Medical AI Validation Network MAIVAN.ai (Thomas Morscher)

Competency-based Evaluation · Independent Clinical Societies · Clinician Adjudication · Outcome-linked Postmarket Monitoring

Generative and agentic AI systems cannot always be evaluated adequately through static measurements of model accuracy alone. Medical practice itself is organised around bounded competencies, defined scopes of practice, specialist standards, supervised clinical application, outcome measurement and continuing assessment. A comparable - but appropriately adapted - approach to medical AI could provide a scientifically rigorous and innovation-compatible framework for assuring safety and effectiveness.

We recommend extending FDA’s proposed direction through three complementary principles:
Specialty-defined competency frameworks. Independent medical societies and clinically governed expert organisations should be able to define and maintain clinically meaningful competency domains, evaluation criteria and reference standards for specific fields of medicine.

Independent clinical confirmation. Benchmark performance should be complemented by structured adjudication by qualified clinicians who are independent of the device manufacturer and, where appropriate, independent of the developer of the underlying foundation model.

Outcome-linked postmarket monitoring. Postmarket monitoring should assess not only whether an AI system continues to reproduce its benchmark performance, but also whether its deployment is associated with appropriate clinical processes and meaningful patient outcomes in real-world care.

Orthopaedic trauma provides a concrete regulatory-science example. At one end of the patient journey, the AO/OTA Fracture and Dislocation Classification provides a structured clinical language for characterising fractures. At the other, the ICHOM Major Injury Set provides a standardised framework for measuring outcomes that matter to patients after major injury.
Competency-based evaluation should not end at the correctness of an AI output. Where feasible, the regulatory evidence architecture should connect a defined clinical competency to clinician-confirmed decisions and, over time, to standardised real-world patient outcomes.

This architecture allows three distinct questions to be addressed:
Can the device perform the defined clinical competency?
Does it continue to perform that competency safely in real-world practice?
Does use of the device within the human clinical workflow support appropriate patient outcomes?

International relevance. Generative-AI-enabled medical devices, foundation models, clinical standards and postmarket evidence infrastructures routinely cross national borders; FDA-cleared devices may be deployed globally; and both AO/OTA classification and ICHOM outcome standards are international in scope. International clinical and outcomes organisations can therefore contribute directly relevant evidence, standards and implementation experience to FDA’s regulatory-science questions.
Conflict-of-interest and illustrative implementation disclosure. MAIVAN.ai PBC / Medical AI Validation Network is developing ORTHO-X and OrthoSkills and therefore has a commercial and intellectual-property interest in related clinical-AI infrastructure. The core recommendations in this submission are intended to stand independently of any particular commercial implementation.

Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Feedback on the Discussion Paper ​
Considerations for the Regulation of Generative AI-Enabled Medical Devices​

Response to FDA Docket FDA-2026-N-7874
Submission date: October 19, 2026

Competency-based evaluation · independent clinical societies ·
clinician adjudication · outcome-linked postmarket monitoring

submitted by

Medical AI Validation Network (MAIVAN.ai)

Contents
Executive Summary..................................................................................................................................... 2
I. Competency-Based Evaluation Is an Appropriate Framework..................................................................3
II. Professional Societies Can Help Define the Clinical Meaning of Competence........................................4
III. AO/OTA Classification as an Orthopaedic Competency Example.......................................................... 5
IV. Clinical Confirmation Should Include Structured Clinician Adjudication..................................................6
V. Multimodal Vision-Language Models Require Clinically Integrated Evaluation....................................... 6
VI. Patient Outcomes Should Become a Complementary Postmarket Evidence Layer.............................. 7
VII. ICHOM as an Outcome-Standard Layer................................................................................................8
VIII. A Longitudinal Clinical Evidence Thread.............................................................................................. 9
IX. Example: AO/OTA Classification to ICHOM Outcome Measurement...................................................10
X. A Federated Postmarket Evidence Infrastructure.................................................................................. 10
XI. Re-Benchmarking Should Be Triggered by Clinically Meaningful Change........................................... 11
XII. Agentic AI Should Be Evaluated at Skill and Workflow Level.............................................................. 12
XIII. Outcome-Linked Monitoring and Risk-Proportionate Regulation........................................................12
XIV. Relationship to TEMPO and Real-World Evidence.............................................................................12

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 1
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

XV. Five Recommendations to FDA...........................................................................................................13
Conclusion..................................................................................................................................................14
Signatory.................................................................................................................................................... 15
Appendix A — Illustrative Orthopaedic Regulatory-Science Architecture.................................................. 16
Appendix B — Regulatory Evidence Matrix............................................................................................... 16
Annex C — Illustrative ORTHO-X / OrthoSkills Implementation and Conflict-of-Interest Disclosure......... 18
References and Source Links.................................................................................................................... 18

Executive Summary
We welcome the U.S. Food and Drug Administration (FDA) Center for Devices and Radiological Health
discussion paper on generative AI-enabled medical devices and its exploration of competency-based
evaluation, clinical confirmation and risk-proportionate postmarket monitoring.

Generative and agentic AI systems cannot always be evaluated adequately through static measurements
of model accuracy alone. Medical practice itself is organised around bounded competencies, defined
scopes of practice, specialist standards, supervised clinical application, outcome measurement and
continuing assessment. A comparable - but appropriately adapted - approach to medical AI could provide
a scientifically rigorous and innovation-compatible framework for assuring safety and effectiveness.​

We recommend extending FDA's proposed direction through three complementary principles:
1.​ Specialty-defined competency frameworks. Independent medical societies and clinically governed
expert organisations should be able to define and maintain clinically meaningful competency
domains, evaluation criteria and reference standards for specific fields of medicine.​

2.​ Independent clinical confirmation. Benchmark performance should be complemented by structured
adjudication by qualified clinicians who are independent of the device manufacturer and, where
appropriate, independent of the developer of the underlying foundation model.​

3.​ Outcome-linked postmarket monitoring. Postmarket monitoring should assess not only whether an AI
system continues to reproduce its benchmark performance, but also whether its deployment is

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 2
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

associated with appropriate clinical processes and meaningful patient outcomes in real-world care.​

Orthopaedic trauma provides a concrete regulatory-science example. At one end of the patient journey,
the AO/OTA Fracture and Dislocation Classification provides a structured clinical language for
characterising fractures. At the other, the ICHOM Major Injury Set provides a standardised framework for
measuring outcomes that matter to patients after major injury.

Competency-based evaluation should not end at the correctness of an AI output. Where
feasible, the regulatory evidence architecture should connect a defined clinical competency to
clinician-confirmed decisions and, over time, to standardised real-world patient outcomes.
This architecture allows three distinct questions to be addressed:

●​ Can the device perform the defined clinical competency?
●​ Does it continue to perform that competency safely in real-world practice?
●​ Does use of the device within the human clinical workflow support appropriate patient outcomes?

International relevance. Generative-AI-enabled medical devices, foundation models, clinical standards
and postmarket evidence infrastructures routinely cross national borders; FDA-cleared devices may be
deployed globally; and both AO/OTA classification and ICHOM outcome standards are international in
scope. International clinical and outcomes organisations can therefore contribute directly relevant
evidence, standards and implementation experience to FDA's regulatory-science questions.
Conflict-of-interest and illustrative implementation disclosure. MAIVAN.ai PBC / Medical AI Validation
Network is developing ORTHO-X and OrthoSkills and therefore has a commercial and
intellectual-property interest in related clinical-AI infrastructure. The core recommendations in this
submission are intended to stand independently of any particular commercial implementation.

I. Competency-Based Evaluation Is an Appropriate Framework
We support FDA's exploration of a competency-based framework combining non-clinical benchmarking
and clinical confirmation. This is particularly appropriate for medical GenAI because a foundation model
may possess broad capabilities, while a regulated medical function should have a clearly bounded
intended use and clinical scope.

foundation-model capability → device function → clinical competency → clinical workflow →
outcome

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 3
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

A medical device should not be regarded as clinically competent merely because its underlying model
performs strongly on general medical benchmarks. The relevant question is whether the deployed device
function reliably performs the specific competency for which it is intended.

In orthopaedic trauma, separate competencies might include:

●​ anatomical routing of an injury;
●​ assessment of whether diagnostic imaging is sufficient for the requested task;
●​ application of a recognised fracture classification;
●​ communication and propagation of uncertainty;
●​ identification of plausible differential classifications;
●​ recognition of when additional imaging or expert review is required;
●​ evidence retrieval and treatment mapping within a bounded scope;
●​ structured follow-up and outcome measurement;
●​ detection of clinically relevant deviations requiring review.

Each competency can have defined inputs, outputs, users, clinical context, exclusions, performance
criteria, uncertainty requirements, escalation rules, human-confirmation requirements and postmarket
monitoring thresholds.

Recommendation 1. FDA should encourage the development of clinically bounded competency
specifications for GenAI-enabled medical-device functions, with evidence requirements scaled to the risk
and autonomy of each function.

II. Professional Societies Can Help Define the Clinical Meaning of
Competence
Public AI benchmarks face well-recognised limitations, including contamination, saturation and
insufficient representation of real clinical complexity. Independent professional societies and clinically
governed organisations can provide the clinical reference layer against which product-specific evaluation
occurs.

Potential contributions include:

●​ defining clinically meaningful competencies and clinically material errors;
●​ identifying authoritative clinical reference standards;
●​ constructing or governing sequestered test datasets;

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 4
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

●​ defining representative case distributions and edge cases;
●​ establishing scoring rubrics, uncertainty requirements and deferral behaviour;
●​ convening specialist adjudication panels;
●​ defining appropriate outcome measures;
●​ maintaining benchmarks as standards of care evolve.

Recommendation 2. Where a professional society or independent organisation participates in
competency evaluation, FDA should consider governance independence, financial conflict disclosure,
transparent methodology, separation of standard-setting from product promotion, expert qualification,
benchmark security, dataset provenance, version control, reproducibility and conflict-of-interest
management. Manufacturer accountability must remain with the manufacturer.

III. AO/OTA Classification as an Orthopaedic Competency
Example
The AO/OTA Fracture and Dislocation Classification demonstrates how a specialty-standard clinical
ontology can support competency evaluation. It provides a structured framework for describing fractures
according to anatomy and morphology and is widely used in clinical communication, research and
registry contexts.

An AI competency operating within such a system can be evaluated across several dimensions:

Task fidelity. Can the device produce the correct classification at the clinically appropriate level of
specificity?

Uncertainty calibration. Does the device recognise when available imaging does not support a sufficiently
confident classification?

Differential reasoning. When several classifications are plausible, does it identify clinically relevant
alternatives rather than expressing unjustified certainty?

Information sufficiency. Can it recognise when additional views, CT or other information would materially
affect classification?

Boundary adherence. Does the classification function stay within its defined scope rather than silently
converting classification into autonomous treatment advice?

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 5
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Clinical deferral. Does it escalate uncertain or consequential cases for specialist confirmation?
​
Illustrative implementation examples are provided separately in Annex C so that the substantive
recommendation remains technology- and vendor-neutral.
Illustrative reference: ORTHO-X OrthoSkills - AO/OTA Classification Skill

IV. Clinical Confirmation Should Include Structured Clinician
Adjudication
Even comprehensive benchmarking may not establish how a GenAI-enabled device behaves in
representative clinical environments. Structured specialist adjudication is particularly valuable where the
task admits several defensible answers, interobserver variability exists, expertise materially affects
interpretation, or the device is intended to augment rather than replace clinicians.

AI alone | clinician alone | clinician + AI
For many decision-support systems, human-AI team performance may be the most clinically relevant
endpoint because that is the intended deployment configuration.

Recommendation 3. For clinically consequential GenAI functions, FDA should encourage prespecified
adjudication protocols defining required expertise, independence, number of adjudicators, consensus
procedures, handling of legitimate disagreement, blinding where feasible, criteria for uncertainty and
deferral, subgroup representation and the intended comparator.

V. Multimodal Vision-Language Models Require Clinically
Integrated Evaluation
Clinical reasoning rarely occurs in a single modality. In orthopaedic trauma, an intended-use system may
integrate radiographs, CT, clinical history, physical examination, operative information, laboratory data,
longitudinal follow-up and patient-reported outcomes.

A multimodal vision-language model should therefore not be evaluated solely as an image classifier
where its intended function combines imaging and language. Multimodality also creates an
error-propagation risk:

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 6
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

inadequate imaging → erroneous classification → inappropriate downstream reasoning →
misleading treatment support
Competency evaluation should test not only final-answer accuracy, but whether the system detects
missing information, propagates uncertainty, identifies contradictory modalities, maintains provenance,
preserves boundaries between competencies and escalates appropriately.

Technical consideration — multimodal error propagation. For multimodal devices, competency evaluation
should include cross-function error propagation and test whether uncertainty identified upstream remains
visible downstream.

VI. Patient Outcomes Should Become a Complementary
Postmarket Evidence Layer
We support periodic re-benchmarking, sample-based clinician review and monitoring for performance
degradation. We recommend an additional complementary layer where clinically and methodologically
appropriate: outcome-linked postmarket monitoring.

A tiered timing structure is essential because patient-centred outcomes and model-safety signals occur
on different timescales. We propose: (1) near-term operational and safety signals, including device/model
version, information sufficiency, classification performance, confidence calibration, clinician
acceptance/override, escalation, workflow deviations and early complications; (2) intermediate
patient-outcome signals, including the relevant ICHOM Major Injury measures at approximately 6
months; and (3) confirmatory patient-centred outcomes at approximately 1 year. The exact device,
foundation-model and workflow version should be recorded at the index clinical event so that later
outcomes remain retrospectively attributable to the correct technical configuration.

Evidence layer Question
Benchmarking Does the device still perform the defined competency?

Clinician review Are its real-world outputs clinically appropriate?

Drift monitoring Has system behaviour materially changed?

Outcome-linked monitoring What happens to patients in care pathways in which the
system is used?

Patient outcomes should not automatically be attributed causally to an AI system. Outcomes depend on
injury severity, patient factors, clinical decisions, treatment quality, rehabilitation, health-system

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 7
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

performance and other confounders. Nevertheless, standardised outcomes can provide valuable sentinel
evidence and generate hypotheses for investigation.

Where the regulatory or clinical question is whether an AI-enabled care pathway actually improves
outcomes, observational postmarket surveillance should be complemented by prospective comparative
evaluation. Where feasible and appropriate, randomized controlled trials should use standardized
patient-centred outcomes — such as the relevant ICHOM Major Injury measures — as prespecified
clinical-effectiveness endpoints. This separates rapid safety/performance surveillance from the stronger
causal question of whether AI-supported care produces better outcomes that matter to patients.

Recommendation 4. Add tiered outcome-linked surveillance as a complementary layer of postmarket
evidence. Near-term process, safety, clinician-override and complication signals should support timely
surveillance; 6-month ICHOM measures may provide intermediate patient-centred signals; and 1-year
outcomes may provide confirmatory evidence. All outcome data should remain linked to the exact
device/model/workflow version used at the index event, with appropriate case-mix adjustment and
predefined investigation triggers.

VII. ICHOM as an Outcome-Standard Layer
The ICHOM Major Injury Set provides a relevant example. It standardises outcomes important to patients
after major injury, including physical health and global functioning, musculoskeletal and ambulatory
function, activities of daily living and autonomy, changes in employment, and other patient-centred
domains.

Licensing note. Where EQ-5D-5L is used within the ICHOM Major Injury framework, implementation must
comply with EuroQol Research Foundation registration, copyright and licensing requirements. EuroQol
states that licensing terms depend on the study/project context and that non-commercial use may be free
after registration; commercial use may attract licence fees. See EuroQol, EQ-5D User License Policy and
'How to obtain EQ-5D'.
For regulatory surveillance and pragmatic implementation, the ICHOM Major Injury primary subset should
be considered before the full data set, reducing collection burden while preserving a standardized core of
patient-centred outcomes and case-mix variables. Additional orthopaedic clinical, process or
device-specific variables may be collected alongside the ICHOM Set, but should be clearly identified as
separate data elements rather than described as extensions of the ICHOM Set.

AO/OTA can describe the injury phenotype. ICHOM can describe what ultimately happened to
the patient. A longitudinal digital evidence layer can connect the two.
A vendor-neutral implementation principle is that the outcome record should remain traceably linked to
the index phenotype, clinician confirmation, treatment pathway and exact device/model version used. An
illustrative implementation is provided in Annex C.

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 8
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Illustrative reference: ORTHO-X OrthoSkills - Outcome Measurement Skill

VIII. A Longitudinal Clinical Evidence Thread
We propose consideration of a general technical concept: a Clinical Evidence Thread. This is not a
proposed regulatory requirement or a specific product design; it is an architecture for maintaining
provenance across a GenAI-assisted clinical episode.

1. Clinical baseline
●​ patient characteristics and case-mix variables
●​ initial clinical state
●​ data provenance

2. AI competency event
●​ intended function
●​ device and relevant foundation-model version
●​ input modalities
●​ output, confidence and uncertainty
●​ source provenance where applicable

3. Human confirmation
●​ accepted, modified, rejected or escalated
●​ reason for override where appropriate

4. Intervention
●​ clinical decision
●​ treatment, procedure or medical device where relevant

5. Outcome
●​ standardised clinical outcome
●​ patient-reported outcome
●​ complications and revisions
●​ follow-up time point

6. Postmarket analytics
●​ competency performance and disagreement rate

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 9
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

●​ calibration and subgroup performance
●​ safety signals and drift
●​ outcome deviations

Such an architecture distinguishes device performance from clinical workflow performance and patient
outcome performance while preserving the possibility of analysing their relationships.

IX. Example: AO/OTA Classification to ICHOM Outcome
Measurement
Stage Illustrative evidence
Clinical input History + examination + imaging + contextual data

AO/OTA classification competency Structured injury phenotype + confidence + alternatives +
uncertainty

Specialist confirmation Treating surgeon confirms, corrects or escalates

Treatment pathway Clinical team selects treatment and retains responsibility

Outcome measurement ICHOM Major Injury primary subset + separately identified
orthopaedic clinical, process and device-specific data
elements

Postmarket evaluation Agreement, calibration, override rates, subgroup
performance, drift, complications, functional recovery and
unexpected outcome patterns

​
The resulting evidence model moves beyond a single accuracy question and separately asks whether the
competency was performed correctly, uncertainty was appropriately managed, clinicians interpreted and
acted on the output appropriately, real-world performance remained stable, and patient outcomes were
consistent with expected care trajectories.

X. A Federated Postmarket Evidence Infrastructure
Where multinational real-world comparators are useful, national and regional trauma registries are more
directly relevant to this orthopaedic-trauma use case than elective hip-and-knee datasets and should be
considered as complementary reference infrastructures.

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 10
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

A federated architecture does not remove data-governance obligations. Implementations should address
jurisdiction-specific privacy and health-data law, lawful access, purpose limitation, data minimisation,
security, provenance, data-subject rights, controller/processor responsibilities and cross-border transfer
requirements. Where feasible, federated analytics should keep identifiable patient-level data under the
control of the originating healthcare institution and exchange only the minimum information required for
the approved evidence purpose. International experience with federated outcome measurement indicates
that these governance challenges are material but surmountable when addressed explicitly in the
architecture.

Stakeholder Illustrative role
Manufacturers Remain responsible for product safety, effectiveness and
regulatory compliance.

Healthcare institutions Provide real-world deployment environments and
appropriately governed clinical-performance data.

Clinicians Provide clinical confirmation, override information and expert
review.

Professional societies Define clinical reference standards, competency taxonomies,
adjudication methods and clinically meaningful failure
categories.

Outcomes organisations Define standardised patient-centred outcomes, time points
and case-mix variables.

Independent public-interest or research organisations May operate governance, benchmarking, surveillance or
evidence infrastructure.

FDA Retains regulatory authority and determines the evidence
appropriate for particular devices and intended uses.

This distributes evidence generation without distributing away manufacturer accountability.

XI. Re-Benchmarking Should Be Triggered by Clinically
Meaningful Change
A GenAI-enabled medical device may change even where its visible application layer has not.
Re-benchmarking should therefore be considered after predefined events such as:

●​ foundation-model version change;

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 11
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

●​ material fine-tuning;
●​ changes to retrieval sources, prompts or orchestration;
●​ new tool integrations or guardrails;
●​ interface changes that may affect clinical interpretation;
●​ expansion to a new intended population, modality or setting;
●​ detected subgroup degradation, clinically meaningful drift or new failure modes.

The relevant question is not simply 'Did the underlying model change?' but 'Could the change
materially alter a validated clinical competency or its associated risk?'

XII. Agentic AI Should Be Evaluated at Skill and Workflow Level
Agentic systems introduce risk not only through isolated outputs but through sequences of individually
plausible actions. We recommend two complementary levels of assessment:

Level 1 - Competency assessment. Each clinically meaningful skill or function is evaluated independently.
Level 2 - Workflow assessment. Representative end-to-end trajectories are evaluated for sequencing
errors, propagation of uncertainty, inappropriate tool use, failures of human confirmation, scope deviation,
cumulative risk and escalation behaviour.

In orthopaedic trauma, an illustrative trajectory might be: imaging adequacy → classification → clinical
reasoning → human confirmation → treatment support → follow-up measurement.

XIII. Outcome-Linked Monitoring and Risk-Proportionate
Regulation
FDA's question whether stronger postmarket monitoring might, in certain circumstances, support greater
premarket uncertainty deserves careful exploration. Such an approach may be more appropriate where
qualified clinician oversight is present, errors are detectable before irreversible harm, systems can be
rapidly disabled or rolled back, deployment is traceable, outputs and overrides are auditable, robust
monitoring exists and predefined reassessment triggers are in place.

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 12
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

It is less appropriate where autonomous actions carry immediate serious risk, errors cannot readily be
detected, supervision is absent, outcome latency defeats timely surveillance, or provenance and
versioning are inadequate.

Outcome monitoring should complement, not replace, adequate premarket evidence.

XIV. Relationship to TEMPO and Real-World Evidence
FDA's Technology-Enabled Meaningful Patient Outcomes (TEMPO) for Digital Health Devices Pilot is a
specific voluntary pilot and should not be extrapolated automatically to GenAI-enabled trauma devices.
Nevertheless, it is a relevant parallel development because it explicitly connects access to selected
digital-health technologies with collection, monitoring and reporting of real-world performance data and
alignment to patient outcomes.

As of July 2026, FDA has announced the first participant selected for TEMPO, demonstrating that the
pilot has moved from policy design to operational implementation.

Digital-health evidence generation can increasingly become continuous rather than episodic:
competency-based evaluation can define an initial reference state; real-world monitoring can
assess whether that state is preserved; standardised outcomes can reveal clinically meaningful
signals over time.

XV. Five Recommendations to FDA
1. Recognise clinically bounded competencies as a useful unit of evaluation. Assess deployed device
functions according to the specific clinical competencies required for intended use rather than relying
primarily on general foundation-model performance.

2. Enable independent specialty organisations to contribute clinical reference standards. Professional
societies and other qualified organisations can contribute benchmarks, sequestered datasets, scoring
rubrics, specialist adjudication and standards while manufacturer accountability remains intact.

3. Make clinician confirmation and human-AI team performance explicit components of clinical
evaluation. For clinician-facing decision support, evaluation should reflect the intended real-world
human-AI configuration.

4. Add tiered outcome-linked surveillance and prospective effectiveness evaluation. Use rapid
process/safety signals for timely monitoring, 6-month patient-centred outcomes as intermediate evidence,

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 13
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

1-year outcomes as confirmatory evidence, and prospective randomized or otherwise appropriately
controlled studies where the objective is to determine whether AI-enabled pathways causally improve
outcomes that matter to patients.
5. Build interoperability between premarket competency evidence and postmarket monitoring. The
competency framework used before deployment should, where appropriate, become the baseline against
which updates, workflow changes and real-world performance are reassessed.

Cross-cutting implementation principle. Federated evidence generation should incorporate privacy,
security, provenance and cross-border governance by design, while preserving manufacturer
accountability and enabling independent clinical/outcomes organisations to contribute standards and
adjudication.

Conclusion
Generative AI may require a regulatory-science architecture that resembles neither conventional static
software testing nor clinician licensure alone. It may instead require a synthesis of defined clinical
competencies, independent clinical standards, structured benchmarking, real-world clinical confirmation,
human accountability, continuous postmarket assessment and meaningful patient outcomes.

Orthopaedic trauma provides a concrete demonstration of how these elements can be connected. The
AO/OTA classification framework can provide a standardised description of what injury occurred.
ICHOM's Major Injury Outcome Set can provide a standardised description of what ultimately happened
to the patient. Clinicians and healthcare institutions provide the judgement and care environment
connecting the two.

patient → phenotype → AI competency → clinician judgement → treatment → outcome
We believe this longitudinal evidence structure could complement FDA's proposed competency-based
approach and contribute to a robust total-product-lifecycle framework for generative and agentic
AI-enabled medical devices.

We would welcome further engagement with FDA, CDRH and the Digital Health Center of Excellence on
specialty-specific competency frameworks, independent clinical adjudication, standardized outcomes and
real-world postmarket evidence architectures.

Signatory

Thomas Morscher

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 14
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Dr. iur., MPA (Harvard)​
Chairman and Cofounder​
Medical AI Validation Network (MAIVAN.ai / ORTHO-X)

Date: October 19, 2026

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 15
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Appendix A — Illustrative Orthopaedic Regulatory-Science
Architecture
Layer Function
Clinical Input History + examination + imaging + contextual data

Competency 1 — Information Sufficiency Is the available information sufficient for the requested task?

Competency 2 — AO/OTA Fracture Classification Structured injury phenotype + confidence + alternatives +
uncertainty + clinician confirmation

Competency 3 — Clinical Reasoning Evidence retrieval + differential reasoning + treatment
mapping + explicit scope limits

Human Clinical Decision Treating clinician / MDT retains responsibility

Treatment Procedure / non-operative care / implant / rehabilitation

Outcome Measurement ICHOM Major Injury primary subset + separately identified
orthopaedic clinical, process and device-specific data
elements

Longitudinal Clinical Evidence Thread AI version + competency outputs + clinician decisions +
overrides + treatment + outcomes

Postmarket Intelligence Re-benchmarking + clinician adjudication + drift detection +
subgroup performance + safety signals + outcome deviation

Continuous Reassessment Competency maintained / investigation / modification /
retraining / regulatory action as appropriate

Appendix B — Regulatory Evidence Matrix
Evidence layer Principal question Potential contributor Orthopaedic example
Clinical standard What constitutes the clinical Professional society AO/OTA fracture
task? classification

Competency benchmark Can the system perform the Independent benchmark Classification accuracy,
task? body / society uncertainty, deferral

Clinical confirmation Does it work in Independent clinicians / Surgeon adjudication of real
representative practice? hospitals trauma cases

Human-AI performance Does augmentation improve Clinical investigators Clinician alone vs clinician +
or preserve care? AI

Postmarket competency Is performance maintained? Manufacturer + independent Periodic re-benchmarking
surveillance

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 16
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

Drift monitoring Has deployment behaviour Manufacturer / healthcare Site/model/version-specific
changed? institutions degradation

Patient outcome What happened to the ICHOM / providers Major Injury outcome
patient? trajectory

Lifecycle evidence Are meaningful signals Multi-stakeholder evidence Classification × treatment ×
emerging? network outcome analysis

Annex C — Illustrative ORTHO-X / OrthoSkills Implementation
and Conflict-of-Interest Disclosure
This annex is provided only to illustrate how the vendor-neutral concepts described in the submission
could be implemented technically. MAIVAN.ai PBC / Medical AI Validation Network is developing
ORTHO-X and OrthoSkills and has a commercial and intellectual-property interest in this work. Inclusion
here does not imply endorsement, validation, clearance or authorization by FDA, AO Foundation, OTA,
ICHOM, or any other regulatory or clinical organisation.​

Illustrative AO/OTA Classification Skill: ​
https://github.com/MAIVAN-ai/OrthoSkills/blob/main/OrthoSkills-v0.1.1_cl/Ortho-Skills/skills/04-aoota-clas
sification/SKILL.md
Illustrative Outcome Measurement Skill: ​
https://github.com/MAIVAN-ai/OrthoSkills/blob/main/OrthoSkills-v0.1.1_cl/Ortho-Skills/skills/10-outcomemeasurement/SKILL.md
Illustrative ORTHO-X OrthoFlow concept: ​
https://maivan.ai/ortho-x-orthoflow-guiding-the-agentic-orthopaedic-patient-journey/

References and Source Links
1. FDA — Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback

2. FDA — Discussion Paper PDF: Considerations for the Regulation of Generative AI-Enabled Medical
Devices

3. Regulations.gov — Docket FDA-2026-N-7874

4. FDA — Launches TEMPO: A First-of-Its-Kind Digital Health Pilot to Expand Access to Chronic
Disease Technologies

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 17
 Response to FDA-2026-N-7874 Regulatory Approach for Generative AI-Enabled Medical Devices
by the Medical AI Validation Network, MAIVAN.ai, Switzerland

5. FDA — TEMPO for Digital Health Devices Pilot

6. FDA — First Participant Selected for TEMPO for Digital Health Devices Pilot (22 July 2026)

7. AO Foundation — AO/OTA Fracture and Dislocation Classification

8. ICHOM — Major Injury Set of Patient-Centered Outcome Measures

9. ORTHO-X OrthoFlow™ — Guiding the Agentic Orthopaedic Patient Journey

10. MAIVAN-ai / ORTHO-X OrthoSkills — AO/OTA Classification Skill

11. MAIVAN-ai / ORTHO-X OrthoSkills — Outcome Measurement Skill

Nothing in this document asserts that the illustrative ORTHO-X/OrthoSkills architecture has
been validated, endorsed, cleared or authorised by FDA, AO Foundation, OTA, ICHOM, or
any other regulatory or clinical organisation unless separately established.

© MAIVAN.ai Medical AI Validation Network, 10-2026 coordinator@maivan.ai 18