for the first time at it. The delay between the
image and the overlay should be specified, tested and kept short, and an overlay that can no longer be
trusted to match the image should be removed. The momentary pause before clipping, one of the
strategies in the SAGES Safe Cholecystectomy Program [46], is the manual equivalent, and a device can
be designed to prompt it. In navigation, checking the registration against visible landmarks before the
overlay is used near a critical structure serves the same purpose. Rehearsal in simulation, including
rehearsal of device failure, belongs here too.

We recommend:
• Treat the reversibility of the enabled action and the time available to act as explicit considerations
when an intraoperative function is placed on the framework. The SAGES white paper on regulating
digital surgery places any technology that can independently initiate irreversible actions in its
highest risk tier [43].
• Assess consequence at the highest-consequence moment of the intended use, rather than as an
average over the procedure.
• Ask sponsors to describe the design measures that address irreversible steps and limited time,
along the lines set out above.
• Include procedural examples in future versions of the framework. The Table at the end of this
document translates Section IV into surgical terms.

Question 2
How should FDA account for informational functions that vary in how strongly they direct the user toward an
action, and which characteristics of an output should modify risk?

We support the approach CDRH is considering, in which directiveness depends on the substance and
context of an output and not solely on its wording. In procedural AI what matters is what the output
asserts, when it appears and where it is displayed. A statement that an anatomical safety criterion has
or has not been met directs the surgeon's next action, whatever disclaimer accompanies it.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 6
Surgical overlays span FDA's continuum, often within a single product. Outlining a structure is
informational, and naming it already asserts an identification. Shading the zones where dissection is and
is not safe [22,23], or reporting whether the critical view of safety [47] has been achieved [24], tells the
surgeon whether to proceed. None of these uses the word “recommend,” and a label stating that an
output is for information only does not change its effect on a surgeon in the middle of a dissection. This
is the point CDRH raises about disclaimers in patient-facing functions.

FDA lists wording, specificity, personalization and context as possible modifiers. In procedures, context
needs to be spelled out. Timing matters, since the same output is more directive at a decision point than
during exposure. So does the display: an output drawn over the operative view is, in our view, harder to
set aside than one shown on a separate monitor beside it, which is how at least one system approved
abroad presents its output [26,48]. The audience matters too, because an output seen by the whole
team is not the same as one seen by the surgeon alone. CDRH's concern about measurement functions,
that the user cannot independently evaluate the basis for the output, applies here as well. A surgeon
looking at the same image the model analyzed is not an independent check, and a navigation overlay
registered from preoperative imaging cannot readily be verified by eye.

We would add a design consideration that the paper does not discuss, which is deliberate friction
between the AI and the surgeon. An interface can be built so that the surgeon has to stop and judge an
output before using it. The surgeon may be asked to commit to his or her own identification of a
structure before the overlay is revealed, the overlay may be shown on request instead of continuously,
or the surgeon may be asked to confirm that a navigation overlay still lines up with visible landmarks
before relying on it. In the human–computer interaction literature such measures are called cognitive
forcing functions. In a controlled experiment they reduced over-reliance on AI advice more than
explanations did, although users rated them less favorably [49]. Others have described the same idea as
frictional AI and have tested a protocol in which the clinician reads first and the AI gives a second
opinion [50]. A recent systematic review of automation bias came to a similar view: explanations by
themselves are often insufficient, and engaging the user in verification is the most practical intervention
[51]. Friction matters most in real-time navigation, where the overlay cannot be checked by eye and the
temptation is simply to follow it. It also costs time and attention, so it should be reserved for the few
points where an error cannot be undone (Question 6).

We recommend:
• Characterize directiveness by the action an output makes available at the moment it is shown.
• Treat timing within the procedure, display location and audience as modifiers of risk, and do not
treat disclaimers or labeling language as reducing directiveness.
• Ask sponsors of real-time guidance and navigation to describe where the interface requires the
surgeon to verify an output, and to show that this was tested with users.
• For predictability, publish worked procedural examples along the continuum (outline, label, safe
and unsafe zones, criterion met or not met).

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 7
Question 4
Should the distinction between generalist and specialist physicians inform the assessment of risk, and what
safeguards might mitigate that risk?

Yes, although in surgery the relevant distinction is experience more than specialty. A trainee and an
attending, an occasional and a high-volume operator, a surgeon new to a platform and one fluent in it
will each be exposed to a different risk by the same output.

Experience will affect risk, but perhaps not in the way one would first assume. Procedural AI is often
justified as extending expertise, for instance by helping a general surgeon in a low-volume hospital to
recognize what a specialist would. That user is also the one least able to tell when the system is wrong.
The clearest evidence comes from radiology. When an AI suggestion was deliberately incorrect, very
experienced radiologists were correct in 45.5% of cases, against 82% when the suggestion was correct,
and inexperienced readers fell from about 80% to under 20% [52]. Experience reduced the effect but did
not remove it. A similar pattern has been reported in dermatology, where faulty AI misled clinicians
across the whole range of experience [53], and in a small study of spine radiographs in which
automation bias was two and a half times larger in residents than in specialists [50]. The human factors
literature has long held that automation bias occurs in both naive and expert users and cannot be
prevented by training or instructions [54].

In surgery the evidence is thinner. Two studies have found that surgeons and trainees change their
identification of anatomical landmarks after seeing an AI overlay [55,56]. In the more recent of them, in
laparoscopic gastrectomy, the rate at which participants reported a change in their recognition of
landmarks in the infrapyloric area was 52% for junior trainees, 38% for senior trainees and 9% for
trainers [56]. Neither study was designed to test incorrect output. A study from our laboratory of how
surgeons at different levels of experience respond to AI output is in press [57]. To our knowledge no
published surgical study has yet measured what each group does when the overlay is wrong, and that is
the study the field most needs.

Experience protects, then, but only in part, and it is a proxy. What matters most, in our view, is the
clinician's ability to appraise the model's output critically, that is, to recognize when it should not be
followed. That ability can be measured, but only with benchmarks designed for the purpose, and the
necessary human factors research in procedural settings has barely begun.

Training adds a second issue. Residents operate under supervision, so an output visible to trainee and
attending is interpreted by two users of different experience at once, and a trainee who learns with
assistance from the start may not acquire the unaided judgment the device assumes its user has
(Question 15).

We recommend:
• Intended-user statements should specify experience (training level, procedural volume, experience
with the platform) as well as specialty.
• Where the labeled population spans very different levels of experience, evidence should be
stratified by experience or the indication narrowed.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 8
 • Evaluation should include the least experienced intended users and deliberately incorrect outputs,
so that the ability to reject a wrong output is measured directly.
• Postmarket data should record operator experience, so that what was expected of each group can
be checked in use.
• FDA should encourage, and where it can support, human factors research on the critical appraisal of
AI output in procedural settings.

Question 5
How should risk be assessed for devices whose behavior moves from non-directive to action-directing in the
course of an exchange, and how can intended use be characterized when behavior is emergent?

The closest counterpart in procedural care to a multi-turn conversation is the procedure itself. A device's
function changes as the operation progresses, and we think risk should be assessed across that course.

A system that recognizes surgical phases is informational while it follows progress during exposure.
Having recognized that dissection is complete, the same system may display whether a safety criterion
has been met at the moment the surgeon decides to clip, which is an action-directing output, and a
robotic platform may later use the same recognition to enable or withhold an instrument function.
Nothing in the model has changed, only the stage of the operation. An intended-use statement written
for a single function, such as “identifies anatomy,” does not capture this.

Two further features of procedures are relevant. Reliance builds up within a case, so that a surgeon who
has watched an overlay agree with his or her own judgment for forty minutes is likely to give it more
weight afterward. Errors also propagate, since a phase that is misrecognized early can mistime every
later output. Conversational surgical assistants built on language models will show both kinds of
migration, across turns and across the procedure.

We therefore suggest that devices be evaluated over complete procedures, including the highestconsequence decision points, rather than on isolated frames or clips, and that a device be classified by
the most directive function it can reach in its intended use.

Question 6
How should manufacturers characterize and weigh under-escalation and over-escalation?

Both directions of error matter in procedures, and they interact. By escalation we mean a prompt to the
surgeon to stop, obtain another view, change strategy, convert to open surgery or call for help.

Under-escalation is the familiar failure, in which the system stays silent as the dissection approaches the
wrong structure. The costs of over-escalation are easier to underestimate. Each unnecessary alert
interrupts a task that demands continuous attention, in an environment that is already distracting. One
study recorded a median of 138 auditory distractions per operation [58]. Unnecessary prompts to stop
or convert carry their own harms in longer operations, more open surgery and more imaging.

Two problems of interface design need to be considered here. One is alert fatigue. Clinicians override
medication safety alerts in 49% to 96% of cases [59], the likelihood that a clinician accepts a reminder

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 9
falls by about 30% with each additional reminder in the same encounter [60], and research on clinical
alarms has found that 72% to 99% of them are false [61]. The more alerts clinicians receive, and the
more of them are repeats, the less likely they are to act on any one, and over-escalation then turns into
under-escalation at the moment it counts.

The other is overcrowding of the display. Every overlay, label, confidence bar and warning competes
with the operative image for the surgeon's attention, and an augmented view can hide what lies
beneath it or draw attention away from it. In a cadaver study of endoscopic navigation, participants
using an augmented-reality display navigated more accurately, but only 1 of 15 noticed either of two
unexpected findings that were plainly in view, compared with 12 of 17 using the conventional display
[62]. A later study found that cognitive load, rather than the augmented-reality display itself, was the
significant factor in inattention blindness [63]. A crowded screen is in our view one source of that load,
and its effect should be tested.

Surgical practice already defines the appropriate escalations for a difficult cholecystectomy: additional
imaging, a momentary pause before clipping, a bail-out procedure, or help from another surgeon
[45,46]. A device can be evaluated on whether it prompts these when they are warranted and stays
quiet when they are not. Escalation also need not depend only on what the model sees in the image. In
our own work the smoothness of the surgeon's instrument motion varied with the difficulty of the case
and may mark the moments at which a surgeon is struggling [14], which suggests that what the operator
is doing could also inform when a device escalates.

We recommend:
• Report both missed-escalation and false-escalation rates per procedure as well as per frame.
• Weigh the two by reversibility. Where the action at stake is irreversible a higher false-alert rate may
be acceptable, but its burden must be measured.
• Treat alert burden and display clutter as design risks, by justifying each element shown during
critical steps and testing what users fail to notice with the display on.
• Evaluate the response to alerts at realistic alert volumes and over repeated use, and monitor alert
dismissal after deployment.

Section V. Premarket evaluation

Question 9
Would the proposed benchmarking structure provide adequate evidence? Are any elements missing or
wrongly categorized, and are there external standards that could be used?

For procedural AI we would make two changes. Competencies in perception, spatial accuracy and realtime performance should be added, because the elements in Appendix A are written for language and
numerical reasoning. And automation bias, which the paper rightly names in element E.4, should not be
listed as something device benchmarking can establish, because benchmarking as the paper describes it
has no users in it.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 10
A system that analyzes live procedural video needs competencies that the proposed elements do not
cover. These include accuracy of localization against expert annotation; robustness to smoke, bleeding,
a soiled lens, instruments crossing the view and deformation of tissue; the delay between image and
output; stability from one frame to the next; behavior when the view is inadequate, which is the
perceptual counterpart of “clinical deferral” in S.3; and, for navigation, accuracy of registration as the
anatomy moves. These properties are not yet routinely reported. In a scoping review of 113 studies of
surgical computer vision, 82% were retrospective, 12% described external validation and 28%
documented latency [64]. Our own experience shows why they should be. When we ran a phaserecognition model in real time on an edge computing device, it made one prediction per second with a
processing delay of about 250 ms, and the accuracy of its real-time predictions averaged 68.7% across
ten operations, ranging from 59.8% to 78.2% [6]. Figures of this kind, measured on the hardware that
will be used, tell a reviewer more than accuracy on recorded video.

On standards, navigation has a bench standard for positional accuracy [65] and usability engineering has
IEC 62366-1 [66]. The first illustrates the limit of benchmarking, since it measures what the system does
on a phantom rather than what a surgeon does when the overlay is wrong. E.4 names “the risk of
automation bias,” which SAGES has also identified as a risk of surgical AI [44], but the paper describes
benchmarking as “non-clinical” evaluation. Whether users accept an incorrect output can only be
measured with users.

We recommend:
• Define a set of perceptual and real-time competencies for devices that analyze procedural video or
guide instruments.
• Add a step between benchmarking and clinical confirmation in which intended users perform the
task in simulation with outputs that include deliberate errors, and acceptance of incorrect outputs
and time to decision are measured.
• State how competency evaluation relates to human factors validation, which the paper does not
mention.

Question 10
How should a sponsor establish that performance on a benchmark predicts real-world behavior, and what
role should sponsor-developed benchmarks play?

Public benchmarks in surgical AI are small, concentrated in a few centers and procedures, and heavily
reused. We do not think performance on them should gate device evaluation without evidence that it
predicts performance in the intended setting.

Much of the field's progress in surgical video analysis has been measured on a small number of public
datasets of laparoscopic cholecystectomy, the best known of which contains 80 videos from one
institution [67]. Reuse of the same data across the literature produces the saturation and contamination
FDA describes, and external validation remains the exception. It was described in 12% of studies in a
recent review [64]. When it is done the results can be sobering. In one study a model for instrument
segmentation lost accuracy even under slightly different surgical conditions [68].

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 11
Several problems are specific to procedural data. Frames from one operation are highly correlated, so
that dividing data by frame rather than by patient or institution inflates results. Reference standards are
expert judgments, and experts disagree, particularly on labels that require interpretation [40,69], so a
benchmark needs documented annotator qualifications and agreement. Selected cases are also cleaner
than practice. Datasets under-represent inflammation, bleeding, aberrant anatomy and poor
visualization, which are the conditions in which injuries occur and in which assistance would matter
most. The SAGES Critical View of Safety Challenge was designed with these problems in mind, with 1,000
videos curated by 54 institutions and each video assessed by three independent annotators under a
consensus protocol [3].

We recommend:
• Use sequestered, multi-institutional test sets held independently of developers (Question 16).
• Report the diversity of sites, surgeons, platforms and imaging systems, with data divided by patient
and institution.
• Include difficult cases deliberately and report performance on them separately.
• Ask for evidence of construct validity, meaning that benchmark performance tracks performance in
prospective use, and accept sponsor-developed benchmarks where annotation protocols, rater
agreement and independence of the test set are documented.

Question 11
How might a sponsor select a clinical confirmation approach proportionate to risk? Are some approaches
insufficient, and are there other methods to consider?

FDA's range of approaches translates well to procedures, with one substitution. “Standardized patient
interactions” have no procedural equivalent, and simulation takes their place. We also have two
requests. The first goes beyond the question as asked: there need to be clear cutoffs for when an
investigational device exemption is required. The second responds to CDRH's request for input on data
collected outside the United States: there need to be clear guidelines on which of those data are
acceptable for regulatory purposes.

Retrospective evaluation on recorded procedures establishes perceptual accuracy but not use. Shadow
deployment suits surgery well, because the system can run on the live video while its outputs are
withheld from the surgeon. An early clinical evaluation of real-time assistance for laparoscopic
cholecystectomy took this form, in three patients, on a secondary video signal [70]. Shadow deployment
tests a system under real conditions without affecting care, but it cannot show how clinicians respond.
Simulation, phantom, cadaveric and animal models take the place of standardized patients and allow
testing with deliberately incorrect outputs at no risk to patients (Question 9). Clinician adjudication of
recorded procedures fits surgery naturally, and prospective studies should follow staged evaluation
[71,72].

On the investigational device exemption, the paper notes that 21 CFR Parts 50, 56 and 812 “may apply”
but leaves the question out of scope. For intraoperative software this is a real obstacle. The regulation
defines a significant risk device in general terms [73]. The sponsor makes the first determination and

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 12
presents it to the institutional review board, and FDA is the final arbiter if asked [74]. In our experience
investigators and review boards are often unsure whether showing an investigational overlay to the
operating surgeon makes a study significant risk. A sponsor can request a study risk determination
through the Q-Submission program [75], but we are not aware of guidance specific to software that
displays information during a procedure. Published cutoffs would help. They might state, for example,
how shadow deployment with outputs withheld is treated, how display on a separate monitor that is
not used for intraoperative decisions is treated, and that display in the operative view at an irreversible
step, or any function that moves an instrument, requires an exemption.

On data from outside the United States, FDA already accepts foreign clinical data collected in
accordance with good clinical practice [76] and asks sponsors to explain how the data apply to the US
population and to US medical practice [77]. For procedural AI that second requirement is where the
difficulty lies, and it is not spelled out. More than one real-time surgical image-recognition program has
received marketing approval in Japan since 2024 [26], and a randomized trial has been conducted there
[25]. Whether such data can be applied to US practice depends on the imaging equipment and platform,
the operative technique, the case mix and the training of the surgeons. Sponsors need to know in
advance which of these FDA will examine and what evidence of comparability it will accept.

We recommend:
• Recognize simulation-based evaluation as a confirmation approach for procedural devices.
• Do not treat retrospective or shadow evaluation alone as sufficient for a function that presents an
inference at an irreversible step, since neither involves the user.
• Publish clear criteria, with worked examples, for when studies of intraoperative AI require an
investigational device exemption, covering at least shadow deployment, display-only studies and
functions that act.
• Publish guidance on the acceptability of procedural AI data from outside the United States, stating
what sponsors must show about the comparability of equipment, technique, case mix and operator
training.

Question 14
How should comparators and acceptance criteria be selected, should generalists or specialists set the
standard, and when should human–AI team performance be the basis for evaluation?

We consider this the most important question in the paper for procedural AI. For any function that
presents an inference to a clinician during a procedure, the performance of the clinician using the device
should be the primary basis for evaluation, and what finally matters is the clinical outcome for the
patient. Performance of the device alone is necessary but not sufficient.

Standalone accuracy does not predict what happens when a clinician uses an output. In mammography,
readers at every level of experience were far less accurate when an AI suggestion was deliberately
wrong than when it was right [52]. In colonoscopy, endoscopists followed a diagnostic system's advice
more often when it was correct than when it was not, and clinician and system together outperformed

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 13
either alone [78]. The interaction can help or harm, and neither result could have been predicted from
the accuracy of the device.

The research that would explain such differences is scant, especially in procedural settings, and more of
it is needed. A recent systematic review by Romeo and Conti examined 35 studies of automation bias in
human–AI collaboration published between 2015 and 2025, about a third of them in healthcare [51].
The authors found broad agreement that professional experience is the most important protection
against automation bias. They also report that users with only a moderate knowledge of AI were the
most likely to over-rely on it. High workload, time pressure and task complexity all increased overreliance, and these are the ordinary conditions of an operating room. Explanations of the AI's reasoning,
on which much hope has been placed, were often insufficient to improve decisions, and explanations
that were either too technical or too simple could reinforce misplaced trust, especially among less
experienced professionals with little knowledge of AI. What helped was engaging the user in
verification. Greater verification effort reduced complacency toward wrong recommendations, and the
authors point to protocols in which the clinician decides first and the AI serves as a second opinion. We
draw two conclusions for regulation. Explainability should not be accepted as a substitute for testing
with users, and the design of the interaction (Question 2) deserves as much scrutiny as the accuracy of
the model.

Nor does performance on a task guarantee benefit to the patient. In a multicenter randomized pilot trial
of intraoperative anatomical recognition in 90 patients, the system shortened the time surgeons took to
recognize autonomic nerves, but intraoperative and postoperative outcomes did not improve [25]. A
guideline panel that reviewed more than forty randomized trials of computer-aided detection in
colonoscopy found higher adenoma detection but very low certainty of evidence for the critical
outcomes, and made no recommendation for or against its use [79]. The accuracy of the model, the
performance of the team and the outcome for the patient have to be established separately.

On comparators, the median clinician in practice is the realistic reference for assistive tools, and
specialist performance is the right reference where a device is meant to extend specialist judgment to
non-specialists. We would avoid calling a regulatory comparator “the standard of care,” which is a legal
term decided case by case. The paper's other phrase, “a reference clinician,” is more accurate.

We recommend:
• Use human–AI team performance as the primary endpoint for intraoperative inference, with testing
that includes incorrect and uncertain outputs.
• Where outcome trials are impractical, accept process endpoints tied to a named clinical
consequence. Major bile duct injury, for example, occurs in roughly 0.15% to 0.4% of
cholecystectomies [45,80], which is too rare to power most trials.
• Encourage staged evaluation and reporting under IDEAL and DECIDE-AI [71,72], as SAGES has
recommended [43].
• Treat human factors research on human–AI interaction during procedures as a priority, and do not
accept explanation features in place of testing with users.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 14
Question 15
Might performance be assessed relative to the care likely to occur in the absence of the device, such as
unaided clinical judgment?

Yes. Comparison with unaided practice is essential for assistive procedural AI. The difficulty is that the
unaided comparator does not stay still. If routine use erodes unaided skill, both the benefit and the risk
of a device change over time, and a single premarket comparison will not show it.

The phenomenon is known as deskilling, and it is worth distinguishing its forms. Deskilling in the strict
sense is the loss of a skill a clinician once had. Never-skilling is the failure of a trainee to acquire the skill
in the first place because the task was assisted from the start. Mis-skilling is the learning of incorrect
behavior from the errors or biases of the AI [81]. The first two follow from a mechanism that the human
factors literature described decades ago. An operator who is asked only to monitor an automated
system loses practice at the task and is least prepared at the moment the system fails [82]. In aviation,
airline pilots tested in a simulator had largely kept their hand-flying skills, but had more frequent and
significant problems with the cognitive tasks of manual flight, such as keeping track of the aircraft's
position without the map display and deciding which navigational step came next [83]. Surgery has its
own example. As laparoscopy became routine, the mean number of open cases logged by general
surgery residents for the procedures studied fell by about a third between 2000 and 2018 [84], and the
open operation is what a surgeon falls back on when the minimally invasive approach fails.

For AI the most direct evidence of which we are aware comes from endoscopy. In an observational
study at four centers, after computer-aided detection entered routine use, adenoma detection in
colonoscopies performed without it fell from 28.4% to 22.4% [85,86]. The finding is observational and
comes from endoscopy rather than surgery. SAGES has nevertheless identified deskilling as a risk of
over-reliance on surgical AI [44], and no surveillance mechanism exists at present that would detect it.

The implications for evaluation are direct. A device assessed against today's unaided performance may
later be compared with a weaker baseline that it helped to create. The safety case for most assistive
devices rests on the clinician as the fallback when the device fails or is unavailable, and if unaided skill
declines so does the fallback. Trainees who learn a procedure with assistance from the outset may never
acquire the unaided baseline at all.

For robotic procedures there is already a practical way to watch for this. The robot records the motion
of every instrument, and objective performance indicators derived from that record distinguish
surgeons by experience [11] and can benchmark trainees on standardized tasks [12]. Tracked over time,
with and without assistance, such measures could show whether unaided technical performance is
changing.

We recommend:
• Where a device assists a perceptual or psychomotor task that is performed repeatedly, the
postmarket plan should include periodic assessment of performance with the assistance off, using
objective measures such as the robot's kinematic record where they are available.
• Device-unavailable and device-failure scenarios should be evaluated before authorization.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 15
 • Effects on unaided operator performance, including in trainees, should be recognized as a
foreseeable device-related effect, with a statement of who is responsible for monitoring them.
• Longitudinal effects on surgical skill should be identified as a research priority.

Question 16
Should independent third parties be involved in benchmarking and clinical confirmation, and with what
qualifications and safeguards?

In our view third parties will be indispensable. The evaluation described in the paper and in these
comments, with sequestered datasets, expert adjudication and testing with users across many
procedures and platforms, cannot be undertaken by FDA alone. It will need both professional societies
and companies. The Agency already relies on outside organizations in comparable ways, through
accredited testing laboratories, third-party review of certain 510(k) submissions and the qualification of
development tools [1,87,88].

Third parties would add most where procedural AI is weakest, in sequestered test data and qualified
adjudication, and professional societies and independent auditors are natural contributors to both. We
see three roles. One is custody of sequestered benchmark datasets, meaning multi-institutional
procedural video with expert annotation that is never released to developers, which addresses the
contamination described under Question 10. Another is adjudication. Deciding from video whether a
safety step was achieved requires procedural expertise, published criteria and a process for resolving
disagreement, all of which societies already maintain through their guidelines and educational
programs. The last is independent testing with users (Question 9), run by simulation centers that have
access to surgeons across levels of experience. FDA's Medical Device Development Tools program offers
one route, since a sequestered dataset or a validated assessment instrument could be proposed for
qualification [88]. Collaborative communities offer another [43], and FDA's list of such communities
already includes one on robotics, telesurgery and surgical AI [89].

Third parties should not replace the manufacturer's responsibility for its device or FDA's judgment on
authorization, and certifying whole devices would place societies in conflict with their members. On the
safeguards FDA asks about, access criteria should be open and published; fees should not exclude small
companies; adjudicators' industry relationships should be disclosed and managed, since many expert
surgeons have them; governance should be independent of any manufacturer; and a third party should
have no role in choosing among competing products.

We recommend:
• Use third parties, both professional societies and independent auditors, for sequestered datasets,
adjudication and independent testing with users, but not for the certification of whole devices.
• Qualify surgical datasets and assessment instruments through existing FDA programs.
Our laboratory and our collaborators at Columbia would welcome discussion with FDA about
contributing to such a program, particularly to the design of user testing in simulation and to benchmark
datasets of operative video and robot kinematic data.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 16
Question 17
Could the competency-based approach be applied to other architectures, such as multimodal vision-language
models and generative or predictive world models?

It should be, because surgical AI is moving to these architectures. For systems that perceive and act in
physical space, however, the approach needs physical and perceptual competencies, and it needs
simulation as its non-clinical tier.

Before turning to those, we want to make a point about scope. Generative models increasingly enter
surgical AI through the development pipeline as well as in the finished product. Large language models
are used to write the question-and-answer pairs on which surgical vision-language models are trained
[90], and diffusion models are used to synthesize annotated images for training segmentation models
[91]. Annotation itself is increasingly assisted, with foundation models proposing labels that human
annotators then correct [92]. A conventional model may then be trained by supervised learning on the
result. The device that reaches the operating room may contain no generative component at all, yet its
behavior depends on labels or data that a generative model produced, and some of those may be wrong
in ways that no one has examined. Our position is that any device whose pipeline includes a visionlanguage or other generative model, whether in annotation, in the creation of training or test data, or in
the deployed function, should be treated as a generative AI-enabled device for the purposes of this
framework, even where the model that finally acts on the image was trained by supervised learning. At a
minimum the sponsor should disclose where such models were used and show how their output was
checked.

The use case we have most in mind is augmented dexterity, and automated retraction in particular.
Retraction and exposure are repetitive tasks that are needed in almost every operation, and they have
been studied as targets for autonomy on research platforms for some years [35,39]. Dr. Krieger's group
has shown that a policy learned by imitation can lift tissue, pick up and hand over a needle and tie a knot
on a research surgical robot [31], and that a language-conditioned system can perform the clipping and
dividing step of cholecystectomy on ex vivo porcine tissue while taking corrections from the team in
natural language [32]. A limiting factor for such systems is data. Demonstrations on real robots are slow
and expensive to collect, and the rare events that matter most, such as bleeding or the tearing of tissue,
cannot be produced on demand.

World models are a possible answer. A world foundation model is trained on very large amounts of
video to predict how a scene will evolve, and can then be fine-tuned for a particular domain [93].
Cosmos 3, released in 2026, is an example. According to its developers it was trained on billions of
samples of text, image, video, sound and action trajectories, and it combines visual reasoning, world
generation and action prediction in a single model [94,95]. Surgical versions of such models already
exist. One has been used to generate synthetic demonstrations that improved a surgical robot policy
over training on real demonstrations alone [96]. Another, of which Dr. Krieger is a co-author, has been
used to evaluate robot policies in simulation, and on tabletop suture-pad tasks its results correlated
strongly with outcomes on the physical robot [97]. Our own group has fine-tuned video-generating
models on clips of robotic suturing exercises, as a first step toward world models of suturing [9]. A

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 17
retraction policy could in principle be trained and stress-tested through thousands of simulated cases,
including the failures, before it touches tissue. This is where FDA's question about synthetic data
produced by models of the same class as the device (Question 13) becomes concrete. A simulator that is
itself a generative model has to be validated against physical experiments before results obtained in it
can be relied on.

These systems differ from the text-based devices the paper has in mind in ways that matter for
evaluation. Their outputs are trajectories and forces rather than sentences, and their errors are physical,
so that the counterpart of a confabulated answer is an instrument in the wrong place. Language also
becomes a control channel. An instruction spoken to the system is both a user interaction and a safetycritical input, and R.1's concern with consistency across paraphrases applies directly.

We recommend:
• Treat any device with a generative model in its development or deployment pipeline as within the
scope of this framework, and ask sponsors to disclose that use and to show how it was validated.
• Extend the competency approach with a set of physical and perceptual competencies for systems
that guide or move instruments.
• Recognize high-fidelity simulation, phantom and cadaveric evaluation as the non-clinical tier for
such systems, with evidence that performance there predicts performance in patients. Where the
simulator is a generative world model, ask for evidence that it has been validated against physical
experiments.
• Treat spoken or typed instructions to a system that acts as safety-critical inputs, tested for
misrecognition and ambiguity.

Section VI. Postmarket monitoring

Question 18
Under what conditions might greater premarket uncertainty be accepted through greater reliance on
postmarket monitoring, and for which devices would this not be appropriate?

We support the principle, on one condition. Accepting more uncertainty before authorization is
reasonable only where something exists after authorization that can detect what was left uncertain. For
most intraoperative AI nothing of the kind exists today. We therefore believe that the cornerstone of
postmarket oversight for procedural AI should be a registry in which adverse events and near misses
involving these devices are recorded, and that creating it should be treated as a precondition for the
trade-off FDA is considering.

The existing system is not adequate to the task. Medical device reporting relies on manufacturers and
user facilities to recognize and report events [98], and FDA itself describes the resulting database as a
passive surveillance system limited by under-reporting, by inaccuracies and by the absence of
information on how often a device is used [99]. An analysis of 14 years of reports on robotic surgery
found 10,624 reports, including 144 deaths and 1,391 injuries, and noted that the reports often lacked

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 18
the detail needed to establish what had happened [100]. An error in an AI overlay is even less likely to
be reported, because it leaves no trace unless the video and the output of the device were kept.

There is a good precedent for the alternative. The Transcatheter Valve Therapy Registry was created by
two professional societies in collaboration with FDA and the Centers for Medicare & Medicaid Services
[101], and FDA has since relied on its data in approving a labeling change [102]. Coordinated registry
networks, including one for robotic-assisted surgical devices, have been organized along similar lines
[103,104].

Building such a registry will not be easy, and we see four main difficulties.

Recording in the operating room. In a procedure the input to the device is a video stream and the
consequence of an incorrect output is an action on tissue, and neither is routinely recorded. Whether an
overlay was right usually cannot be established afterward without the video. Recording is feasible but
not simple. In one implementation study 80% of operations were captured, and of those missed, 58%
were lost because staff declined consent and 10% because patients did [105]. Hospital information
systems were also not designed to store video on this scale [106].

Synchronization of data sources. The relevant data sit in separate systems with separate clocks: the
endoscopic video, the device's own log, the kinematic record of the robot, the anesthesia record and the
electronic health record. Perioperative data are distributed across information systems that seldom
exchange data [106], and without a common time base it is not possible to say what the device was
showing at the moment something went wrong. Research consortia have shown that robotic video and
kinematics can be collected in synchrony across many institutions [4], but routine clinical recording does
not do this.

Extraction from the electronic record. The outcomes that matter, such as a bile leak, a return to the
operating room or a readmission, are documented largely in free text. Registries such as the American
College of Surgeons National Surgical Quality Improvement Program depend on trained reviewers who
abstract charts by hand, and investigators working with that registry have described the abstraction as
laborious and a barrier to sustaining it [107]. Automated extraction is improving but is not yet a
substitute.

Medicolegal repercussions. Surgeons and hospitals will hesitate to record and report if the record can
be used against them. Concerns about privacy, data use and litigation have limited the adoption of
recording in the operating room [108], and a registry of AI-related events raises the same concerns.
Federal law offers a partial answer, in that information assembled for reporting to a patient safety
organization is privileged and confidential [109]. Whether that protection reaches device logs and video
is the kind of question that FDA, working with the Agency for Healthcare Research and Quality, is in a
position to clarify.

Rare, severe harms compound the problem. A monitoring program cannot wait for bile duct injuries to
accumulate, and it has to rely on leading indicators defined in advance. On the last part of the question,
the approach is least appropriate for functions that bear on an irreversible step where the sponsor
cannot show how an incorrect output would be detected in use.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 19
We recommend:
• Make a postmarket registry of adverse events and near misses the foundation of oversight for
procedural AI, built with professional societies on the model of existing device registries (Questions
19 and 21).
• Require devices to keep a time-stamped log of their outputs, software version and imaging
configuration in an exportable format, so that device data can be aligned with video and with the
clinical record.
• Define, before authorization, the measures that will be observed in the field, including measures of
user interaction, with thresholds, responsible parties and actions (including suspension of the
function) prespecified, and with clear escalation procedures when thresholds are not met [43].
• Provide a route to ground truth, such as adjudicated samples of procedural video.
• Make sure that monitoring can detect a change in how clinicians rely on the device as well as a
change in the model.
• Address the medicolegal barrier directly, for example by clarifying how reporting through patient
safety organizations applies to device and video data.

Questions 19 and 21
Which postmarket approaches are appropriate, and how should cadence and triggers be set? What roles
might clinicians, institutions and professional societies play without diffusing manufacturer accountability?

The three approaches FDA describes examine the inputs and outputs of the device. Procedural AI needs
a fourth, which is to monitor what clinicians do with them. Professional societies can play a part in
creating the postmarket adverse-event datasets described under Question 18, and can supply reference
standards and adjudication, while the manufacturer remains accountable for the device.

Re-benchmarking, sample-based review and degradation monitoring would not detect a change in
which the model is stable but clinicians accept its outputs more readily and check them less. They may
also miss the kind of drift that is characteristic of surgery, where what shifts is often not the patient
population but the equipment, the technique and the operators: a new scope or video processor, a
different energy device, a new group of trainees. A minimum dataset should therefore include whether
the function was active; the software version and imaging configuration; when outputs were displayed;
what the operator did next, where the device can observe it; the experience of the operator; and a
sampled, adjudicated video record. Automated video review is a plausible “supervisory agent” (Question
20), but it is itself an AI system that requires validation, and it should not leave manufacturers assessing
their own devices.

On roles, we would divide the work as follows. Manufacturers are responsible for the performance of
the device and for making interaction data available. Institutions are responsible for governing
deployment and for local monitoring. Professional societies can supply reference standards, as they
have for the critical view of safety [45], provide adjudication, and host or govern the registries that link
device use to adverse events and outcomes. The transcatheter valve registry, which is run by the Society
of Thoracic Surgeons and the American College of Cardiology, shows that this can be done in partnership

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 20
with FDA [101]. SAGES has suggested that surgical societies form a collaborative community for work of
this kind [43]. The division follows the principle that incentives should be aligned so that those “best
positioned to know the AI system risks and best positioned to avert or mitigate harm” do so [110].

We recommend:
• Add monitoring of user interaction to the postmarket approaches, and expect devices to record and
export a minimum interaction dataset.
• Set cadence by volume of use rather than by the calendar alone, and include new sites, new user
populations and changes to imaging equipment or platform among the triggers for reassessment
(Questions 22 to 24).
• Involve professional societies such as SAGES in building and governing postmarket adverse-event
datasets for procedural AI, together with FDA, existing registries and academic groups that record
and analyze operative video and robot data. We would be glad to help define a minimum dataset
for surgical and endoscopic AI.

Questions 22 to 24
How should re-benchmarking scale with the impact of a modification, how can change control work when
changes cannot be fully prespecified, and how can manufacturers manage changes initiated by third parties?

The paper's discussion of modifications is about the model. For procedural AI the changes that matter
are often elsewhere, in what the clinician sees and in the imaging chain that feeds the model. A change
to how and when a device communicates with the clinician is relevant to safety, a change anywhere in
the imaging chain is the procedural counterpart of an update to a foundation model, and re-evaluation
after either has to include the user as well as the model.

The paper does list the user interface among the features of a device that may change after
deployment. For a device used during a procedure the interface is part of the device's effect. Moving an
overlay from a separate monitor to the operative view, changing its color or persistence, adding a
confidence display or altering an alert threshold can change how often clinicians act on an incorrect
output with no change to the model at all. FDA's 2026 human factors guidance allows a sponsor, in
some circumstances, to justify not submitting validation test data, taking into account the history of the
user interface [111]. We suggest that experience with an existing interface should not by itself support
that conclusion once the interface presents a new inference.

On third-party changes, the input to a procedural AI passes through components its manufacturer may
not control: the endoscope, light source, camera, video processor and its image enhancement, and the
display pipeline of a robotic platform. A firmware update to any of these alters the model's input
without any change to the device. Navigation depends in the same way on imaging protocols and
tracking hardware. Where changes cannot be prespecified (Question 23), a change control plan [112]
can define an envelope, meaning bounds on performance and on measures of user interaction, within
which changes are verified against a fixed, sequestered test set.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 21
We recommend:
• Scale re-evaluation by what changed: the model, its input, or what the user sees. Model changes
within a prespecified envelope, verified on a fixed sequestered set, suit a change control plan.
• Changes to what the user sees, or changes that bring an output closer to an irreversible step, should
repeat testing with users.
• Labeling should state the imaging configurations with which a device has been validated, and
devices should detect and flag an unvalidated configuration where feasible.

Section VII. Other topics

Question 26
What additional considerations apply to agentic devices, and how should the risk of autonomous multi-step
action and reduced opportunity for human review be reflected in acceptance criteria and oversight?

In surgery an agentic system is one that acts physically. The agentic systems the paper describes support
care coordination, documentation, patient outreach and clinical workflow, but it notes that such a
system may be a device where its “action sequences result in control of another medical device.” That is
the surgical case: a learned model that moves a camera, retracts tissue, places a clip or divides a
structure.

Our view is that autonomous multi-step action in soft-tissue surgery is still far off, and that the ways in
which it might be achieved are too varied and too unpredictable for a general methodology to be
written now. The most advanced demonstrations of which we are aware have been in animal and ex
vivo models [30,32]. The prospects are nearer in fields where the anatomy is more constrained or the
task can be planned from imaging, notably endovascular intervention, orthopedics and urology. FDAcleared surgical robots already span levels of autonomy. A systematic review of 49 systems cleared from
2015 to 2023 classified 86% as robot assistance, 8% as task autonomy and 6% as conditional autonomy,
and the three most autonomous systems were for bone milling in orthopedics, prostate biopsy in
urology and hair follicle extraction [113]. Robotic knee arthroplasty systems have for years confined the
saw within haptic boundaries set by the preoperative plan [114]. In 2026 ARPA-H committed up to
$175.3 million over five years to a program on autonomous interventions, which includes endovascular
robots for stroke and aims to demonstrate fully autonomous interventions in models, animals or
cadavers by the end of that period [37].

These approaches differ a great deal from one another in what is automated, in how the plan is formed
and in how the clinician supervises. For that reason we believe the regulatory framework for
autonomous action should, for the present, be developed case by case through early engagement
between FDA and each sponsor, and should be generalized only once there is experience to generalize
from.

Some elements will, we think, be needed in every case. Appendix A lists “compliance with humanoversight checkpoints before irreversible or high-consequence actions” among the things to be
evaluated. Surgery can say what those actions are (applying a clip, dividing a structure, firing a stapler,

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 22
activating energy near a critical structure, crossing a planned boundary), and established practice
already marks some of them, as the critical view of safety and the pause before clipping do [45,46].
Oversight during a procedure also differs from reviewing a draft. The supervising surgeon must perceive
the problem, decide and take control within seconds, and a supervisor who rarely needs to intervene
becomes less practiced at it [82].

We recommend:
• Develop requirements for autonomous surgical functions case by case for now, and say so, with an
expectation of early pre-submission engagement.
• In each case, ask the sponsor to specify the irreversible actions the function can take and the
checkpoint before each.
• Evaluate takeover, meaning the time and success of human intervention under realistic conditions,
including by less experienced intended users.
• Require that transfers of control from system to surgeon be logged with their cause and
summarized in postmarket reporting.
• Scale the evidence to an established taxonomy of autonomy [115], drawing on existing standards
[116,117].

Closing
We thank the Agency for seeking the views of clinicians and engineers at this stage, and we would be
glad to provide further detail on any of the points above.

Respectfully submitted,

Filippo Filicori, MD, FACS, FASMBS
Director, Intraoperative Performance Analytics Laboratory, Department of Surgery, Lenox Hill Hospital,
Northwell Health, New York, NY
System Chief, Surgical Innovation, Northwell Health, New York, NY
Professor of Surgery, Zucker School of Medicine at Hofstra/Northwell
Zoran Kostić, PhD
Professor of Engineering, Department of Electrical Engineering, Columbia University, New York, NY

The views expressed are those of the signatories and do not necessarily represent the official positions of Northwell
Health or Columbia University.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 23
Table. FDA's risk considerations, translated for the operating room
What the discussion paper considers The surgical counterpart

Activity axis: informational non-directive → Post-hoc phase analytics → anatomy or navigation overlay
action-directing → action-taking, HCP-supervised → “do not clip” → a navigation-defined boundary that
→ fully autonomous holds the instrument → autonomous clip-and-divide

Consequences of relying on an incorrect output Often irreversible within seconds: a divided common bile
duct cannot be undone

Directiveness depends on substance and context, “For information only” does not change what a binary
not solely on wording; a disclaimer may not safety call does to a surgeon in mid-dissection
reduce it

Measurement and signal-processing outputs the Navigation and augmented-reality registration of
user cannot independently evaluate preoperative imaging to moving anatomy; perfusion and
tissue characterization; segmentation of structures the
surgeon cannot yet see

Patient-facing versus HCP-facing: does the user Who sees the output: the operating surgeon alone, or the
have the knowledge to judge the output? whole team, including trainees and staff who are not
placed to judge it

Generalist versus specialist HCP Trainee versus attending; occasional versus high-volume
operator; new to a platform versus experienced

Multi-turn conversations migrating toward action- The procedure itself: informational during exposure,
directing directive at the moment of clipping

Care escalation, under and over Stop, obtain another view, bail out, convert, call for help;
and the cost of alerts that fire too often

References
1. US Food and Drug Administration, Center for Devices and Radiological Health. Considerations for the Regulation of
Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. August 2026. Docket FDA2026-N-7874.
2. Morais MC, Godbole AA, Iqbal E, et al. Standardization of surgical gesture taxonomy: a SAGES Delphi consensus
study. Surg Endosc. 2026;40(6):4514–4528.
3. Alapatt D, Eckhoff J, Lyu Z, et al. The SAGES Critical View of Safety Challenge: a global benchmark for AI-assisted
surgical quality assessment. arXiv:2509.17100 [preprint]. 2025.
4. Open-H-Embodiment Consortium, Nelson N, Chen JT, et al. Open-H-Embodiment: a large-scale dataset for
enabling foundation models in medical robotics. arXiv:2604.21017 [preprint]. 2026.
5. Zang C, Turkcan MK, Narasimhan S, et al. Surgical phase recognition in inguinal hernia repair—AI-based
confirmatory baseline and exploration of competitive models. Bioengineering (Basel). 2023;10(6):654.
6. Choksi S, Szot S, Zang C, et al. Bringing artificial intelligence to the operating room: edge computing for real-time
surgical phase recognition. Surg Endosc. 2023;37(11):8778–8784.
7. Choksi S, Narasimhan S, Ballo M, et al. Automatic assessment of robotic suturing utilizing computer vision in a drylab simulation. Artif Intell Surg. 2025;5:160–169.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 24
8. Narasimhan S, Turkcan MK, Ballo M, et al. Monocular 3D tooltip tracking in robotic surgery—building a multi-stage
pipeline. Electronics. 2025;14(10):2075.
9. Turkcan MK, Ballo M, Filicori F, Kostic Z. Towards suturing world models: learning predictive models for robotic
surgical tasks. arXiv:2503.12531 [preprint]. 2025.
10. Nath V, Tejero JG, Kumar AS, et al. G-SHARP: Gaussian surgical hardware accelerated real-time pipeline.
arXiv:2512.02482 [preprint]. 2025.
11. Bitner DP, Choksi S, Carsky K, et al. Kinematic metrics and surgeon experience in robotic cholecystectomies: a pilot
study on breaking down technical performance. Surg Endosc. 2024;38(2):913–921.
12. Choksi S, Ballo M, Profant C, et al. Standardizing surgical training with objective performance indicators: a
prospective cohort study. Surg Endosc. 2025;39(4):2316–2323.
13. Choksi S, Hirachan B, Ballo M, et al. Characterizing advanced stapling technique using objective performance
indicators in robotic-assisted sleeve gastrectomy: a retrospective cohort study. Surg Endosc. 2025;39(8):4848–
4856.
14. Tindal EW, Ballo M, Nussbaum JE, et al. Quantifying surgical complexity: how smoothness metrics vary in robotic
cholecystectomy. Surg Endosc. 2026;40(2):1686–1693.
15. Addison P, Bitner D, Chung P, et al. Blinded intraoperative skill evaluations avoid gender-based bias. Surg Endosc.
2022;36(11):8458–8462.
16. Addison P, Bitner DP, Addy J, et al. Does surgeon experience correlate with crowd-sourced skill assessment in
robotic bariatric surgery? Am Surg. 2023;89(12):5253–5262.
17. Addison P, Bitner D, Carsky K, et al. Outcome prediction in bariatric surgery through video-based assessment. Surg
Endosc. 2023;37(4):3113–3118.
18. Bitner DP, Kutana S, Carsky K, et al. The surgical learning curve: does robotic technical skill explain differences in
operative performance? J Laparoendosc Adv Surg Tech A. 2023;33(5):471–479.
19. Ballo M, Profant C, Tindal EW, et al. Video-based assessment and prediction of surgical outcomes in sleeve
gastrectomy. J Am Coll Surg. 2025. doi:10.1097/XCS.0000000000001595.
20. Yuan K, Srivastav V, Yu T, et al. Learning multi-modal representations by watching hundreds of surgical video
lectures. Med Image Anal. 2025;105:103644.
21. Schmidgall S, Kim JW, Kuntz A, Ghazi AE, Krieger A. General-purpose foundation models for increased autonomy in
robot-assisted surgery. Nat Mach Intell. 2024;6(11):1275–1283.
22. Madani A, Namazi B, Altieri MS, et al. Artificial intelligence for intraoperative guidance: using semantic
segmentation to identify surgical anatomy during laparoscopic cholecystectomy. Ann Surg. 2022;276(2):363–369.
23. Laplante S, Namazi B, Kiani P, et al. Validation of an artificial intelligence platform for the guidance of safe
laparoscopic cholecystectomy. Surg Endosc. 2023;37(3):2260–2268.
24. Mascagni P, Vardazaryan A, Alapatt D, et al. Artificial intelligence for surgical safety: automatic assessment of the
critical view of safety in laparoscopic cholecystectomy using deep learning. Ann Surg. 2022;275(5):955–961.
25. Kitaguchi D, Kosugi N, Ishikawa Y, et al. A multicentre randomized controlled trial exploring the clinical usefulness
of the intraoperative use of an artificial intelligence–based anatomical navigation system. BJS.
2025;112(6):znaf121.
26. Pharmaceuticals and Medical Devices Agency, Japan. Approved devices under the generic name “surgical image
recognition support program,” including approval no. 30600BZX00061000 (Ministry of Health, Labour and Welfare,
April 12, 2024).
27. Bernhardt S, Nicolau SA, Soler L, Doignon C. The status of augmented reality in laparoscopic surgery as of 2016.
Med Image Anal. 2017;37:66–90.
28. Schneider C, Allam M, Stoyanov D, et al. Performance of image guided navigation in laparoscopic liver surgery: a
systematic review. Surg Oncol. 2021;38:101637.
29. Shademan A, Decker RS, Opfermann JD, et al. Supervised autonomous robotic soft tissue surgery. Sci Transl Med.
2016;8(337):337ra64.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 25
30. Saeidi H, Opfermann JD, Kam M, et al. Autonomous robotic laparoscopic surgery for intestinal anastomosis. Sci
Robot. 2022;7(62):eabj2908.
31. Kim JW, Zhao TZ, Schmidgall S, et al. Surgical Robot Transformer (SRT): imitation learning for surgical tasks.
Proceedings of the 8th Conference on Robot Learning (CoRL 2024). Proc Mach Learn Res. 2025;270:130–144.
32. Kim JW, Chen JT, Hansen P, et al. SRT-H: a hierarchical framework for autonomous surgery via languageconditioned imitation learning. Sci Robot. 2025;10(104):eadt5254.
33. Li Y, Richter F, Lu J, et al. SuPer: a surgical perception framework for endoscopic tissue manipulation with surgical
robotics. IEEE Robot Autom Lett. 2020;5(2):2294–2301.
34. Richter F, Shen S, Liu F, et al. Autonomous robotic suction to clear the surgical field for hemostasis using imagebased blood flow detection. IEEE Robot Autom Lett. 2021;6(2):1383–1390.
35. Liang X, Wang CP, Shinde NU, et al. MEDiC: autonomous surgical robotic assistance to maximizing exposure for
dissection and cautery. arXiv:2409.14287 [preprint]. 2024.
36. Yip M, Salcudean S, Goldberg K, et al. Artificial intelligence meets medical robotics. Science. 2023;381(6654):141–
146.
37. Advanced Research Projects Agency for Health (ARPA-H). ARPA-H funded teams to transform life-saving stroke
interventions with autonomous robotic surgeries. August 6, 2026. Available at arpa-h.gov.
38. Goldberg K, Guthart G. Augmented dexterity: how robots can enhance human surgical skills. Sci Robot.
2024;9(95):eadr5247.
39. Attanasio A, Scaglioni B, Leonetti M, et al. Autonomous tissue retraction in robotic assisted minimally invasive
surgery: a feasibility study. IEEE Robot Autom Lett. 2020;5(4):6528–6535.
40. Meireles OR, Rosman G, Altieri MS, et al. SAGES consensus recommendations on an annotation framework for
surgical video. Surg Endosc. 2021;35(9):4918–4929.
41. Eckhoff JA, Rosman G, Altieri MS, et al. SAGES consensus recommendations on surgical video data use, structure,
and exploration (for research in artificial intelligence, clinical quality improvement, and surgical education). Surg
Endosc. 2023;37(11):8690–8707.
42. Ali JT, Yang G, Green CA, et al. Defining digital surgery: a SAGES white paper. Surg Endosc. 2024;38(2):475–487.
43. Guzman NM, Marwaha JS, Keller DS, et al. Regulating digital surgery to balance safety and innovation: a SAGES
white paper. Surg Endosc. 2026;40(7):5480–5490.
44. Hashimoto DA, Marwaha JS, Lee SA, Schwaitzberg S, Duffourc MN. Risk and liability in the deployment of AI
systems for surgery: a SAGES white paper. Surg Endosc. 2026;40(7):5468–5479.
45. Brunt LM, Deziel DJ, Telem DA, et al. Safe cholecystectomy multi-society practice guideline and state of the art
consensus conference on prevention of bile duct injury during cholecystectomy. Ann Surg. 2020;272(1):3–23.
46. Society of American Gastrointestinal and Endoscopic Surgeons. The SAGES Safe Cholecystectomy Program:
strategies for minimizing bile duct injuries. Available at: https://www.sages.org/safe-cholecystectomy-program/
47. Strasberg SM, Hertl M, Soper NJ. An analysis of the problem of biliary injury during laparoscopic cholecystectomy. J
Am Coll Surg. 1995;180(1):101–125.
48. Ryu S, Imaizumi Y, Goto K, et al. Artificial intelligence-enhanced navigation for nerve recognition and surgical
education in laparoscopic colorectal surgery. Surg Endosc. 2025;39(2):1388–1396.
49. Buçinca Z, Malaya MB, Gajos KZ. To trust or to think: cognitive forcing functions can reduce overreliance on AI in
AI-assisted decision-making. Proc ACM Hum Comput Interact. 2021;5(CSCW1):188.
50. Cabitza F, Natali C, Famiglini L, et al. Never tell me the odds: investigating pro-hoc explanations in medical decision
making. Artif Intell Med. 2024;150:102819.
51. Romeo G, Conti D. Exploring automation bias in human–AI collaboration: a review and implications for explainable
AI. AI Soc. 2026;41(1):259–278.
52. Dratsch T, Chen X, Rezazade Mehrizi M, et al. Automation bias in mammography: the impact of artificial
intelligence BI-RADS suggestions on reader performance. Radiology. 2023;307(4):e222176.
53. Tschandl P, Rinner C, Apalla Z, et al. Human–computer collaboration for skin cancer recognition. Nat Med.
2020;26(8):1229–1234.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 26
54. Parasuraman R, Manzey DH. Complacency and bias in human use of automation: an attentional integration. Hum
Factors. 2010;52(3):381–410.
55. Endo Y, Tokuyasu T, Mori Y, et al. Impact of AI system on recognition for anatomical landmarks related to reducing
bile duct injury during laparoscopic cholecystectomy. Surg Endosc. 2023;37(7):5752–5759.
56. Hasegawa T, Etoh T, Nakano T, et al. Impact of artificial intelligence intervention on awareness of anatomical
landmarks in laparoscopic gastrectomy: an exploratory study. Surg Endosc. 2026;40(9):7523–7532.
57. Morais MC, et al. Surg Endosc. In press.
58. 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.
59. van der Sijs H, Aarts J, Vulto A, Berg M. Overriding of drug safety alerts in computerized physician order entry. J Am
Med Inform Assoc. 2006;13(2):138–147.
60. Ancker JS, Edwards A, Nosal S, et al. Effects of workload, work complexity, and repeated alerts on alert fatigue in a
clinical decision support system. BMC Med Inform Decis Mak. 2017;17(1):36.
61. Sendelbach S, Funk M. Alarm fatigue: a patient safety concern. AACN Adv Crit Care. 2013;24(4):378–386.
62. Dixon BJ, Daly MJ, Chan H, et al. Surgeons blinded by enhanced navigation: the effect of augmented reality on
attention. Surg Endosc. 2013;27(2):454–461.
63. Hughes-Hallett A, Mayer EK, Marcus HJ, et al. Inattention blindness in surgery. Surg Endosc. 2015;29(11):3184–
3189.
64. Buchanan J, Connor S, Pearson J, Carey-Smith B, Eglinton T. Surgical computer vision for intraoperative decisionsupport: a scoping review on performance metrics and readiness for real-time deployment. Artif Intell Surg.
2026;6:150–170.
65. ASTM International. ASTM F2554-22, Standard Practice for Measurement of Positional Accuracy of ComputerAssisted Surgical Systems.
66. International Electrotechnical Commission. IEC 62366-1:2015+AMD1:2020, Medical devices, Part 1: Application of
usability engineering to medical devices.
67. Twinanda AP, Shehata S, Mutter D, et al. EndoNet: a deep architecture for recognition tasks on laparoscopic
videos. IEEE Trans Med Imaging. 2017;36(1):86–97.
68. Kitaguchi D, Fujino T, Takeshita N, et al. Limited generalizability of single deep neural network for surgical
instrument segmentation in different surgical environments. Sci Rep. 2022;12:12575.
69. Ward TM, Fer DM, Ban Y, et al. Challenges in surgical video annotation. Comput Assist Surg (Abingdon).
2021;26(1):58–68.
70. Mascagni P, Alapatt D, Lapergola A, et al. Early-stage clinical evaluation of real-time artificial intelligence assistance
for laparoscopic cholecystectomy. BJS. 2024;111(1):znad353.
71. McCulloch P, Altman DG, Campbell WB, et al. No surgical innovation without evaluation: the IDEAL
recommendations. Lancet. 2009;374(9695):1105–1112.
72. Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision
support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28(5):924–933.
73. Investigational Device Exemptions, 21 CFR Part 812, §§ 812.2 and 812.3(m).
74. US Food and Drug Administration. Information Sheet Guidance for IRBs, Clinical Investigators, and Sponsors:
Significant Risk and Nonsignificant Risk Medical Device Studies. January 2006.
75. US Food and Drug Administration. Requests for Feedback and Meetings for Medical Device Submissions: The QSubmission Program. Guidance for Industry and Food and Drug Administration Staff. May 2025.
76. Acceptance of data from clinical investigations conducted outside the United States, 21 CFR 812.28 (final rule, 83
FR 7366, February 21, 2018).
77. US Food and Drug Administration. Acceptance of Clinical Data to Support Medical Device Applications and
Submissions: Frequently Asked Questions. Guidance for Industry and Food and Drug Administration Staff. February
2018.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 27
78. Reverberi C, Rigon T, Solari A, et al. Experimental evidence of effective human–AI collaboration in medical
decision-making. Sci Rep. 2022;12:14952.
79. Sultan S, Shung DL, Kolb JM, et al. AGA living clinical practice guideline on computer-aided detection–assisted
colonoscopy. Gastroenterology. 2025;168(4):691–700.
80. Mangieri CW, Hendren BP, Strode MA, Bandera BC, Faler BJ. Bile duct injuries (BDI) in the advanced laparoscopic
cholecystectomy era. Surg Endosc. 2019;33(3):724–730.
81. Abdulnour RE, Gin B, Boscardin CK. Educational strategies for clinical supervision of artificial intelligence use. N
Engl J Med. 2025;393(8):786–797.
82. Bainbridge L. Ironies of automation. Automatica. 1983;19(6):775–779.
83. Casner SM, Geven RW, Recker MP, Schooler JW. The retention of manual flying skills in the automated cockpit.
Hum Factors. 2014;56(8):1506–1516.
84. Bingmer K, Ofshteyn A, Stein SL, et al. Decline of open surgical experience for general surgery residents. Surg
Endosc. 2020;34(2):967–972.
85. Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling risk after exposure to artificial intelligence in
colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896–903.
86. Ahmad OF. Endoscopist deskilling: an unintended consequence of AI-assisted colonoscopy? Lancet Gastroenterol
Hepatol. 2025;10(10):872–873.
87. US Food and Drug Administration. 510(k) Third Party Review Program and Third Party Emergency Use
Authorization (EUA) Review: Guidance for Industry, Food and Drug Administration Staff, and Third Party Review
Organizations. November 2024.
88. US Food and Drug Administration. Qualification of Medical Device Development Tools: Guidance for Industry, Tool
Developers, and Food and Drug Administration Staff. July 2023.
89. US Food and Drug Administration. Collaborative Communities: Addressing Health Care Challenges Together.
Content current as of April 22, 2026. Available at fda.gov.
90. Li J, Skinner G, Yang G, et al. LLaVA-Surg: towards multimodal surgical assistant via structured surgical video
learning. arXiv:2408.07981 [preprint]. 2024.
91. Venkatesh DK, Rivoir D, Pfeiffer M, Kolbinger F, Speidel S. Data augmentation for surgical scene segmentation with
anatomy-aware diffusion models. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer
Vision (WACV). 2025:2280–2290.
92. Alabi O, Toe KKZ, Zhou Z, et al. CholecInstanceSeg: a tool instance segmentation dataset for laparoscopic surgery.
Sci Data. 2025;12:825.
93. NVIDIA. Cosmos world foundation model platform for physical AI. arXiv:2501.03575 [preprint]. 2025.
94. NVIDIA. NVIDIA launches Cosmos 3, the open frontier foundation model for physical AI [press release]. June 1,
2026.
95. NVIDIA. Cosmos 3: omnimodal world models for physical AI. arXiv:2606.02800 [preprint]. 2026.
96. He Y, Guo P, Xu M, et al. Cosmos-H-Surgical: learning surgical robot policies from videos via world modeling.
arXiv:2512.23162v4 [preprint]. 2026 (earlier versions titled “SurgWorld”).
97. Zbinden L, Nelson N, Chen JT, et al. Cosmos-Surg-dVRK: world foundation model-based automated online
evaluation of surgical robot policy learning. arXiv:2510.16240 [preprint]. 2025.
98. Medical Device Reporting, 21 CFR Part 803.
99. US Food and Drug Administration. About Manufacturer and User Facility Device Experience (MAUDE) Database.
Available at fda.gov.
100. Alemzadeh H, Raman J, Leveson N, Kalbarczyk Z, Iyer RK. Adverse events in robotic surgery: a retrospective study
of 14 years of FDA data. PLoS One. 2016;11(4):e0151470.
101. Carroll JD, Edwards FH, Marinac-Dabic D, et al. The STS-ACC Transcatheter Valve Therapy National Registry: a new
partnership and infrastructure for the introduction and surveillance of medical devices and therapies. J Am Coll
Cardiol. 2013;62(11):1026–1034.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 28
102. US Food and Drug Administration. FDA Facts: Postmarket Patient Registry Ensures Access to Safe and Effective
Devices. Content current as of December 21, 2017. Available at fda.gov.
103. Krucoff MW, Sedrakyan A, Normand SLT. Bridging unmet medical device ecosystem needs with strategically
coordinated registries networks. JAMA. 2015;314(16):1691–1692.
104. Medical Device Epidemiology Network (MDEpiNet). Coordinated Registry Networks. Available at mdepinet.net.
105. Møller KE, Sørensen JL, Topperzer MK, et al. Implementation of an innovative technology called the OR Black Box:
a feasibility study. Surg Innov. 2023;30(1):64–72.
106. Maier-Hein L, Eisenmann M, Sarikaya D, et al. Surgical data science: from concepts toward clinical translation. Med
Image Anal. 2022;76:102306.
107. Skube SJ, Hu Z, Simon GJ, et al. Accelerating surgical site infection abstraction with a semi-automated machinelearning approach. Ann Surg. 2022;276(1):180–185.
108. van Dalen ASHM, Legemaate J, Schlack WS, Legemate DA, Schijven MP. Legal perspectives on black box recording
devices in the operating environment. Br J Surg. 2019;106(11):1433–1441.
109. Patient Safety and Quality Improvement Act of 2005, Pub. L. 109-41, 42 U.S.C. 299b-21 to 299b-26; implementing
regulation at 42 CFR Part 3.
110. American Medical Association. Augmented Intelligence Development, Deployment, and Use in Health Care.
November 2024.
111. US Food and Drug Administration. Content of Human Factors Information in Medical Device Marketing
Submissions: Guidance for Industry and Food and Drug Administration Staff. May 2026.
112. US 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 2025 (originally issued December 2024).
113. Lee A, Baker TS, Bederson JB, Rapoport BI. Levels of autonomy in FDA-cleared surgical robots: a systematic review.
npj Digit Med. 2024;7:103.
114. Kayani B, Konan S, Ayuob A, et al. Robotic technology in total knee arthroplasty: a systematic review. EFORT Open
Rev. 2019;4(10):611–617.
115. Yang GZ, Cambias J, Cleary K, et al. Medical robotics: regulatory, ethical, and legal considerations for increasing
levels of autonomy. Sci Robot. 2017;2(4):eaam8638.
116. International Electrotechnical Commission. IEC 80601-2-77:2019+AMD1:2023, Medical electrical equipment, Part
2-77: Particular requirements for the basic safety and essential performance of robotically assisted surgical
equipment.
117. International Electrotechnical Commission. IEC/TR 60601-4-1:2017, Medical electrical equipment, Part 4-1:
Guidance and interpretation. Medical electrical equipment and medical electrical systems employing a degree of
autonomy.

IPAL, Northwell Health, and Columbia University comments, Docket FDA-2026-N-7874, page 29