ARISE Research Network (Adam Rodman)

See attached file(s)

Comments from Adam Rodman, MD, MPH, Sophia Li, and David Wu, MD, PhD, on behalf of the
ARISE Research Network

6. For GenAI-enabled care escalation functions, CDRH is considering how the evaluation may
account for both under-escalation and over-escalation. How could manufacturers characterize
and weigh these two directions of error, given that they may not be commensurable and that
acceptable trade-offs may vary by clinical context?

We agree that under-escalation and over-escalation should be measured separately and weighted
according to the severity of their potential consequences. A missed escalation may delay necessary
care and result in substantial patient harm. Unnecessary escalation may expose patients to
avoidable testing or treatment, consume limited resources, and contribute to inflationary health
costs.

Because some degree of under- and over-triage is unavoidable, errors should be evaluated
according to the potential harm they could cause (Wu et al. 2025). Severity-weighted evaluation
would allow regulators, clinicians, and other stakeholders to determine an acceptable balance of
under- and over-escalation for a particular clinical use case.

Degree of harm Description (adapted from WHO Harm Severity Ratings) (The conceptual
framework for the inte...)

None GenAI recommendations lead to a patient outcome that is not symptomatic or
no symptoms detected and no treatment is required.

Mild GenAI recommendations lead to a patient outcome that is symptomatic,
symptoms are mild, loss of function or harm is minimal or intermediate but
short term, and no or minimal intervention (e.g., extra observation,
investigation, review or minor treatment) is required.

Moderate GenAI recommendations lead to a patient outcome that is symptomatic,
requiring intervention (e.g., additional operative procedure; additional
therapeutic treatment), an increased length of stay, or causing permanent or
long term harm or loss of function.

Severe GenAI recommendations lead to a patient outcome that is symptomatic,
requiring life-saving intervention or major surgical/medical intervention,
shortening life expectancy or causing major permanent or long term harm or
loss of function.

Death On balance of probabilities, death was caused or brought forward in the short
term by following GenAI recommendations.
7. Is the competency-based approach described above, i.e., device benchmarking followed by
clinical confirmation, a useful and appropriate framework for evaluating GenAI-enabled devices?

We agree that device benchmarking, followed by clinical confirmation, is the most appropriate path
forward. Benchmarking can provide a standardized, context-specific snapshot of model capabilities
across multiple functions at a given point in time, and it should be followed by clinical confirmation
and longitudinal assessment to evaluate real-world performance and model evolution. GenAI
presents several challenges that make a regulatory regime more like licensing a healthcare
professional rather than a device or drug. Of the existing licensure pathways that have been
suggested, there is no clear mechanism by which GenAI could fit. In fact, the overwhelming
evidence suggests that the traditional evaluation types used in licensure pathways collapse when
used to evaluate GenAI for several compelling reasons:

First, GenAI outputs are general-purpose and open-ended. Even when prompting or clinical
harnesses constrain them, GenAI tools perform multiple tasks such as history-taking, diagnostic
reasoning, recommendation generation, and patient communication. This makes traditional
classification measurements (e.g. sensitivity, specificity, PPV, NPV) particularly unhelpful.

Second, output depends heavily on context. Because patient-facing systems will reason over large
amounts of electronic health record-derived information, small differences or perturbations can
have large impacts (just as they would for a human clinician). This contextual complexity expands
further when interacting with real patients. Patients can interact with GenAI systems in many
different ways, and share data differently than they do with human healthcare providers (Brodeur
et al. 2026)

More advanced systems that involve audio and visual inputs only increase this complexity (Shah et
al. 2026). Modern vision-language models (VLMs) can pick up on body language and nonverbal cues
and perform physical exams. However, these increasing abilities also lead to novel failure modes
such as “contextual completion:” the tendency of these systems to infer visual findings expected
from the initial presentation rather than what is actually present. Reliance on context is not a
negative but a strength of these systems, similar to the well-defined concept of context specificity
in human clinicians (Durning et al. 2012; Konopasky et al. 2020). However, this again supports the
need for human-like oversight mechanisms.

Third, as model capabilities continue to improve, a regulatory regime should avoid model lock-in.
Model capabilities grow on an exponential curve, offering huge opportunities to improve human
health. In particular, advances from just the past several months now enable audiovisual consults,
real-time interpretation in multiple languages, agentic reasoning over longitudinal health records
(Singh et al. 2026; Nagda et al. 2026; Jiang et al. 2025). Persistent AI agents that can coach and
assist with health outside of traditional healthcare venues (Clinical AI ). Any proposed regulatory
mechanism must allow the flexibility to upgrade existing workflows as models improve (or become
more efficient and can benefit more people). Traditional regulatory mechanisms that mandate a
locked model would not allow these benefits. However, a pre-clearance benchmarking mechanism
enables rapid testing and ensures model improvements quickly reach patients.

8. How might the two-axis risk framework described in Section IV be considered within the
competency-based approach described above to help determine the level of evidence needed for
a premarket submission?

The two-axis framework, organized by activity and clinical consequence of incorrect outputs, fits
well within a competency-based approach to premarket submission. However, we believe that it is
currently overly narrow because it focuses on incorrect outputs rather than clinical harm more
broadly. Harm, safety, and risk are broader constructs than the presence of inappropriate outputs
(errors of commission). These concepts also involve the absence of beneficial outputs (errors of
omission), which benchmarks often score separately under another performance metric such as
accuracy. The frequent dichotomization of safety as distinct from overall performance does not
reflect how these systems affect patients.

For many conventional therapeutics, efficacy and adverse effects can be evaluated as distinct
outcomes. However, when the endpoint reflects the overall clinical consequence (such as patient
survival), benefit and harm become inseparable. The same principle applies to medical AI: the
clinically relevant endpoint is the consequence of the system’s behavior, regardless of whether the
underlying failure is labeled as a safety error, an omission, or an inaccuracy.

Evaluating or optimizing measures of safety, harm, and performance in isolation can therefore be
misleading. Approaches that try to reduce risk by penalizing harmful outputs alone may cause
systems to avoid beneficial outputs. Aggregate accuracy is not a reliable proxy for safety. For
example, in a radiology benchmark with clinician-labeled high-risk answers, strategies could raise
accuracy without reducing high-risk errors or dangerous overconfidence (Wind et al. 2026).

We have argued elsewhere that safety is inseparable from capability in medical AI (Wu et al., under
review). Applied to the two-axis framework, evidence requirements should scale with the severity
of potential harm, including harm from omission. Sponsors should score omissions and
commissions within a common harm-severity framework and map errors to established harmseverity taxonomies such as those of AHRQ and WHO (The conceptual framework for the inte...;
Measure Dx: A Resource To Identify, A...). They should report both severity-weighted aggregate
performance and high-severity failure rates, with acceptance criteria set on the latter for functions
toward the upper right of the framework.
These approaches have increasingly been adopted across the industry, and we hope regulatory
regimes will enforce them as well.

9. Would a benchmarking structure such as the one described above be likely to provide
adequate evidence of clinical knowledge, analytic capabilities, safety behavior, communication,
and generalizability to support a reasonable assurance of safety and effectiveness? Are there
elements that are missing, redundant, or inappropriately categorized? Are there externally
developed standards that could be leveraged?

The benchmarking structure mentioned above is flexible and can encode the listed domains above
(generalizability is more complex; generalizability theory suggests that we probably can design
benchmarks that are broadly applicable, but current approaches would still require real world
testing and validation). Modern approaches to benchmarking are flexible: rubric-based
benchmarking can assess knowledge and safety behavior and psychometric benchmarking (eg, the
PDSQI-9) can determine communication skills. The science of evaluation continues to advance at a
rapid rate. Agentic abilities (in which AI systems interact both with charts and patients directly)
offer new challenges for evaluation. However, there is no reason to believe that this current system
will not be able to adapt to technological innovation.

External standards for human clinicians can also provide useful reference points. Licensing boards
such as the American Board of Internal Medicine or the American Board of Pediatrics have the
ethical and legal responsibility to define the clinical knowledge and performance-based standards
expected of physicians. While it is understood that these standards in humans are only a minimum
bar and that supervised practice is also necessary, the same fundamental principles can be applied
to AI systems. Many other human-based competencies are encapsulated in entrustable professional
activity (EPA) frameworks, used as part of a holistic licensing regimen for US physicians. Each
medical specialty has its own EPA framework that decomposes competencies by task. These
standards could also inform clinical benchmark design.

10. CDRH recognizes that publicly available benchmarking assets may be subject to data
contamination, saturation, and limited real-world representativeness. How should a sponsor
establish that performance on a given benchmark predicts safe and effective real-world behavior
for the device’s intended use? What evidence should support the construct validity of a
benchmark used to gate device evaluation, and what role should sponsor-developed benchmarks
play given potential concerns around independence and optimization to the test?

Data contamination, benchmark saturation, and limited real-world representativeness are distinct
problems and should be addressed separately.
Data contamination is primarily a benchmark security problem and can be mitigated by releasing a
limited public development set while keeping the larger evaluation set secure. Major benchmarks
outside medicine, such as MMLU and ARC-AGI, have already adopted this approach.

Saturation is not inherently a problem and instead reflects the rapidly progressing nature of model
capabilities. While there certainly is some degree of “benchmaxxing”—the process by which certain
model families may be overoptimized on certain benchmarks rather than unmeasured tasks—
benchmark saturation appears to result from GenAI improvements. Therefore, the answer to
saturation is creating a flexible enough regime that can adapt as the science of evaluation also
adapts. For example, state-of-the-art benchmarks just one year ago involved vignettes or
completed conversational turns (such as HealthBench). However, as agentic capabilities of models
have increased, the field has rapidly developed new methods of benchmarking agentic retrieval and
reasoning over EHRs, and is currently developing benchmarks for computer tool use of EHRs
(PhysicianBench/PhysicianBenchPro).

The third and final category, however, is by far the most challenging. Construct validity—the degree
to which a benchmark measures what it intends to measure—is related to, but still a distinct
concept from, real-world efficacy and effectiveness. For example, researchers have studied the
content of high-quality clinical conversations for decades, and multiple psychometric measures and
rubrics define their quality. Creating a benchmark using these existing scales, or validating a new
benchmark with well-established consensus methods, can create a multi-turn benchmark that rates
the overall quality of a patient-facing system. Many current benchmarks completely forego
construct validity by ignoring existing standards on concordance and appropriateness. However,
even when benchmarks are designed with scientific gold standards, this does not mean the same
system will lead patients to make better health decisions. Simply using established
frameworks/ontologies does not guarantee construct validity, but disregarding them should be a
red flag. In addition, meta-analyses have found that current benchmarks rarely use data drawn
directly from real patient charts (Bedi et al. 2025), and do not capture the complexity of real-world
practice, including coordination across care teams, clinical ambiguity, and loss to follow-up (Wu et
al. 2026) though this is likely to change as benchmarking improves.

The goal of pre-deployment benchmark use, therefore, is not real-world efficacy, but to
appropriately measure what the sponsor claims it can do. This distinction is important. A
competency-based framework will require a defined taxonomy of clinical constructs, likely drawing
on the standards described in Question 9, from which relevant benchmarks can be selected or
developed. The intended construct and evaluation criteria should be specified a priori. There should
not be a single “master” benchmark intended to determine overall performance. Once the
competency of interest is specified, established psychometric, clinical, and increasingly
computational methods can support development and validation. For example, emerging efforts
such as the Medical AI Superintelligence Test (MAST) project curate clinically grounded benchmarks
across multiple domains, including clinical reasoning (including uncertainty and safety), multimodal
image interpretation, and agentic capabilities (ARISE ).

Construct validity, however, is only the first step. The CDRH should encourage linking benchmarks
to real-world efficacy in the post-deployment phase. As the pathway moves forward, individual
benchmarks can be linked to post-deployment data, which can provide important information
about which constructs are most closely linked to real-world performance. Importantly, a lack of
linkage to real-world performance does not inherently mean the linkage is invalid. Real-world use
can fail for many reasons (such as workflow integration), many of which go beyond the device's
capabilities. These approaches are not substitutes for clinical trials.

11. CDRH is considering that clinical confirmation for a GenAI-enabled device might not require a
prospective clinical study in every case, and has described above a range of approaches of
increasing rigor and patient exposure. How might a sponsor select and justify a confirmation
approach tailored to a device’s intended use and proportionate to the device’s risk profile? Are
there device types or risk profiles for which one or more of these approaches would be insufficient
or inappropriate? How might the anticipated distribution of real-world inputs be taken into
consideration? Are there other methods of clinical confirmation that might help inform the
evaluation of GenAI-enabled devices?

We broadly agree that not all clinical use cases require clinical trials, and there are many ways in
which clinical trials are insufficient given the general-purpose nature of workflows augmented by
GenAI and the continued pace of model development.

Several inputs must be considered when deciding a confirmation approach. The first is the risk of
harm. Many AI-powered interventions have very low risk of harm and may warrant limited clinical
confirmation, such as a device intended to counsel patients about diet and exercise. At the other
end, devices that influence high-risk care (such as a system operating at ICU-level care) will require
rigorous clinical assessment such as a randomized clinical trial. In practice, most interventions fall in
between, where benchmarking and post-deployment measures, tagged to existing ontologies, may
be appropriate. However, no frameworks currently determine where those cutoffs might be.

The second factor is the degree of human oversight. Systems that are designed as human-in-theloop (HITL)—that is, every output is confirmed by a human professional—likely require less
intensive modes of confirmation, while human-on-the-loop (HOTL) or fully autonomous systems
have a higher threshold. HOTL systems are not a monolith either, with a range of different
interaction systems that confer different risks. A regulatory regime must account for the variety of
interaction mechanisms and their implications for evaluation.
Thirdly, clinical confirmation should reflect populations, disease prevalence, care settings, and
workflows expected under the device’s intended use. Some GenAI may be deployed in settings with
increased heterogeneity and require broader validation.

In general, we support three graded approaches: pre-deployment benchmarking, post-deployment
monitoring, and clinical trials. While other proposed methods of model safety exist, such as “model
cards,” which give aggregate performance, or “centers of excellence,” we do not believe that these
are sufficient regulatory mechanisms. Model cards do not give comprehensive information and lack
construct validity. Centers of excellence cost far more than a benchmarking mechanism, yet give far
less information than clinical trials.

13. For which clinical domains, device functions, or subpopulations is synthetic data particularly
well-suited, or particularly inadequate, as a supplement to real-world evidence? What safeguards
would mitigate the risk that synthetic data generated by models of the same class as the device
under evaluation reproduces the very performance gaps the evaluation is intended to detect,
particularly for underrepresented subgroups?

Synthetic data is widely used in benchmark evaluations and evaluation development. Its usage falls
along a spectrum, so it's important to distinguish the nature and degree of synthetic data use. A few
major pitfalls of synthetic data must be considered: circularity, bias, and limited real-world
representativeness.

Synthetic Class Description Pitfalls Appropriate Effort
uses

De novo Generated by LLMs or Highly susceptible to Known-hazard Minimal
generation (fully similar models from zero- circularity, bias, and probes
synthetic) shot prompting (e.g., limited real-world
"Generate realistic representativeness.
patient cases that a rural
physician may encounter
in routine practice").

Population- Synthetic data informed May miss clinical Rare-scenario Moderate
informed (fully by population-level data. nuances and individual coverage
synthetic, patient complexities.
knowledge- Subgroup combinations
driven) may be wrong even
when overall
prevalence matches.
 Synthetic Class Description Pitfalls Appropriate Effort
uses

Reconstruction Synthetic data based on Requires rigorous Performance High
(partially individual patient charts validation to ensure estimates
synthetic) processed through complete
deidentification and deidentification while
reconstruction pipelines avoiding data loss.
to prevent Circularity risk if an LLM
reidentification while performs the
maintaining fidelity. reconstruction.

Circularity arises when a model of the same class as the device writes, selects, or grades test items,
so the evaluation inherits the device's blind spots. Bias arises when generators reproduce
inaccurate pictures of patient populations, which impacts underrepresented subgroups the most.
Limited representativeness arises when synthetic cases lack the ambiguity and incompleteness of
real records. Moreover, every real case, however unusual, has actually occurred, and reconstruction
alters only non-essential details. A generated case, however plausible, describes a sequence of
events that may never have occurred and may be biologically impossible. Physician review can
catch implausible details but cannot confirm that such a sequence is itself possible, except for fairly
generic presentations. For the same reason, a physician's inability to distinguish a synthetic case
from a real record cannot be the bar for validity.

De novo data is therefore best suited to probing known hazards, such as prompt injection or missed
escalation, while population-informed data is best suited to cover rare scenarios. Neither should be
used to estimate real-world performance: such estimates should rest on real data or on
reconstructed data validated against its source records. Synthetic data alone should not establish
performance for any subgroup. Sponsors should document the models used at each step, extend
the independence requirement for LLM adjudicators to generation and filtering, have clinicians
validate content and reference standards, and compare results against held-out real cases by
subgroup. Validation on held-out real patients is a long-standing standard in both machine learning
and clinical prediction. It is the ultimate test of a synthetic dataset's validity as the only check whose
content no model has shaped.
14. For open-ended device outputs, how should performance comparators and acceptance
criteria be selected? When a panel of qualified clinicians serves as the comparator, how should
the applicable standard (for example, the standard of care versus the performance of a median
clinician in practice) be defined and justified? Should generalist or specialist physicians be used as
a performance standard? When should human-AI team performance, rather than the device
operating alone, serve as the basis for evaluation?

Benchmarking rests on a fundamental assumption: that experts set the standards by which we
judge systems (or other humans) through concordance. This assumption has always been imperfect.
There is wide variation in practice behavior, even in the age of “evidence-based medicine,” and
expert concordance is usually modest at best. These are not new issues, and evaluation specialists
have developed numerous methods to increase the reliability of concordance measures. However,
GenAI systems, especially agentic ones, challenge many of these assumptions, and may differ in
meaningful ways from human experts.

While “move 37” for AlphaGo is often cited here, more prosaic examples can show the limits of this
approach. Agentic AI systems can now perform large-scale retrieval across patient charts, much of
which is impossible for an individual clinician to read. The information necessary to make a medical
decision, say, the appropriate first antihypertensive to start, will naturally be based on the chart
review that a physician can reasonably accomplish. However, because AI systems can parse far
more information, their outputs may differ considerably from humans in ways biased toward AI
output.

While theoretical approaches to transcend human expert concordance (such as outcomes-based
labeling) exist, they are currently hypothetical and should not form the basis of a regulatory regime.
For the time being, we believe that the human baseline for AI systems should be set with the
highest possible standards – specialist physicians without time restrictions, under idealized
conditions. While this does not abrogate the underlying evaluation problem, setting the highest
standard gives this approach more time to work. While we consider this the bar, it may still be
useful to concurrently measure average physician performance under real-world practice
conditions. Such data can contextualize performance and understand existing trade-offs, similar to
comparisons of autonomous vehicles against average human drivers—not professional drivers—in
terms of crash rates per mile.

When a device is intended for collaborative use with clinicians, human–AI team performance should
also be evaluated directly. The relevant question is not only whether the model performs as well as
a clinician in isolation, but whether the human–AI dyad performs better and more safely than a
clinician working without the device (Rodman and Zwaan 2026). Standalone model performance
remains important for characterizing failure modes, but the human–AI team should be the primary
unit of clinical evaluation when it is the intended use.
As clinicians increasingly train and practice with AI, the human comparator may also change over
time. For example, cognitive offloading and deskilling may alter physicians’ independent abilities.
GenAI systems may also exceed expert performance on specific tasks, making specialist
concordance a less meaningful ceiling (Goh et al. 2026). Long-term evaluation should therefore
consider how AI changes clinician competence and whether expert performance remains an
appropriate reference standard.

15. Are there ways in which performance might be assessed relative to the care, technology, or
course of action likely to occur in the absence of the device, rather than to the comparators
described in this section? What approaches might be used to identify and justify a comparator
such as unaided clinical judgment, delayed specialist review, or no intervention?

Yes, clinical confirmation should assess a device against the care pathway patients would
reasonably experience in its absence. This comparison addresses a different question from
competency benchmarking: benchmarks characterize whether a system can perform specified
clinical tasks, while comparison with usual care assesses whether introducing the device improves
patient-relevant outcomes.

The appropriate comparator depends on intended use. For clinician-facing systems, it may be
clinicians providing usual care without the device. For systems intended to expand access, it may be
delayed specialist review, existing triage services, or the care available while patients await
evaluation. “No intervention” should be used only when it accurately represents the alternative;
patients may otherwise seek care, use other technologies, or manage symptoms themselves.

Sponsors should justify the comparator using evidence from the intended population and setting,
including clinical workflows, referral completion, wait times, service availability, and existing careseeking behavior. Where feasible, comparative clinical studies should measure outcomes across the
relevant care pathway, including time to appropriate care, harms from omission and commission,
and downstream healthcare use. Comparisons should account for differences in patient populations
and access to care rather than attributing those differences to the device. Performance relative to a
constrained care pathway should contextualize clinical benefit without replacing an independently
justified minimum standard of safety and effectiveness.
16. Should independent third parties be involved in some or all aspects of a competency-based
approach, including device benchmarking and clinical confirmation? If so, in what ways might
qualified, independent third-party participation contribute to this approach, and what
qualifications and independence criteria should apply? Are there aspects of the assessment for
which third-party involvement would be impractical or inadvisable? What safeguards or program
design features would be critical to prevent third-party participation from limiting competition or
preventing innovation?

We support qualified, independent third-party participation in a competency-based approach,
particularly in benchmark development, validation, and administration. Evaluating GenAI-enabled
devices requires a combination of clinical, technical, and measurement expertise. Shared
independent infrastructure could improve the credibility and comparability of evaluations, reduce
duplication, and help evaluation methods keep up with technological advances.

Physician and patient stakeholders should participate in defining the clinical constructs benchmarks
assess, which include defining the relevant patient contexts and potential harms. These
specifications should be established before benchmark development, with independent scientific
experts responsible for ensuring that the proposed measurements are defensible. Independent
academic groups and other qualified organizations could develop benchmarks using these
specifications, while a separate evaluation group assesses benchmark quality and administers
secure testing. Benchmark performance should remain distinct from evidence of clinical efficacy.
For clinical confirmation, third parties could contribute to study design, independent outcome
adjudication, and analysis of relationships between benchmark results and clinical performance.

Third party qualifications should reflect the assigned role and demonstrate relevant clinical
expertise, experience in evaluation science, statistical and technical competence, and the ability to
protect sensitive patient data and confidential testing assets. Independence should be
demonstrated through disclosed conflicts of interest, scientific autonomy, and safeguards
preventing sponsors or funders from controlling methods, suppressing unfavorable findings, or
influencing evaluations. Industry funding need not preclude independence if these protections are
credible and enforceable.

Third-party involvement should be proportionate to the device’s intended use and risk, rather than
be required for every development activity. Mandating external review of every internal
experiment or iteration could introduce unnecessary delays. Clinical confirmation also requires
expertise in the intended care setting and workflow; an external evaluator should not substitute
generic benchmark performance for context-specific clinical evidence. Manufacturers should retain
accountability for their devices, and FDA should retain regulatory decision-making authority.

To support competition and innovation, evaluation specifications and quality criteria should be
transparent, while held-out testing assets remain protected. Access should be available on
equitable terms, with affordable or subsidized options for startups and smaller developers.
Programs should avoid exclusive dependence on a single evaluator or proprietary benchmark,
permit scientifically justified alternative methods, and provide a fair appeals process. Benchmarks
should be updated as capabilities and evaluation science advance. Properly designed, independent
evaluation can support innovation by providing credible evidence of clinically meaningful
performance and clearer targets for development.