Siemens Healthineers (Surya Kumar)

See attached file(s)

Surya Kumar, RAC

Siemens Healthineers,Regulatory Affairs

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

Docket: FDA-2026-N-7874
Submitted by: Siemens Healthineers
Function: Regulatory Affairs

Executive Summary

FDA's proactive initiative to engage stakeholders regarding an appropriate regulatory framework
for Generative AI (GenAI)-enabled medical devices is appreciable. The discussion paper
appropriately recognizes that traditional software validation paradigms may not adequately
address the unique characteristics of GenAI-enabled systems, including open-ended outputs,
non-deterministic behavior, continuous adaptation, foundation model dependencies, and
agentic capabilities. FDA’s continued development of a Total Product Lifecycle (TPLC) based
regulatory framework for GenAI-enabled devices is appreciated. However, I believe several
proposed concepts require further clarification to ensure regulatory predictability, consistency
across review divisions, global harmonization, and practical implementation by manufacturers.

The following comments are intended to support development of a risk-based, least
burdensome, scientifically robust, and internationally harmonized approach.

Summary of Key Comments

Section Page No. Topic Commenter’s Position

Two-Axis Risk Support concept; recommend additional risk
IV 10-Jun
Framework dimensions

ActionRequire objective and reproducible classification
IV 9-Jul Directing
criteria
Functions
CompetencyV 16-Oct Based Support concept with modifications
Evaluation
 Section Page No. Topic Commenter’s Position

Benchmarking Requires statistical rigor and predefined acceptance
V 13-18
Framework criteria

Clinical
V 15-17 Risk-based evidence matrix needed
Confirmation

V 15-18 Synthetic Data Appropriate as supplementary evidence only

Postmarket
VI 19-22 Should not replace adequate premarket evidence
Monitoring

PCCP and
VI 20-22 Additional GenAI-specific change categories needed
Modifications

Foundation
VII 22-23 Support concept with defined responsibilities
Model MAF

VII 23-24 Agentic AI Requires additional oversight considerations

Entire International Alignment with IMDRF and global frameworks
Overall
Document Harmonization recommended

Page references correspond to FDA discussion paper pagination.
 Detailed Comments
Comment 1

Two-Axis Risk Framework Requires Additional Risk Dimensions

FDA Section: IV. Considerations for Assessment of Risk
Pages: 6-10

FDA proposes a risk framework based upon:

1. Device activity

2. Consequences of relying upon incorrect output

Commenter's position

I support the proposed framework as a useful organizing principle. However, risk associated
with GenAI-enabled devices extends beyond device activity and consequence severity.

The proposed framework does not explicitly account for several key risk determinants that
significantly influence device safety and effectiveness.

Additional Risk Modifiers

Additional Factor Regulatory Relevance
Explainability Ability to understand rationale for output

Traceability Ability to trace outputs to source information

Human Reviewability User ability to independently verify output

Degree of Adaptation Static versus evolving model
Foundation Model Dependency Reliance on external model provider

Reversibility Reversible versus irreversible clinical impact

Error Detectability Ability to recognize incorrect outputs
Recommendation

FDA should consider incorporation of these factors as:

• Additional risk dimensions, or

• Formal risk modifiers applied to the proposed framework.

Comment 2

Action-Directing Functions Require Objective Regulatory Criteria

FDA Section: Action-Directing Functions
Pages: 7-8

Commenter's position

I agree that outputs that influence decision-making should be distinguished from outputs that
merely provide information.

However, the distinction between:

• non-directive,

• action-directing, and

• action-taking

functions may currently be interpreted subjectively.

Concern

Outputs such as:

• "Clinicians often increase dosage"

• "Clinicians in similar situations increase dosage"

• "Consider increasing dosage"

• "Increase dosage"
represent a spectrum rather than discrete categories. Without objective criteria, sponsors may
encounter inconsistent classification and regulatory expectations.

Recommendation

Characteristic Potential Indicator of Directiveness

Specificity General vs patient-specific

Personalization Population level vs individual

Recommended Action Informational vs explicit action

Availability of Alternatives Multiple options vs single recommendation

Human Confirmation Required Mandatory vs optional
 Comment 3

Competency-Based Assessment is Promising but Requires Further Definition

FDA Section: Competency-Based Approach
Pages: 10-18

Commenter’s Position

I support FDA's recognition that traditional software validation approaches may not fully
capture the performance characteristics of GenAI-enabled devices.

The proposed benchmarking plus clinical confirmation framework is a valuable innovation.

Concern

The discussion paper heavily references analogies to physician credentialing and competency
assessment.

While conceptually useful, AI systems fundamentally differ from human clinicians in terms of:

• accountability,

• professional liability,

• contextual awareness,

• ethical responsibilities.

Recommendation

FDA should position competency-based assessments as:

"A complementary device evaluation framework" rather than direct analogue to human
clinician credentialing.
 Comment 4

Benchmarking Framework Requires Statistical Expectations

FDA Section: Device Benchmarking
Pages: 12-15

FDA proposes benchmarking across:

• Safety

• Clinical Proficiency

• Generalizability

• Agentic Capabilities

Commenter’s Position

The proposed domains cover important aspects of GenAI-enabled device performance.
However, the discussion paper does not specify expectations for:

Characteristic Potential Indicator of Directiveness

Specificity General vs patient-specific

Personalization Population level vs individual

Recommended Action Informational vs explicit action

Availability of Alternatives Multiple options vs single recommendation

Human Confirmation
Mandatory vs optional
Required

Recommendation

FDA should develop benchmark qualification principles comparable to established performance
testing expectations used for AI-enabled SaMD.
 Comment 5

Synthetic Data Should Remain Supplementary

FDA Section: Clinical Confirmation
Pages: 15-17

Commenter’s Position

Synthetic data may provide important benefits including:

• evaluation of rare events,

• subgroup representation,

• privacy protection,

• scalability.

However, synthetic datasets may also inherit limitations of source models and fail to fully
represent real-world clinical variability.

Recommendation

For moderate and high-risk functions:

Synthetic data should supplement but generally not replace real-world clinical evidence.

Additional validation should demonstrate that synthetic datasets adequately represent
intended-use populations.
 Comment 6

Clinical Confirmation Requires Defined Risk-Based Expectations

FDA Section: Clinical Confirmation Approaches
Pages: 15-17

Commenter’s Position

The range of proposed confirmation methods reflects an appropriate risk-based philosophy.

However, greater predictability is needed regarding evidence expectations.

Recommended Framework

Risk Category Potential Evidence Approach

Low Risk Information
Retrospective validation
Functions
Moderate Risk Clinical
Retrospective + Shadow Deployment
Support

High Risk Decision Support Prospective Clinical Evaluation

Autonomous Clinical Action Controlled Clinical Study
Recommendation

FDA should publish risk-based evidence expectations to improve consistency across submissions
and reviews.
 Comment 7

Postmarket Monitoring Should Not Substitute for Premarket Evidence

FDA Section: Postmarket Monitoring
Pages: 19-22

Commenter’s Position

I support robust postmarket monitoring for GenAI-enabled devices.

However, reliance on postmarket evidence should not replace adequate premarket
demonstration of safety and effectiveness, particularly for:

• autonomous systems,

• therapeutic recommendations,

• high-severity clinical decisions.

Recommendation

Reduced premarket evidence should be limited to low risk functions and accompanied by:

• predefined monitoring plans,

• performance thresholds,

• corrective action triggers,

• manufacturer accountability.
 Comment 8

Change Management and PCCPs Require GenAI-Specific Clarification

FDA Section: Postmarket Modifications
Pages: 21-22

Commenter’s Position

Current PCCP concepts may not fully address GenAI-specific modifications.

GenAI systems may change through:

Modification Type
• Prompt updates
• RAG knowledge-base updates
• Foundation model changes
• Safety guardrail modifications
• Orchestration changes
• Workflow integration updates

Recommendation

FDA should establish a GenAI-specific modification taxonomy that clearly identifies:

• QMS-managed changes

• PCCP-manageable changes
 Comment 9

Foundation Model Master Files Are a Valuable Concept

FDA Section: Foundation Model MAF
Pages: 22-23

Commenter’s Position

I support exploration of Foundation Model Master Files.

Manufacturers increasingly depend upon third-party foundation models while having limited
visibility into:

• training data,

• architecture,

• updates,

• safety controls.

Recommendation

FDA should define minimum expectations for:

Area Expected Information

Version Traceability Model release documentation

Change Notification Update communication process

Performance Characterization Evaluation summaries

Known Limitations Documented failure modes

Safety Controls Guardrails and restrictions

Auditability Availability of logs and records
 Comment 10

Additional Considerations for Agentic AI Systems

FDA Section: Agentic AI Systems
Pages: 23-24

Commenter’s Position

Agentic AI systems introduce risks beyond those of conventional GenAI-enabled devices.

Additional concerns include:

• multi-step autonomous actions,

• tool delegation,

• cascading errors,

• emergent behaviors,

• reduced human oversight opportunities.

Recommendation

Additional evaluation criteria should address:

Assessment Area

• Planning reliability
• Tool-use integrity
• Human override effectiveness
• Failure recovery
• Escalation mechanisms
• Prompt injection resilience
 Comment 11

International Harmonization Should Be Explicitly Addressed

Applicable Throughout Document

Commenter’s Position

Given the global nature of AI-enabled medical device development, regulatory convergence
remains critical.

I encourage FDA to align future regulatory expectations with:

• IMDRF SaMD Frameworks

• FDA PCCP Guidance

• EU MDR

• EU AI Act

• Health Canada AI initiatives

• MHRA AI regulatory roadmap

Recommendation

FDA should continue collaboration through IMDRF and related international initiatives to avoid
duplication of evidence requirements and support global patient access.
Conclusion

I support FDA's efforts to establish a future-ready regulatory framework for GenAI-enabled
medical devices. The proposed concepts of risk-based oversight, competency-based evaluation,
lifecycle monitoring, and foundation model governance provide a strong foundation for future
policy development.

To facilitate consistent implementation, I recommend additional clarification regarding:

1. Risk classification methodology.

2. Action-directing function criteria.

3. Statistical expectations for benchmarking.

4. Clinical confirmation requirements.

5. Appropriate use of synthetic data.

6. Postmarket monitoring expectations.

7. GenAI-specific change management.

8. Foundation model governance.

9. Agentic AI oversight.

10. International harmonization.

I appreciate FDA's engagement with stakeholders and welcome continued collaboration to
develop scientifically sound, predictable, least burdensome, and globally harmonized regulatory
approaches for GenAI-enabled medical devices.

Prepared by:
Surya Kumar RAC
Regulatory Affairs
Siemens Healthineers
Response to FDA-2026-N-7874 Discussion Paper on Generative AI-Enabled Medical Devices.