Japan Digital Health Alliance (JaDHA)

Please See attached the file.

Submitted to: Dockets Management Staff, U.S. Food and Drug Administration
Docket No.: FDA-2026-N-7874
Title: Comments on "Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback" (August 2026)
Submitted by: Regulatory Affairs Working Group, Committee on Regulatory and Reimbursement
Frameworks for SaMD, Japan Digital Health Alliance (JaDHA)
Date: October 16, 2026

Introduction
The Japan Digital Health Alliance (JaDHA) appreciates the opportunity to comment on CDRH's
discussion paper, "Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback". JaDHA is an industry association of companies
developing and commercializing digital health technologies in Japan, including Software as a Medical
Device (SaMD).
On September 14, 2026, Japan's Pharmaceuticals and Medical Devices Agency (PMDA) published
draft Review Points for Medical Devices Utilizing Generative AI to gather a wide range of opinions.
Our comments draw on that document and on the work of the PMDA Science Board. We hope this
helps CDRH identify where the approaches of the two regulators already converge. Each comment
below states our recommendation, identifies the relevant discussion questions in Appendix B of the
paper, and explains our rationale.1,2
General Comments
JaDHA supports CDRH's intent to pursue a nimble regulatory approach for GenAI-enabled devices
based on three foundations:
• least burdensome principles,
• a risk-based approach that considers the product's intended use and technological
characteristics, and
• a total product life cycle (TPLC) approach.
We also agree with CDRH's statement that FDA does not regulate GenAI as such, but regulates
products that meet the definition of a device.
These directions are consistent with PMDA's draft Review Points:1
• PMDA states that the Total Life Cycle Approach becomes more important for GenAI-enabled
medical devices.1
• PMDA calls for actively implementing a "rebalancing" of efficacy and safety evaluation before
and after marketing.1
• PMDA states that a device using a GenAI model is likely to follow conventional evaluation
policies if its use is limited and its procedure for generating outputs is fixed.1

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
Several general principles in the paper are already expected of medical devices more broadly. CDRH
itself notes that prespecifying test methods and acceptance criteria is consistent with expectations
for non-clinical performance testing across devices. Retrospective evaluation on real patient inputs
is another example. These principles apply equally to SaMD in general, to hardware devices, and to
in vitro diagnostics.
We therefore recommend that CDRH clearly distinguish requirements specific to GenAI from those
that existing frameworks already address. This would make regulation more predictable and support
international harmonization, including through IMDRF.

1. Action-directing and action-taking functions involving medication dosing should be explicitly
placed high on the consequences axis
Relevant Discussion Questions: 1, 2, 18
Recommendation: JaDHA supports the two-axis risk framework, which organizes risk around device
activity and the consequence of relying on an incorrect output. We recommend the following:
• Placement on the consequences axis: Action-directing and action-taking functions that
specify or change medication doses (e.g., insulin, anticoagulants) should be explicitly
identified as positioned high on the consequences axis.
• Reliance on postmarket monitoring: These functions should be identified as a category for
which it would not be appropriate to accept greater premarket uncertainty through greater
reliance on postmarket monitoring (Question 18).
Rationale:
• The paper itself contrasts an incorrect recommendation about an over-the-counter remedy
with an incorrect recommendation about adjusting an insulin dose. Whether a dose is
appropriate depends heavily on the target disease, the drug, and the patient's condition.
• A physician sets a dose after clinical examination and testing. The inputs are fundamentally
different from those available when a patient asks a general-purpose model the same
question.
• Dosing errors may be irreversible. We therefore support adding the "reversibility of a
resulting action," which Question 1 lists as an example, as an explicit dimension of the
framework.
• PMDA's draft Review Points state that premarket evaluation under specified conditions may
be acceptable provided a post-marketing system actively identifies and responds to incidents.
For dosing functions, we believe such a system alone is not enough to offset premarket
uncertainty.1
• We agree with CDRH that a patient-facing function does not become any less directive
because it includes a "talk to your doctor" statement. PMDA similarly recommends

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
 implementing functions that reject requests outside the intended use (e.g., guardrails),
rather than relying on cautionary statements. For patient users, PMDA also calls for safety
design based on foolproof and failsafe concepts.1
2. The generalist/specialist distinction should be decided based on disease characteristics, not on
the use of GenAI
Relevant Discussion Questions: 4, 14
Recommendation: Whether a function should be restricted to specialist HCPs should be decided on
the basis of the disease and the intended use. The fact that the device uses GenAI should not, by
itself, be the deciding factor. We agree with CDRH that a function helping a generalist HCP approach
a question typically handled by a specialist may meaningfully extend access to specialty knowledge.
Rationale:
• In Japan, some approved SaMD products carry approval conditions limiting use to specialists.
In our experience, these restrictions reflect the characteristics of the disease more than the
fact that the product is software.
• In regions with too few specialists, functions that support generalist HCPs can greatly
improve access to care. Consistent with least burdensome principles, uniformly strict
requirements should be avoided.
• The PMDA Science Board has noted that narrowing the conditions of use too far during
review can itself become an obstacle to clinical use of AI. It recommends setting such
conditions from the perspective of risk-benefit balance.2
• PMDA's draft Review Points describe a stepwise policy: obtain initial approval for a narrowly
defined intended use, then gradually expand the claims. A similar approach could allow use
by generalist HCPs to be expanded step by step.1
3. We support the general principles for device benchmarking but request flexibility on adjudicator
independence and clarity on re-benchmarking
Relevant Discussion Questions: 7, 9, 10, 22
Recommendation: JaDHA supports the general principles for benchmarking, including:
• prespecifying test methods and acceptance criteria, and
• grounding scoring rubrics in established clinical guidelines.
We also request the following:
• Adjudicator independence: The paper calls for expert adjudicators who are structurally
independent from the sponsor and from the third-party model developer. Where full
structural independence is impractical, we ask that alternative safeguards also be accepted.
Examples include blinding, disclosure of conflicts of interest, and restricting developers'
access to test data.
• Re-benchmarking: The paper states that the initial device benchmarking may serve as a

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
 reference against which later changes are evaluated. We ask CDRH to give concrete examples
of how re-benchmarking results would be assessed. One example would be demonstrating
non-inferiority or equivalence to premarket performance using prespecified, clinically
justified margins.
• Training data integrity: In addition to the limitations of benchmarking assets, we
recommend that the provenance and integrity of training and fine-tuning data be included
as an evaluation consideration.
Rationale:
• Securing fully independent expert adjudicators can be difficult in practice.
• PMDA's draft Review Points describe three ways to judge whether outputs are appropriate:
rule-based evaluation, assessment by physicians or other experts, and LLM-as-a-Judge.
Methods should be selected or combined according to the task. PMDA also presents
"property-based testing" as useful where no single ground truth exists.1
• The PMDA Science Board identifies physically restricting developers' access to test data as a
lower-risk practice. It also warns that repeatedly reusing the same test data can introduce
bias into performance estimates. For evaluating performance after retraining, the Science
Board describes non-inferiority or equivalence to premarket performance, with endpoints
and margins set on clinical grounds.2
• A recent review in Nature points to a study showing that data poisoning can be done with
very small amounts of data. In that study, altering as little as 0.001% of a training corpus
(approximately 1,500 pages) was enough to make LLMs produce clinically dangerous outputs.
One example was recommending beta blockers for asthma, a condition in which they are
contraindicated.5,6
• The same review identifies data version control (e.g., checksums) and systematic data
curation as essential steps before training medical AI models. It also notes that closed-weight
models accessed through APIs are more difficult to audit externally.5
• Because the training datasets of third-party foundation models may not be disclosed to
device developers, benchmark performance alone may not reveal this kind of vulnerability.
Such vulnerabilities add to the data contamination, saturation, and limited real-world
representativeness that the paper already recognizes for public benchmarking assets.
4. CDRH should clarify what counts as sufficient clinical confirmation and recognize the limits of
shadow deployment for therapeutic functions
Relevant Discussion Questions: 11, 12, 13
Recommendation: JaDHA supports CDRH's position that clinical confirmation might not require a
prospective clinical study in every case. We request the following:
• Sufficiency criteria: CDRH should provide criteria for judging when the type and rigor of

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
 evidence are consistent with the device's intended use and proportionate to its risk profile.
• Therapeutic functions: For therapeutic functions such as dose recommendations, shadow
deployment alone has limited value. Simulation tools, including virtual patient avatars and
digital twins, should be accepted as complementary options, provided that their credibility
for the specific context of use is demonstrated.
Rationale:
• Retrospective evaluation on real patient inputs is already established in Japan. MHLW
Notification PSEHB/MDED No. 0929-1 (September 29, 2021) covers performance evaluation
of diagnostic devices using existing medical image data without additional invasiveness or
intervention. This creates an opportunity for international harmonization.3
• The paper defines shadow deployment as follows: device outputs are not shown to clinicians
or patients and do not affect care, and they are compared against decisions actually made
and, where appropriate, downstream outcomes. For diagnostic functions, this comparison
can be informative. For therapeutic functions, however, the downstream outcomes reflect
the treatment that was actually given, not the device's recommendation. The clinical
consequences of following the device's output therefore cannot be observed directly.
• The paper already recognizes simulation tools, including virtual patient avatars, as possible
benchmarking methods. The PMDA Science Board, however, considers that simulationgenerated data are not yet accepted as test data unless their validity for the evaluation
purpose is scientifically demonstrated. It discusses this with reference to the context of use
and validation concepts of ASME V&V 40. We suggest that CDRH set out comparable
conditions for synthetic data and simulation-based evidence. FDA's guidance Assessing the
Credibility of Computational Modeling and Simulation in Medical Device Submissions may
provide a useful basis.2
5. Independent third parties should be used, with international mutual recognition in view
Relevant Discussion Question: 16
Recommendation: JaDHA supports exploring whether third-party frameworks analogous to the ASCA
and MDDT programs could support the assessment of GenAI-enabled devices. We also encourage
CDRH to consider future mutual use of such frameworks with other jurisdictions.
Rationale:
• ASCA works in a similar way to the IECEE CB Scheme.
• In Japan, when testing is performed by laboratories accredited under ISO/IEC 17025, full test
reports do not have to be submitted. This has meaningfully reduced the burden on applicants.
6. We request clarification of triggering events for reassessment and of changes manageable within
the quality management system
Relevant Discussion Questions: 19, 21, 22

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
Recommendation: JaDHA recommends the following:
• Cadence and triggers: In addition to triggering events such as changes to the underlying
model, we ask CDRH to consider a defined cadence of periodic re-benchmarking as one
option, similar to a renewal-type reassessment.
• Changes managed internally: We ask CDRH to illustrate, with concrete examples (e.g., in a
Q&A document), categories of change that could be appropriately managed within a
sponsor's quality management system.
• Adverse event reporting: As the foundation of shared ecosystem responsibility, we believe it
is important to establish a reliable pathway for reporting adverse events to regulators. Care
should be taken that manufacturer accountability is not diffused in the process.
Rationale:
• Some jurisdictions, such as the EU under the CE marking framework and China under its
registration renewal system, renew marketing authorizations at fixed intervals. This is
separate from the procedures required when a change is made. Under these mechanisms,
regulators periodically check whether changes and product updates made during the
authorization period were based on appropriate risk management. We believe this periodic
renewal concept may serve as one possible reference when considering how GenAI-enabled
devices should be continuously reassessed.
• PMDA's draft Review Points expect applicants to actively collect information on postmarketing performance, adverse events, and hallucinations, and to report periodically.
Similar expectations in the U.S. would support a consistent global approach to postmarket
monitoring.1
• For models that learn after deployment (model-evolution changes), errors in training data
can directly affect outputs. This makes performance degradation monitoring particularly
important.
7. Drawing on Japan's experience with IDATEN, JaDHA supports CDRH's approach to applying PCCPs
Relevant Discussion Questions: 23, 24
Recommendation: Drawing on Japan's experience with the Improvement Design within Approval for
Timely Evaluation and Notice (IDATEN) system, JaDHA supports CDRH's approach of using PCCPs to
address modifications to GenAI-enabled devices. Where the nature or scope of future modifications
cannot be fully prespecified, we further recommend a mechanism through which sponsors and CDRH
decide, through consultation, whether a PCCP applies. For unplanned changes arising from updates
to a third-party foundation model, the baseline expectations could include:
• identifying and locking the model version or inference environment where feasible,
• limiting the device's operating scope,
• re-benchmarking when the underlying model changes, and

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
 • defining contractual arrangements, such as Service Level Agreements, with the model
developer.
Rationale:
• Japan's IDATEN system addresses the plasticity of machine learning. Whether it applies is in
practice determined through consultation with the regulator.2
• IDATEN has been used only to a limited extent in Japan so far. In our members' experience,
even among imaging diagnostic systems, some product designs fit IDATEN well and others
do not. Many members currently find it difficult to use.
• PMDA itself recognizes that IDATEN is limited in its ability to handle products whose
characteristics change in real time. Japan and the U.S. share this challenge. We believe there
is much to gain from international dialogue with CDRH, which has more experience in
operating PCCPs.2
• PMDA's draft Review Points explain that the areas an applicant directly controls differ by
development model:1
o Using a proprietary model via a cloud API differs from operating an open-weight
model in the applicant's own environment.1
o Updates to a foundation model provided via a cloud API can irreversibly change
outputs.1
o Contracts and Service Level Agreements should therefore be specified according to
the development model.1
• There have also been cases where a change made by an external platform provider, such as
a renamed service, caused the product to no longer match the description in its approval.1
8. JaDHA supports voluntary Foundation Model MAFs
Relevant Discussion Question: 25
Recommendation: JaDHA supports the feasibility of voluntary Foundation Model MAFs that leverage
the existing Device Master File program. To make them useful for premarket review, we recommend
the following:
• Content: The minimum content of a Foundation Model MAF should be clearly defined,
focusing on information needed to identify the specific model and version.
• Update notifications: The update notification commitments that the paper lists as possible
content should be a core element. The program should also make clear how an update to a
referenced MAF affects the marketing authorizations of devices that rely on it.
• Incentives: The benefits to participating foundation model developers, such as more
consistent and efficient premarket review, should be communicated clearly.
Rationale:
• Japan has a comparable Master File (MF) system under Article 80-6 of the Pharmaceuticals

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
 and Medical Devices Act. Since 2005, it has also been available for raw materials used in
medical devices.4
• MHLW Notification PSEHB/MDED No. 0530-1 (May 30, 2019) changed the system in two
ways:4
o The predecessor system covered only polyvinyl chloride, polyethylene, and
polypropylene. The notification extended the registrable scope to other device
materials where registration is appropriate.4
o The notification specified that registered content consists of information that
identifies the material.4
• The notification also requires that, when a material is re-registered under the current MF
system and receives a new MF number, devices using that material must update their
approvals through a minor change notification. This experience shows that changes to
master file information directly affect the regulatory status of the devices that reference it.
For foundation models, which may be updated frequently, the link between MAF updates
and device authorizations should be designed in advance.4
• In our members' experience, the Japanese MF system for device materials has not been
widely used. To be effective, a voluntary program needs clear incentives for participation.
• Timely notification of model updates is also a prerequisite for manufacturers to detect,
evaluate, and respond to third-party developer-initiated changes (Question 24).
9. CDRH should provide further examples of agentic AI systems that meet the device definition
Relevant Discussion Question: 26
Recommendation: The paper notes that an agentic AI system may meet the device definition, for
example when its action sequences result in control of another medical device. We request additional
concrete examples, particularly for agentic AI systems that run on non-device platforms such as
electronic health records but initiate or modify the operation of medical devices.
Rationale:
• In cases such as an EHR-based agent switching a medical device on or off, where the
boundary of regulatory oversight lies is currently unclear.
• PMDA's draft Review Points describe autonomy as a continuum rather than a binary property,
similar to the levels of automation for autonomous vehicles. PMDA treats the level of
autonomy as a design choice. PMDA also calls for:1
o human confirmation, stoppage, and intervention mechanisms (e.g., human-in-theloop), and1
o follow-up and auditing of executed action sequences.1
These are consistent with the human-oversight checkpoints before irreversible or high-consequence
actions described in benchmarking element A.1. We suggest that acceptance criteria and oversight

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.
for agentic GenAI-enabled devices be graded according to the level of autonomy.

Conclusion
JaDHA appreciates CDRH's collaborative and transparent approach to this rapidly evolving field. The
FDA discussion paper and PMDA's draft Review Points share many key principles, including:1
• the TPLC approach,1
• risk-proportionate evaluation, and1
• recognition of the limitations of benchmarks.1
We would welcome continued dialogue between CDRH, PMDA, and other regulators, including
through IMDRF, and we are ready to share practical experience from the Japanese regulatory
environment.

References
1. [1] PMDA Center for Product Evaluation. Review Points for Medical Devices Utilizing
Generative AI (Draft). September 14, 2026 (provisional English translation).
2. [2] PMDA Science Board. Report on AI-based Software as a Medical Device (SaMD). August
28, 2023.
3. [3] MHLW. Notification PSEHB/MDED No. 0929-1, September 29, 2021.
4. [4] MHLW. Notification PSEHB/MDED No. 0530-1, "Handling of the Master File for Raw
Materials of Medical Devices," May 30, 2019.
5. [5] Clusmann J, et al. Safety and security of large language models in healthcare. Nature.
2026;656:577–589. doi:10.1038/s41586-026-10687-1.
6. [6] Alber DA, et al. Medical large language models are vulnerable to data-poisoning attacks.
Nat Med. 2025;31:618–626.

Copyright (C) 2026 The Japan Research Institute, Limited. All Rights Reserved.