SolarisRTC (Esther Abels)

Comment of Esther Abels, Founder and CEO, SolarisRTC, on the GenAI-enabled devices discussion paper, focusing on devices built on pathology foundation models (Q8–Q11, Q16–Q18, Q22–Q25). Full comment attached.

Comment Letter
Submitted by: Esther Abels, Founder and CEO, SolarisRTC
Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback. (Docket FDA-2026-N-7874)

Thank you for the opportunity to comment.

My work focuses on regulatory and clinical strategy for AI-enabled diagnostics, particularly digital
pathology and companion diagnostics. My comments address a category of device anticipated in the
discussion paper: products built on pathology foundation models, often vision or vision-language
models trained on whole-slide images (Q17).

I support FDA's emphasis on evaluating the final device as deployed for its intended use, and
evaluating related risk across the full device life cycle. At the same time, foundation models are
important safety-relevant components whose characteristics, limitations, and change-management
controls influence downstream device performance. I recommend permitting evidence regarding a
foundation model to be generated once and reused where appropriate, while maintaining devicespecific validation for the intended use and clinical claim, highlighting the collaboration between
intended users and device manufacturers when determining evidence requirements. This approach
would reduce unnecessary premarket and postmarket burden, accelerate patient access, and
maintain appropriate assurance of safety and effectiveness.

My comments focus on four areas.

1. Benchmarking should be tied to the intended use and claim, using
independent data (Q9, Q10, Q16)

FDA appropriately identifies contamination, saturation, and limited representativeness as important
risks in benchmarking. These concerns apply whether benchmarks are public or sponsor-developed.

Recommendation:
Expect benchmark datasets that are independent of model development activities and have not been
used for training, tuning, threshold selection, or foundation-model selection. The benchmark should
reflect the device's intended use, target population, deployment conditions, and clinical claim.

Page 1 of 5
SolarisRTC, Silver Spring MD
The benchmark may be sponsor-developed or maintained by an independent third party and could,
for example, consist of adjudicated datasets or longitudinal real-world clinical data. In either case,
sponsors should document provenance, representativeness, and the controls used to preserve
dataset independence. Where provenance is incomplete, sponsors should justify why the dataset
remains fit for use.

For rare diseases and uncommon clinical scenarios, fully independent datasets may be difficult to
obtain. In those situations, FDA should scale evidentiary expectations according to risk and
uncertainty while relying on appropriate postmarket monitoring and reassessment.

Based on risk and intended use, it is to be assessed what clinical evidence is needed, including
whether a clinical study is needed at all. Where one is, it may be retrospective, prospective, or in
silico, using computational evaluation with simulated or synthetic data.

In general, publicly available datasets should be used with caution in device development because
they may not fully represent the intended-use population, real-world deployment conditions, or
clinically relevant subgroups. Public datasets can be valuable for hypothesis generation, exploratory
analyses, and foundation-model pretraining, but benchmark and validation datasets used to support
regulatory claims should reflect the device's intended use, target population, and deployment
environment.

2. Foundation models should be treated as safety-relevant components
and transparently documented (Q8, Q9)

The presence of a foundation model within a device does not by itself establish safety, effectiveness,
or generalizability for a particular diagnostic use. Device performance depends on the foundation
model, downstream adaptation, control mechanisms, intended use, and validation evidence.

Recommendation:
Sponsors should identify the foundation model and exact version used in the device and provide
available information about training-data provenance, limitations, failure modes, and performance
characteristics, and link that to the intended use, such as whether the device provides informational
outputs or action-directing outputs.

As such, sponsors should be encouraged to have transparent documentation of what is known about
the foundation model and to demonstrate that the finished device performs adequately for its
intended clinical use. Evaluation should focus on device behavior and clinical performance, with the
level and type of evidence determined by the intended use, risk profile, and remaining uncertainty,
rather than by a presumption that clinical studies are always required.

Page 2 of 5
SolarisRTC, Silver Spring MD
3. Foundation-model changes require defined change-control
mechanisms (Q22, Q23, Q24, Q25)

Third-party foundation-model updates may occur independently of the device manufacturer.

In pathology and companion diagnostics, changes to a foundation model may alter downstream
outputs even when the intended use remains unchanged. Effective change management is therefore
essential.

Recommendation:

• Implement postmarket performance monitoring mechanisms capable of detecting clinically
relevant performance drift, notifying users when appropriate, and alerting the manufacturer.
Manufacturers should investigate identified signals and follow established quality-system
and change-control procedures to maintain continued safety and effectiveness of the device.
• Identify and version the foundation model used in the authorized device.
• Define, through PCCPs or comparable mechanisms, what types of changes require
reassessment, what evidence must be regenerated, and what performance thresholds must
be met.
• Support voluntary Foundation Model Master Files containing training-data provenance,
known limitations, subgroup performance information, version identification, and updatenotification commitments.
• Distinguish between version-pinned models controlled by the manufacturer and externally
managed models that may change outside the sponsor's direct control.

4. Foundation-model evidence should be reusable when evaluating
downstream devices (Q11, Q16, Q18, Q25)

FDA is considering whether greater reliance on postmarket monitoring could justify accepting more
premarket uncertainty about a GenAI-enabled device’s benefit-risk profile. Independent third parties
can support this. Academic medical centers, non-profit organizations, professional societies and
cooperative study groups can maintain evaluation datasets and benchmarks, and can independently
evaluate clinical utility, generalizability and real-world performance.

Page 3 of 5
SolarisRTC, Silver Spring MD
Building devices on the same foundation model may repeatedly assemble similar evidence regarding
model provenance, limitations, robustness, and performance characteristics. Requiring the complete
recreation of that evidence for every downstream application may add burden without proportionate
regulatory value.

Recommendation:
Permit evidence regarding a foundation model's provenance, performance characteristics,
limitations, and known failure modes to be reused when the same model version, or a modification
covered by each device’s authorized PCCP, is incorporated into multiple devices.

Such evidence may be provided through a voluntary Foundation Model Master File, an independent
third party, or equivalent documentation. Independent third parties, such as academic medical
centers and non-profit organizations, could generate reusable evidence on foundation-model
performance, clinical utility and generalizability across representative clinical populations. However,
reusing foundation-model evidence should not remove the need for device-specific verification and
validation. The required level of clinical evidence, including whether a prospective clinical study is
needed at all, should depend on what verification can already establish, and on the intended use,
clinical claim, target population, workflow and risk profile of the final device.

Where the foundation model changes, reassessment and change-control procedures should apply.
The device manufacturer remains responsible for demonstrating the safety and effectiveness of the
finished device.

Collaboration between intended users and device manufacturers should be considered when
determining evidence requirements, with the goal of reducing unnecessary premarket and
postmarket burden while maintaining appropriate assurance of safety and effectiveness.

Summary of Recommendations:

1. Require intended-use-specific benchmarking using independent datasets that have not been
used for model development activities.
2. Independent third parties, such as academic medical centers, non-profit organizations,
cooperative study groups, or other qualified organizations, could conduct independent
assessments of clinical utility, generalizability, and real-world performance using
representative clinical populations. Such evidence could be leveraged across multiple
downstream devices that incorporate the same foundation-model version.
3. Identify and document the foundation model and version used in the device, including
available information regarding provenance, limitations, and performance.
4. Establish clear change-control expectations for foundation-model updates, including version
management, retesting triggers, and reassessment criteria.

Page 4 of 5
SolarisRTC, Silver Spring MD
 5. Permit reuse of foundation-model evidence for the same model version while maintaining
device-specific validation need and level, focused on the intended use and clinical claim.
6. Consider requiring collaboration between intended users and device manufacturers when
determining evidence requirements, with the goal of reducing unnecessary premarket and
postmarket burden while maintaining appropriate assurance of safety and effectiveness.

Thank you for considering these comments. I would welcome further discussion as CDRH develops
its approach.

Respectfully submitted,

Esther Abels
CEO, Founder SolarisRTC

Page 5 of 5
SolarisRTC, Silver Spring MD