RadFirst PLLC (Nick Tanev)

RadFirst PLLC, an independent teleradiology practice, offers a clinical deployer’s perspective on FDA’s discussion paper. Our attached comment addresses radiologists working across multiple deployments, relevant clinical expertise, feasible postmarket monitoring, and coordination among manufacturers, client facilities, and reading practices while preserving manufacturer accountability. We also recommend clear distinctions between configuration, local evaluation, and model modification, together with assessment of automation bias through evaluations in which clinicians encounter AI outputs under realistic reading conditions.

RadFirst Comment on Generative AI Medical Devices
Docket No. FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback
Submitted by Nick Tanev | EVP, Healthcare Alliances | RadFirst PLLC | Chicago, Illinois

I. Basis of interest

RadFirst welcomes CDRH’s request for feedback. We are commenting from our role as a clinical
deployer. We recommend an approach that accounts for clinicians working across multiple deployments,
places monitoring capabilities with the parties able to provide them, and clearly distinguishes local
evaluation from device modification. Our comments address practical implementation in an independent
radiology practice.

RadFirst provides subspecialty radiology coverage to hospitals, outpatient imaging centers, and
physician-owned groups. We work with a network of approximately 40 radiologists across 15 facilities,
including full radiology department oversight at a Level II trauma and stroke center. We operate a
distributed reading practice in which physicians interact with AI-enabled tools within client systems we
do not own. This experience informs our observations about the proposed evaluation and monitoring
approaches; we do not suggest that all of those tools are generative AI devices.

II. Readers moving between deployments
Questions 1, 7 and 11

The paper appropriately considers the context in which a device is used. We ask that evaluation also
account for readers moving between deployments, rather than examining each environment only in
isolation.

A radiologist in our practice may read studies from multiple client facilities during a single shift,
encountering different AI tools, configurations, and ways of presenting output. Expectations formed while
reading for one facility may carry into work for another. A reader accustomed to frequent flags at one site
may respond differently at a site with more conservative settings or no comparable tool.

This is a potential interaction risk, not an effect whose magnitude we have established. Evaluations
should consider whether representative users encounter multiple configurations and whether they can
recognize the active tool, its scope, and relevant limitations. Characterizing a single deployment remains
useful, but may not fully describe the experience of a physician who moves among several environments.

III. Relevant expertise and user interaction
Question 4

Our practice routes studies to subspecialty-trained readers. That experience informs our concern that an
AI output may serve different functions depending on the user’s relevant expertise. A specialist may
compare the output against established domain knowledge, while a clinician outside that specialty may
rely on it to supply knowledge or judgment they do not independently possess. Evaluation should
consider that difference alongside the user’s ability to detect an incorrect output.
The Mammography Quality Standards Act provides a bounded example of domain-specific interpretingphysician qualifications, including training, continuing education, and interpretation experience. For
example, the continuing qualification requirements include at least 960 interpreted or multi-read
mammographic examinations over the applicable 24-month period and at least 15 category I
mammography CME units taught or completed over the applicable 36-month period, under 21 CFR
900.12(a)(1)(ii)(A)-(B). Its value here is as an analogy to examine, rather than a structure to transfer
directly to devices. Experience thresholds measure exposure rather than diagnostic accuracy, and
qualification tracking creates administrative work. FDA should consider both the benefits and limitations
of such measures when developing competency-based device evaluation.

IV. Postmarket monitoring in independent practices
Questions 18, 19 and 21

We agree that premarket evaluation is unlikely to resolve every question about performance after
deployment. Question 21 expressly asks how other stakeholders can contribute without diffusing
manufacturer accountability. From the perspective of a small reading practice, the three approaches in
Section VI.A require support that cannot simply be assumed to exist locally.

Periodic re-benchmarking requires suitable evaluation cases, reference standards or expert adjudication,
thresholds, and analysis. Independently building and maintaining that infrastructure is not realistic for our
practice. Manufacturer-supported or shared evaluation arrangements could make participation more
feasible.

Sample-based clinician review requires protected physician time to examine previously interpreted
studies. We are not aware of a dedicated reimbursement mechanism for this work in our setting. A
feasible approach should identify how review is funded and how sample size and cadence reflect the risk
and purpose of the assessment.

Performance degradation monitoring requires access to relevant data and instrumentation. Our practice
does not own the client PACS environments in which those data reside and cannot independently
authorize access or install monitoring capabilities. Any approach that depends on those actions must
involve the facility and the relevant technology providers.

If each deploying organization must independently assemble a complete monitoring program, the burden
could limit adoption in smaller practices and hospitals. Institutional size alone does not establish whether
a practice can safely perform a defined activity. Independent practices can contribute clinical observations
and local evidence when the necessary tools, agreements, and support are available. Their participation
can broaden the settings represented in postmarket evaluation.

V. Responsibilities and operational control
Question 21

We recommend explicit, coordinated responsibilities for manufacturers, client facilities, and reading
practices. The arrangements may differ by deployment, but each monitoring function should have an
identified owner, necessary access, and a defined route for escalation.

Manufacturers should provide monitoring capabilities proportionate to device risk and intended use.
These should include logging, version and configuration traceability, support for detecting potential
degradation, and methods for re-benchmarking when appropriate. Such capabilities may be built into the
device or supplied through an associated service. Manufacturers should lead investigation of potential
device defects and maintain clear responsibilities for updates, user notification, and corrective action.

Client facilities control or arrange access to their clinical systems and data. Where monitoring depends on
data access or transfer, the necessary permissions and technical arrangements should be established with
the parties that actually control them, consistent with applicable privacy and security requirements. A
reading practice should not be assigned an obligation to grant access it does not possess.

Our contribution is clinical: identifying unexpected behavior, documenting the circumstances, and
responding to findings. Maintaining the technical infrastructure for ongoing device evaluation requires
support from manufacturers and the facilities that control the relevant systems and data.

Consider a radiologist who notices unexpected AI behavior during an overnight shift: the reporting
pathway should allow the physician to flag the concern without leaving the reading environment.
Reporting tools should be available within the reading workflow and allow an initial report in seconds,
with relevant device state captured automatically where feasible. Follow-up can collect further detail.
Reducing the effort required to flag a concern would support more consistent reporting; a low volume of
reports should not by itself be interpreted as evidence that problems are absent. This proposal concerns
the usability of reporting and does not replace applicable reporting obligations.

Mosaic Clinical Technologies’ citizen petition, FDA-2026-P-9175, filed August 12, 2026, recognizes
both provider responsibilities and the difficulty of carrying them out without adequate performance
information and support (pp. 2-3 and 8). Our recommendation makes that allocation operational:
responsibilities should reflect what each participant controls, while preserving manufacturer
accountability for the device.

VI. Local activities and change traceability
Questions 22, 23 and 24

We ask CDRH to distinguish configuration within manufacturer-specified parameters, evaluation of local
performance without altering the device, and modification of the model itself. Measuring performance
does not itself change device output. Configuration or model changes may do so and require a different
assessment. Neither an intended-use authorization nor the existence of a PCCP should be read as
permission for every local modification.

For Question 22, the evaluation required should reflect what changed, the functions and populations
potentially affected, and the associated risks. Manufacturers should explain which activities fall within the
existing authorization, which modifications are covered by an authorized plan, and what evidence and
acceptance criteria apply. Local users need a practical way to determine when to stop and seek further
assessment.

For Question 23, FDA should clarify how modifications that cannot be fully prespecified, or fall outside
an authorized plan, should be assessed before implementation. That clarification should address
responsibility for determining whether additional regulatory review is needed, the evidence supporting the
change, notification to affected users, and safeguards such as staged implementation or rollback where
appropriate. An existing plan should not be treated as blanket coverage for an unanticipated change.
For Question 24, clinical users may notice altered behavior without knowing its cause. A vendor release,
third-party model update, local configuration change, or shift in case mix may each contribute.
Attribution requires cooperation: manufacturers provide device and upstream model history, while
facilities and reading practices contribute relevant local information.

Manufacturers should retain enough information to reconstruct the device state associated with a
particular output, including relevant model versions and configurations. Local changes should also be
documented by the party making them. Upstream update notification and change-control arrangements
should support timely investigation. Traceability can make field reports substantially more useful and
help distinguish a device change from a change in the setting where it is used.

VII. Automation bias and reading conditions
Element E.4

We support attention to automation bias. Reading volume, time pressure, and fatigue may affect whether
users independently scrutinize a fluent output. Evaluation should therefore include representative
workloads and users rather than relying exclusively on unhurried review.

The paper’s shadow deployment approach keeps outputs hidden from clinicians and patients. It can help
characterize device performance in a real workflow, but cannot directly measure how seeing an output
changes a clinician’s judgment. Assessment of automation bias should include an appropriately designed
evaluation in which representative users encounter the output, such as a controlled simulation under
realistic reading conditions. That evaluation should examine error detection and inappropriate acceptance
as well as task performance. Any clinical investigation must meet its applicable oversight requirements.

VIII. Closing

We appreciate CDRH inviting feedback at this early stage. A framework that makes manufacturer
accountability operational, recognizes divided control in distributed care, and supports proportionate local
participation can improve both the quality and representativeness of evidence. RadFirst would welcome
the opportunity to provide further detail about the implementation issues described here.

Respectfully submitted,

Nick Tanev
EVP, Healthcare Alliances
RadFirst PLLC

Martin Urban, MD
Chief Medical Officer
RadFirst PLLC