Premier Inc. (John Knapp)

See attached file(s)

October 19, 2026

Kyle Diamantas, J.D.
Acting Administrator
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

Submitted electronically to: http://www.regulations.gov

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback

Dear Acting Administrator Diamantas:

Premier appreciates the opportunity to submit comments to the Food and Drug Administration (FDA) on its
discussion white paper and Request for Information (RFI): Considerations for the Regulation of Generative
AI-Enabled Medical Devices. Premier supports federal efforts to responsibly leverage artificial intelligence
(AI)/super intelligence (SI) to improve patient health, quality and outcomes, as well as access to innovations
in safe and effective therapies. Premier also appreciates FDA’s proactive approach to gather feedback from
stakeholders as it determines the best regulatory framework for generative SI-enabled medical devices
(GenSI-enabled devices).

In our detailed comments below, Premier offers the following recommendations:
• Require device manufacturers to demonstrate that benchmark data reflect the populations and care
settings where the device will be deployed, explain and justify material differences, and use multisite, real-world datasets when feasible;
• Use qualified, independent third parties to strengthen evidence generation and evaluation, while
preserving FDA's authority to determine whether a device is safe and effective and making the
manufacturer responsible for the device;
• Permit reduced premarket clinical evidence only when postmarket monitoring is enforceable,
grounded in a rigorous baseline, and fully funded and controlled by the manufacturer;
• Use periodic re-benchmarking, sample-based clinician review, performance degradation
monitoring, and multi-site comparative monitoring to operate as complementary components of
postmarket oversight;
• Make postmarket monitoring event-driven first and calendar-driven second, with risk-based
periodic reviews, independent validation of any automated supervisory agents, and timely
manufacturer notification to healthcare institutions of material device or model changes; and
• Clearly distinguish stakeholders' capacity to observe and report device performance from the
manufacturer's nondelegable duty to monitor and remain accountable for safety and effectiveness.

I. BACKGROUND ON PREMIER INC.

Premier Inc. is a leading healthcare improvement company and national supply chain leader, uniting an
alliance of more than 4,350 hospitals and approximately 325,000 continuum‑of‑care providers nationwide.
Premier leverages integrated data, analytics, collaboratives, and technology solutions to enable better care
and outcomes at lower cost. Premier has experience applying SI-enabled technologies to improve quality
of care and patient outcomes in health care settings. Premier Clinical Decision Support (CDS) leverages
natural language processing technologies to analyze unstructured clinical data and apply evidence-based

200 Massachusetts Avenue NW, Suite 330, Washington, DC 20001 | premierinc.com
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and RFI
Page 2 of 6 pages

guidelines to generate real-time alerts and analytics that improve care quality and reduce costs. The
Premier Quality Enterprise platform also offers health systems targeted quality, financial and regulatory
reporting insights generated through the integration of SI technology. Finally, Premier operates an AI Value
& Governance Collaborative, through which health systems assess SI governance maturity, establish a
baseline before deploying SI technology, and reassess these capabilities regularly against that baseline.
The aim of the collaborative is to produce comparable measurements of maturity across organizations
rather than single-site case studies.

II. A COMPETENCY-BASED APPROACH FOR PREMARKET EVALUATION OF GENSI-ENABLED
DEVICES

Benchmarking (Question 10)

The FDA recognizes that publicly available benchmarking assets may be subject to data contamination,
saturation, and limited real-world representativeness. Premier believes that the benchmarking challenge is
fundamentally a deployment setting challenge. Meaning, a benchmark assembled without data from all the
care settings in which a device will operate cannot establish a baseline for measuring or predicting realworld behavior. As Premier noted in comments to the National Quality Forum, models trained or evaluated
at a single health system or provider network may not generalize to populations that differ by region, income,
prevalence of chronic conditions, workflow or data environment. The FDA should therefore expect sponsors
to describe how benchmark populations and care settings compare with the intended deployment
population and to justify material differences. Benchmark datasets should draw from multiple care settings
when feasible. Organizations holding national, multi-site clinical and administrative data are well-positioned
to support this work.

The Role of Third Parties (Question 16)

The FDA asks about the role of independent third parties to aid FDA’s competency-based processes for
GenSI-enabled device premarket evaluation, e.g., by maintaining sequestered benchmark datasets;
conducting or evaluating parts of the assessment; serving as expert adjudicators; developing FDA-qualified
tools; or creating standards for the FDA. The white paper shares examples of existing programs, such as
the Accreditation Scheme for Conformity Assessment (ASCA) and the Medical Device Development Tool
program, as potential analogues.

Premier supports participation of third parties to aid and facilitate FDA’s regulatory processes but
cautions that the FDA retain final decision-making authority about a GenSI-enabled device’s safety
and effectiveness. While third parties could supply the infrastructure for approval processes, the FDA
should maintain responsibility for regulatory decisions, and the sponsor should retain responsibility for the
device. Independent third parties can add the greatest value in three areas: maintaining sequestered
benchmark datasets, serving as expert adjudicators, and developing standards that FDA can recognize.
These functions can both streamline evaluation processes and assure fairness and integrity of final
decisions. Where evaluation standards already exist, the FDA should consider recognizing them rather than
creating duplicative federal specifications.

However, there are two areas that are poor fits for third party involvement. First, third parties cannot
replicate clinical confirmation in a live workflow because that data needs to come from the deploying
institution using the GenSI-enabled device. Workflow integration, clinician use under time pressure, and
the behavior of users who may over-rely on or disregard an output cannot be fully reproduced in a
laboratory. A third party may design the protocol, hold data or analyze results, but observation must occur
where the device operates. Second, no third party should determine whether a device's benefit-risk profile
supports marketing authorization. That judgment requires application of the statutory standard and should
remain with FDA. Submissions should clearly identify which portions of an assessment were performed or
adjudicated by a third party, as well as the limits of that work.
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and RFI
Page 3 of 6 pages

To create an efficient regulatory process, the FDA should build on market-based evaluation models and
allow multiple pathways for qualified independent participation. For example, Premier operates a national
vendor assessment process in which digital health vendors seeking a Premier group purchasing
organization (GPO) contract complete more than 100 questions addressing security practices, independent
third-party cybersecurity attestations or certifications, company governance, and the use of SI. Premier
members may review that information before contracting, and Premier provides model contract language
addressing SI liability and vulnerability disclosure. This process does not assess clinical competency and
is not a substitute for FDA premarket review. It does, however, demonstrate that structured third-party
disclosure and attestation can operate at scale without a single federal accreditation gatekeeper. Premier
urges the FDA to provide interested third parties with flexibilities in how they structure their
processes, provided they meet the FDA’s requirements.

Additionally, FDA should take steps to avoid reliance within the evaluation process on a single third-party
entity, a scenario which diminishes robust competition. Such an entity could become a de facto toll booth
that advantages sponsors able to absorb its costs and timelines. To maintain competition among third
parties, Premier recommends that the FDA should:
• Recognize multiple qualified bodies for any given assessment function rather than designating a
sole third party and structure qualification criteria so that new entrants can qualify as the field
develops;
• Preserve a pathway for sponsors to submit evidence directly to the agency so that third party
review remains an option rather than a required toll;
• Publish the qualification criteria it applies to third parties along with expected review timelines; and
• Monitor whether costs or wait-times at recognized bodies begin to function as a barrier for smaller
sponsors and retain authority to qualify additional bodies in response.

III. POSTMARKET MONITORING FOR GENSI-ENABLED DEVICES

Reliance on Postmarket Monitoring v. Premarket Certainty (Question 18)

The FDA asks under what conditions is it appropriate to justify reduced premarket evidence when approving
GenSI-enabled devices.

Premier acknowledges that a certain level of premarket uncertainty is necessary for GenSI-enabled device
approval, given the challenges of testing and validating their safety and efficacy prior to deployment.
Reliance on postmarket monitoring is a defensible tradeoff to premarket uncertainty when the
monitoring program is enforceable, the baseline is rigorous, and accountability for device safety
remains with the manufacturer. Premier offers the following principles for postmarket surveillance of GenSI enabled devices.

• Make monitoring an enforceable condition of device authorization. A monitoring commitment
without consequences for nonperformance is not evidence. The FDA should require the sponsor
to identify the party accountable for postmarket performance; define performance measures,
desired outcomes, methods, acceptable ranges, and monitoring cadence before deployment;
specify actions that follow a measure falling outside an acceptable range; and prespecify
thresholds for suspension, withdrawal or other regulatory intervention. The FDA should also state
the consequences when a sponsor fails to monitor, misses a required cadence or does not act on
a detected signal.
• Manufacturers should assume the cost of postmarket monitoring. Reduced premarket
evidence lowers a manufacturer's time and cost to market. The corresponding monitoring
obligation should remain with the manufacturer. The risk of transfer is greatest in the purchase
agreement, which the FDA does not review. A monitoring program can be described to FDA as
the manufacturer’s and then carried out, under the contract, through customer-supplied
performance data, audit log review, and staff time that FDA does not see at authorization. When
a monitoring program depends on data or labor from deploying institutions, the submission should
identify that dependency and demonstrate that the sponsor has arrangements and resources
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and RFI
Page 4 of 6 pages

necessary to execute the program. A program the sponsor cannot perform without uncompensated
customer effort is not a program the sponsor controls.
• Require a rigorous and usable performance baseline. Degradation cannot be detected without
a reference point. The FDA should treat robust premarket benchmarking as the floor, even when
clinical confirmation is appropriately reduced. Benchmarking and clinical confirmation should be
separately scalable rather than treated as one interchangeable pool of evidence. The FDA should
also expect sponsors to provide deploying institutions with a usable description of the performance
baseline and acceptable range. Without that information, a health system cannot determine
whether an apparent error reflects expected behavior or emerging degradation.
• Reversibility and detection latency belong in the risk framework. Postmarket monitoring can
substitute for premarket certainty only when an incorrect output can be recognized by someone
positioned to observe it. The FDA should require sponsors to identify who would detect an error,
through what signal, and within what time period. If no concrete answer exists, reduced premarket
evidence is inappropriate. The FDA should also consider detection latency alongside reversibility:
the period between an incorrect output and recognition of that error, measured against the rate at
which harm may accumulate. A slowly detected but reversible error may affect many patients
before intervention, while an immediately detected irreversible error may at least be contained
after the first event. Premier recommends that the FDA incorporate both reversibility and detection
latency as companion dimensions into the Section IV framework and use them to determine where
reduced premarket evidence is available. Together they establish whether postmarket monitoring
can substitute for premarket certainty.
• Incorrect outputs must be 1) be detectable and 2) not severely harmful or irreversible.
Reduced premarket evidence should not be available when users cannot independently evaluate
outputs, when failure is unlikely to generate a timely observable signal, or when reliance on an
incorrect output may cause high-severity or irreversible harm before detection. If a clinician or
patient cannot independently evaluate whether an output is correct, because the output is not
verifiable against anything the user observes, the monitoring premise fails and reduced premarket
evidence is not appropriate regardless of the severity of the consequence. Premier recommends
that the FDA require a sponsor to identify who would notice an incorrect output, through what
signal, and within what period, and treat a device for which that question has no concrete answer
is an inappropriate use case.

Approaches to Postmarket Performance Evaluation (Question 19)

The FDA seeks feedback on postmarket performance evaluation approaches. Premier supports periodic
re-benchmarking, sample-based clinician review, and performance degradation monitoring, and
recommends that FDA generally expect these approaches to be used in combination rather than as
alternatives. Each addresses a distinct failure mode. Re-benchmarking measures change against a known
reference but only for covered behaviors. Clinician review can identify unanticipated problems but is
constrained by sample size and retrospective evaluability. Drift monitoring detects changes between review
cycles but may not explain their cause. When a sponsor omits one of these elements, FDA should expect
a risk-based rationale.

The FDA should also recognize comparative monitoring across multiple sites as a distinct
postmarket approach. A performance decline in one health system may result from the device, the local
patient population, the data environment, or workflow integration. Cross-site comparison can help
distinguish these causes, but only when common measure definitions, consistent capture methods, and
appropriate adjustment for differences in patient populations and care settings are used. Aggregating data
from multiple sites without comparable measurement does not create a valid comparison. FDA should
expect the measurement methodology and comparison population to be specified, whether monitoring is
performed by the manufacturer across customers or by an independent party.

The cadence for postmarket performance evaluation should be event-driven first and calendar-driven
second, with a minimum periodic review that applies even when no triggering event is reported and is scaled
to the device's risk. The absence of a trigger is not proof of stability. Instead, reassessment should be
triggered by:
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and RFI
Page 5 of 6 pages

• Any change to the underlying foundation model, including a change initiated by a third-party
developer;
• Changes to prompts, retrieval strategies, guardrails, orchestration logic, or other components that
shape device behavior;
• A material shift in the input population, deployment setting, workflow, or data environment;
• A drift metric crossing a prespecified threshold;
• A cluster of user-reported concerns, complaints, near misses, or adverse events; and/or
• A change in the clinical standard of care or underlying evidence reflected in the device's outputs.

If a sponsor uses a machine-based supervisory agent to satisfy monitoring obligations, then the FDA should
require evidence of the agent's independent detection performance against known failures. A supervisory
agent built on the same model class or architecture of the device may share the monitored device's blind
spots. The FDA should therefore consider model and architectural diversity, independence of evaluation
data, and validation against known failure modes. An agent that merely reproduces the device's reasoning
provides monitoring in form rather than substance.

Lastly, event-driven monitoring depends on deploying institutions knowing that a relevant change occurred.
The FDA should expect manufacturers to notify deploying institutions of material changes, including
changes to an underlying third-party foundation model, on a defined timeline and through a channel that
reaches the institution's clinical, technology and SI governance functions. Notices limited to procurement
or routine support contacts may not reach the personnel responsible for monitoring and risk management.

Additional Stakeholder Roles in Postmarket Monitoring (Question 21)

The manufacturer holds the monitoring duty and remains accountable for postmarket safety and
effectiveness. Other stakeholders possess information or expertise the manufacturer cannot supply alone.
Consequently, the FDA should structure their contributions as inputs to the manufacturer's obligation, not
as a transfer of that obligation.

Premier recommends that the FDA define the expected contribution of various stakeholders as follows:
• Healthcare institutions and clinicians: Observe performance in real workflows and across actual
patient populations. Their role is to report what they observe, not to certify that a device remains
safe and effective. Manufacturers should define reportable deviations and provide practical
reporting channels.
• Independent parties with multi-organization data: Can produce comparable measurement
across sites and distinguish device-level from site-level problems, provided common definitions,
documented data-capture methods and appropriate adjustment are used.
• Professional societies: Can define clinically adequate performance within a domain and identify
when the clinical standard reflected by a device has changed.
• Standards-setting bodies: Can harmonize benchmark definitions, performance measures and
reporting formats so results can be compared across manufacturers and settings.
• Payers: May contribute longitudinal utilization and outcomes data that can reveal patterns not
visible to a single institution. However, the FDA would need to address data access challenges and
alignment of the payer's interests before relying on their data.
• Public-private consortia: Can convene stakeholders and sustain shared infrastructure but should
not hold the regulatory monitoring duty because they do not have the manufacturer's direct
accountability for the marketed device.

It is critical that the FDA state explicitly that participation by a clinician or healthcare institution in
postmarket monitoring transfers no portion of the manufacturer's obligation, regardless of the
extent of that participation. A manufacturer should not be permitted to discharge its monitoring duty by
assigning it contractually to customers or conditioning purchase or licensing on customers performing
monitoring that FDA expects of the manufacturer. As noted in response to question 18, this transfer
happens in the purchase agreement, outside anything the FDA reviews. Although the FDA does not
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and RFI
Page 6 of 6 pages

regulate commercial contracts generally, a clear regulatory expectation would give deploying institutions
an important reference point in negotiations.

Finally, an institution cannot report what it cannot see. When the FDA expects health systems or providers
to contribute monitoring data, it should also expect manufacturers to provide timely notification of material
changes, access to performance information and audit logs for the institution's deployment, a clear
description of the authorized performance baseline and acceptable range, and a defined channel for
reporting observations and receiving a response. These should be obligations of the manufacturer's
monitoring program, not optional services. Otherwise, the regulatory framework places cost and operational
responsibility on health systems or providers for a duty the manufacturer holds.

IV. CONCLUSION

Premier appreciates the opportunity to provide feedback on the FDA’s request for input on the
Considerations for the Regulation of Generative AI-Enabled Medical Devices white paper and looks forward
to continued collaboration with the agency. For questions, please contact Somaieh McMullan, Senior
Director for Payment Policy, at Somaieh_McMullan@premierinc.com.

Sincerely,

John Knapp
Vice President, Advocacy
Premier Inc.