Tina Hernandez-Boussard

I appreciate the FDA’s effort to establish a scientifically rigorous framework for evaluating generative AI-enabled medical devices. The proposed competency-based approach provides an important foundation for assessing clinical proficiency, safety, and performance. However, as generative AI systems increasingly participate in clinical decision-making, evaluation must establish more than whether a system produces answers that agree with single-point benchmarks. It must determine whether the outputs from these tools are justified by the information available, remain appropriate as clinical circumstances evolve, and are supported by effective structures of accountability.
The fundamental regulatory objective is to establish whether a device performs its claimed function safely and effectively within its intended use. Evidence must reflect the intended patient population, clinical users, care setting, and workflow in which the device will operate.
My research focuses on the evaluation, implementation, and governance of artificial intelligence in healthcare. Our work on language models, AI-enabled workflows, longitudinal patient trajectories, and governance has highlighted the distinction between technical capability and evidence of clinical readiness. In particular, our research emphasizes the importance of evaluating information sufficiency, whether the information available to an AI system is adequate to support its intended function. This includes assessing whether systems recognize missing or incomplete information, manage uncertainty appropriately, and operate safely within their intended clinical authority. Our published governance research further emphasizes that transparency, independent evaluation, and ongoing oversight are essential to maintaining safety and accountability in AI-enabled healthcare.

FDA-2026-N-7874 | Public Comment

Evaluating Intended Use, Clinical Decision Quality, and Information Sufficiency in
Generative AI Medical Devices
Public Comment to the U.S. Food and Drug Administration
Docket FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices
Tina Hernandez-Boussard, PhD, MPH, MS, FACMI
Associate Dean for Research, Stanford University School of Medicine
Professor of Medicine (Medical Informatics), Biomedical Data Science, and Surgery
Stanford University
Recommendations:
1. Demonstrate performance against intended use. Require evidence that generative AIenabled medical devices perform their claimed clinical functions safely and effectively for the
intended populations, users, and clinical settings. Benchmark performance alone is
insufficient.
2. Establish information sufficiency as an explicit clinical competency. Evaluation should
determine whether AI systems have sufficient information to support their intended functions
and recognize when additional information or human judgment is required.
3. Distinguish agreement-based performance from clinical appropriateness. Reference
standards should assess clinical appropriateness rather than assuming that concordance
with clinicians establishes validity.
4. Evaluate longitudinal clinical decision quality. Conversational and agentic systems should be
assessed across complete decision trajectories, including changes in information,
cumulative errors, escalation, and recovery.
5. Operationalize distributed accountability. Responsibilities should be assigned across
developers, manufacturers, healthcare institutions, and clinical users according to their
ability to observe, control, and intervene in system behavior, while preserving applicable
manufacturer accountability.
These recommendations strengthen the scientific basis of the FDA's proposed framework by
connecting intended-use performance, clinical evidence, decision authority, and ongoing accountability.
Summary
I appreciate the FDA's effort to establish a scientifically rigorous framework for evaluating generative
AI-enabled medical devices. The proposed competency-based approach provides an important
foundation for assessing clinical proficiency, safety, and performance. However, as generative AI
systems increasingly participate in clinical decision-making, evaluation must establish more than
whether a system produces answers that agree with single-point benchmarks. It must determine
whether the outputs from these tools are justified by the information available, remain appropriate as
clinical circumstances evolve, and are supported by effective structures of accountability.
The fundamental regulatory objective is to establish whether a device performs its claimed function
safely and effectively within its intended use. Evidence must reflect the intended patient population,
clinical users, care setting, and workflow in which the device will operate.
My research focuses on the evaluation, implementation, and governance of artificial intelligence in
healthcare. Our work on language models, AI-enabled workflows, longitudinal patient trajectories, and
governance has highlighted the distinction between technical capability and evidence of clinical
readiness. In particular, our research emphasizes the importance of evaluating information sufficiency,
whether the information available to an AI system is adequate to support its intended function. This
includes assessing whether systems recognize missing or incomplete information, manage uncertainty
appropriately, and operate safely within their intended clinical authority. Our published governance
research further emphasizes that transparency, independent evaluation, and ongoing oversight are
essential to maintaining safety and accountability in AI-enabled healthcare.
 FDA-2026-N-7874 | Public Comment

The recommendations below are offered in an individual academic capacity.
1. Information Sufficiency as a Clinical Evaluation Requirement
Relevant FDA questions: 2, 3, 6, 9, and 11
The FDA's proposed competency-based framework appropriately includes information gathering,
clinical analysis, uncertainty communication, and clinical deferral. However, these competencies should
be connected through an explicit assessment of whether the information available to an AI system is
sufficient to support its intended function within the clinical context of use. A clinically appropriate
recommendation does not necessarily demonstrate that the system had sufficient information to justify
it. Therefore, information sufficiency should be assessed independently of whether the resulting output
or decision is correct, based on whether missing or unreliable information could materially affect its
appropriateness or safety.
Consider a generative AI system recommending renewal of a prescription. Its recommendation may
agree with clinical decisions in a retrospective dataset. However, the system may lack recent laboratory
results, medication changes, contraindications, or information about the patient's current clinical status.
Even if the recommendation happens to be correct, the system has not demonstrated that it can reliably
determine when sufficient information is available to support the decision.
Information sufficiency can be defined as the adequacy of available information to justify a specified
output, decision, or action within a defined context of use. This differs from information completeness.
Clinical practice routinely involves uncertainty, and not every decision requires every available patient
datum. The relevant issue is whether missing or unreliable information could materially change the
appropriateness or safety of the proposed action and/or decision. Information should be assessed
relative to the device's intended function, clinical context, and potential consequences of acting without
the information, using prespecified criteria appropriate to the clinical task.
Scope of application
Information sufficiency should be explicitly evaluated for generative AI-enabled medical device
functions, with requirements tailored to the function's intended use. The rigor of evaluation should be
proportional to the consequences and decision authority of the device.
Capabilities to be evaluated
Information sufficiency should be evaluated through five capabilities that assess whether the system
can determine when available information is adequate for its intended function, as applicable to the
device's clinical context and authority:
• Identifying information necessary to support the intended function.
• Recognizing missing, outdated, conflicting, or unreliable information.
• Seeking additional information when it could materially change the appropriateness of the
output or action.
• Deferring or escalating when available information does not justify the intended output or action.
• Revising outputs appropriately when new information becomes available.
These capabilities can be evaluated experimentally. Clinically grounded scenarios can systematically
vary information availability, including laboratory results, medication histories, contraindications, and
patient preferences. Evaluation should measure whether the system changes its behavior appropriately
when clinically consequential information is added, withheld, or corrected. Scenarios should include
cases in which the appropriate response is to proceed, seek additional information, or defer. Proposed
measures:
Measure Proposed definition
Unsupported Proportion of clinically consequential recommendations or actions made without
decision rate information necessary to justify them
Appropriate Proportion of information-insufficient cases in which the system appropriately seeks
deferral rate information, escalates, or defers
 FDA-2026-N-7874 | Public Comment

Information Extent to which recommendations change appropriately when decision-relevant
sensitivity information is modified and remain stable when changes should not affect the decision

These measures are proposed for functions that inform or perform clinical decisions and should be
adapted to the device's intended function. They should undergo independent validation across
representative patient populations, clinical settings, and information conditions. Validation should
assess agreement with independent clinical adjudication, reproducibility, and the ability to identify
clinically consequential failures not captured by conventional accuracy metrics. Acceptance criteria
should be prespecified and justified according to intended use and potential clinical harm.
Implications for directiveness and information asymmetry
Information sufficiency depends on the device's intended use and the clinical authority granted to it.
Information sufficient to recommend treatment may be insufficient to initiate it autonomously. The
information available to the intended user also matters. A clinician with access to a patient's medical
record may be able to assess a recommendation that a patient cannot independently verify.
Conversely, patients may have relevant information unavailable to the AI system or clinician. Evaluation
should therefore consider both the information available to the AI and the information available to the
person expected to act on its output..
Recommendation: FDA should incorporate information sufficiency into clinical proficiency and safety
evaluation, with independently validated measures and evidence requirements proportional to the
device’s intended use, clinical consequences and decision authority.
2. Evaluating Clinical Appropriateness and Effectiveness
Relevant FDA questions: 10, 14, and 15
The FDA appropriately recognizes that open-ended generative AI outputs present challenges for
conventional reference-standard evaluation. In many clinical settings, there is no single correct answer.
Multiple decisions may be reasonable, clinical evidence may be incomplete, and patient preferences
may influence the appropriate course of care. Conventional benchmarks frequently assess whether AI
reproduces a predefined answer, whereas clinical decision-making requires recognizing uncertainty,
identifying missing information, and determining when no definitive recommendation is justified. Our
research evaluating medical AI chatbots highlights the limitations of treating benchmark success as
evidence of clinical decision quality.
For FDA evaluation, this distinction has direct implications for reference-standard construction:
agreement with an expected answer should not substitute for assessment of whether the
recommendation was justified under the information conditions presented to the system. Four
constructs can be distinguished:
Construct Evaluation objective
Clinician agreement Concordance between AI recommendations and clinician judgments
Guideline Consistency with established clinical recommendations
concordance
Clinical Whether a decision is justified by the evidence, patient context, and acceptable
appropriateness clinical alternatives
Clinical effectiveness Whether use of the system produces meaningful clinical benefit relative to an
appropriate comparator

Expert consensus can provide a useful reference standard, but consensus should account for clinically
acceptable alternatives and uncertainty. Clinicians may disagree for legitimate reasons.

Comparator selection and intended-use performance
FDA's Q14 asks how the applicable standard should be defined when a clinician panel serves as the
 FDA-2026-N-7874 | Public Comment

comparator. Three positions follow from the analysis above:
• Standard of care versus median clinician. Guideline-concordant care should inform the
reference where applicable guidelines exist, while allowing justified individualized care. Where
guidelines do not resolve the decision, an independent expert consensus reference should be
constructed before outputs are seen, with inter-adjudicator agreement reported and clinically
acceptable alternatives documented. Median clinician performance should not automatically
become the acceptance threshold because it may incorporate deficiencies in current practice.
• Generalist versus specialist. Comparator expertise should follow the intended use and the
clinical claim. A generalist-facing device should be evaluated in the context of generalist
practice, and a specialist-facing device in the context of specialist practice. Comparator
selection should reflect the clinical task and expertise required to evaluate the device's claims,
rather than assuming that performance equivalent to the intended user group is necessarily
sufficient. The sponsor should specify and justify the comparator.
• Human-AI team versus device alone. For devices intended to assist clinicians, evaluation
should assess performance of the human–AI team relative to an appropriate unassisted
comparator. Device-alone performance should also be assessed when relevant to the device's
intended function and clinical authority.
The reference standard, comparator, and endpoints should reflect the device's intended use and
claimed clinical benefit. Demonstrating clinical appropriateness alone does not establish that
deployment improves patient outcomes or care delivery. For devices intended to support clinician
decision-making, evaluation should assess whether AI assistance improves or degrades clinical
performance relative to unassisted care, including inappropriate reliance, failure to recognize erroneous
recommendations, and variation across intended users and clinical settings.
Clinically indeterminate decisions
Clinical evaluation should distinguish among clinically appropriate, clinically inappropriate, and clinically
indeterminate decisions. The last category is essential because insufficient information or conflicting
evidence may prevent a definitive judgment. Evaluation frameworks that require a single answer in
every case risk rewarding unjustified certainty and penalizing appropriate deferral. Indeterminate cases
should be independently adjudicated and separately reported. Evaluation should distinguish
appropriate deferral, unnecessary deferral, and failure to defer when clinically required. Independent
adjudication should also assess which information is necessary to justify a decision, using prespecified
clinical criteria to inform validation of the information-sufficiency measures proposed in Section 1.
Information asymmetry between AI and comparator
An additional source of evaluation bias arises when the AI system and its comparator operate under
different information conditions. A clinician panel reviewing a complete medical record may reach a
different decision from an AI system interacting with a patient who has disclosed only limited
information. Conversely, an AI system with access to longitudinal records may possess information
unavailable to the clinician at the point of care. Where the purpose is to compare clinical reasoning, the
information available to the AI and comparator should be matched or explicitly accounted for. Where
the purpose is to evaluate the effectiveness of the deployed system, differences in information access
should be retained as part of the intervention being studied. This distinction is essential because
apparent performance differences may reflect information availability rather than differences in clinical
reasoning.
Recommendation: FDA should require sponsors to demonstrate that devices perform their claimed
clinical functions safely and effectively within their intended use. Reference standards, comparators,
and endpoints should reflect those claims, characterize uncertainty and clinically acceptable
alternatives, and distinguish agreement-based performance from clinical appropriateness and
effectiveness.
3. Longitudinal Decision Quality for Conversational and Agentic AI
 FDA-2026-N-7874 | Public Comment

Relevant FDA questions: 5, 6, and 26
Generative AI increasingly operates through conversations and sequences of actions rather than
isolated predictions. A system may gather information, interpret findings, recommend treatment, initiate
follow-up, and revise its recommendations as new information becomes available. Our research on
longitudinal patient trajectories, clinical decision pathways, and digital twins provides a foundation for
examining how patient states, information, and clinical decisions evolve over time. For conversational
and agentic AI, the clinically meaningful sequence of interactions of decisions should be a unit of
evaluation.
Failure modes missed by per-turn assessment
• Cumulative scope drift: Individual responses remain apparently within intended use while the
interaction progressively moves toward an unauthorized clinical function.
• Changing information sufficiency: Information adequate for an initial recommendation becomes
insufficient for a later decision.
• Delayed escalation: A system eventually recommends appropriate escalation but does so after
a clinically important delay.
• Error propagation: An incorrect assumption or intermediate decision influences subsequent
recommendations and actions.
• Failure to revise: New findings invalidate an earlier recommendation without prompting an
appropriate change in the care pathway.
Consider an agent supporting longitudinal medication management. An initial recommendation may be
clinically justified. Subsequent laboratory findings or changes in patient status may alter the
appropriateness of continued treatment. If the agent proceeds without reassessment, individual actions
may remain technically permissible while the overall clinical pathway becomes unsafe. Trajectory
boundaries and evaluation endpoints should be prespecified and justified according to the device's
intended use and clinical workflow. Sponsors should define the events that initiate and terminate a
trajectory and ensure that the evaluation captures clinically consequential dependencies between
interactions, decisions, and actions. Sample sizes should be sufficient to estimate clinically
consequential failure rates with appropriate precision, including targeted evaluation of rare but serious
failures. Compliance with permitted action boundaries does not establish that the sequence of
permitted actions is clinically justified.
Operational requirements
FDA should require trajectory-based evaluation with prespecified decision points, safety-critical events,
and clinically meaningful changes in information availability. The following should be explicit in any
evaluation plan:
• Acceptance criteria should be stated at the trajectory level, such as the proportion of trajectories
containing an unsupported action or delayed escalation beyond a clinically justified window.
• Safety-critical boundary crossings should be reported explicitly, not averaged away across
trajectories.
• Results should be reported by trajectory type (naturalistic, adversarial, long-horizon) and by
information condition (complete, incomplete, corrected, contradictory).
• Duration and complexity should be justified against the intended clinical workflow and the
consequences of the device's actions.
Evaluation scenarios should include naturalistic clinical interactions and controlled perturbations
involving delayed results, contradictory information, changing patient status, and resistance to
appropriate escalation. These perturbations should help identify when and why failures occur.
Recommendation: FDA should incorporate longitudinal decision-quality assessment for devices whose
clinical functions involve multi-turn interactions or interdependent actions. Evaluation should establish
whether the system maintains information sufficiency, recognizes changing circumstances, and
appropriately revises or terminates clinical actions over time.
 FDA-2026-N-7874 | Public Comment

4. Distributed Accountability Across the Clinical AI System
Relevant FDA questions: 16, 21, 24, and 25
The scientific evaluation of generative AI must be connected to the governance structures responsible
for maintaining safe performance after deployment.
Our research on governance of AI-enabled healthcare, including our framework for transparency,
evaluation, and shared accountability in mental health chatbots, emphasizes that safe deployment
depends on more than initial model evaluation. Our work on transparency in biomedical AI further
emphasizes documentation and stakeholder participation. More recently, our analysis of Utah's clinical
AI sandbox highlighted the importance of independent access to evidence when assessing emerging
autonomous clinical applications.
Generative AI-enabled devices may depend on foundation models, retrieval systems, external tools,
institution-specific configurations, clinical workflows, and human oversight. These components may be
developed, controlled, and modified by different organizations. Accountability is inherently distributed
because control over system behavior is distributed.
Manufacturer accountability and shared responsibilities
These roles should be made explicit without weakening manufacturer accountability. Three principles
are essential:
• Manufacturer accountability for the integrated device must be preserved. Assigning observation,
control, or intervention functions to other parties should not diminish the manufacturer's
applicable regulatory obligations for device safety and effectiveness.
• Third-party dependencies must be governed. Where a manufacturer cannot fulfill a
responsibility because a third party controls a component, such as a foundation model update,
the limitation should be disclosed and addressed through contractual, technical, or operational
controls.
• Responsibilities must be actionable. Safety-critical duties should not be assigned to patients or
clinicians who lack the information, resources, or authority to fulfill them. Manufacturers and
deploying institutions should demonstrate that their respective responsibilities are supported by
workable controls.
I propose organizing distributed accountability around four operational functions:
1. Observation: The ability and responsibility to detect performance failures, missing
information, inappropriate actions, and changes in system behavior.
2. Control: The authority to modify system components, restrict functionality, or change the
clinical workflow in which the device operates.
3. Intervention: The ability to suspend, override, roll back, or withdraw clinical functionality
when safety conditions are no longer satisfied.
4. Responsibility: The obligation to investigate failures, document corrective actions,
communicate relevant findings, and demonstrate continued appropriateness for intended
use.
For example, a manufacturer may be responsible for monitoring a device that depends on a third-party
foundation model. If the manufacturer cannot detect changes to that model, its ability to fulfill that
responsibility is compromised. The manufacturer should establish reliable mechanisms for detecting
consequential changes or restrict reliance on components it cannot adequately oversee.
Healthcare organizations also require defined responsibilities for local monitoring, incident reporting,
and intervention when a system becomes unsafe in its deployment environment. These responsibilities
complement manufacturer obligations rather than replacing them. These functions should be
demonstrated through documented and tested procedures, including the ability to detect consequential
changes, exercise control over affected functionality, implement timely corrective actions, and establish
clear responsibility for investigation and remediation.
 FDA-2026-N-7874 | Public Comment

Relationship to information sufficiency
Distributed accountability is directly connected to information sufficiency. A system's ability to make
justified decisions may depend on data interfaces, retrieval tools, and workflows controlled by different
parties. When information becomes unavailable, outdated, or unreliable, the governance system must
detect the change and prevent clinical actions that are no longer justified.
For agentic systems, governance should establish who can restrict or withdraw decision authority when
the conditions supporting that authority are no longer satisfied. Intervention mechanisms should be
technically feasible, assigned to identifiable actors, and tested under realistic failure conditions.
Governance plans should also specify safe fallback procedures, who may authorize restoration of
functionality, and what evidence is required before the system resumes clinical actions.
Patient roles
Patients and patient organizations contribute observation and reporting that technical monitoring cannot
replace. Patients may identify downstream consequences of AI interactions that are not captured in
system logs or clinical records. For patient-facing devices, this should include:
• Mechanisms for capturing patient-reported behavior and outcomes following AI interactions,
where relevant to the device's intended use.
• An accessible, protected patient reporting pathway for postmarket safety concerns
• Defined responsibilities for reviewing patient-reported concerns, investigating potential harms,
and implementing corrective actions.
Patient reporting should strengthen postmarket safety oversight without transferring responsibility for
detecting or preventing device failures to patients. These mechanisms should complement ongoing
manufacturer obligations and institution-level oversight.
Recommendation: FDA should encourage explicit allocation of observation, control, intervention, and
responsibility across the AI-enabled clinical system. These arrangements should preserve
manufacturer accountability while demonstrating that other assigned parties possess the authority and
capabilities necessary to fulfill their responsibilities.
5. Integrating Clinical Decision Validity and Accountability
The FDA's proposed competency-based framework provides a practical foundation for incorporating
these recommendations. Intended use should determine the evidence required to establish that a
device performs its claimed clinical function. It defines the relevant patients, users, settings,
comparators, endpoints, and acceptable risks. Four complementary dimensions of evidence should be
distinguished:
Dimension Evaluation objective
Output validity Establish whether outputs are accurate, appropriate, and supported within the
intended use
Decision-process Establish whether outputs, decisions, or actions are supported by sufficient
validity information and appropriate management of uncertainty
Longitudinal decision Establish whether outputs, decisions, or actions remain appropriate as
validity information and clinical circumstances evolve
Operational Establish whether responsible parties can observe, control, and intervene when
accountability conditions supporting safe use change
Clinical effectiveness remains an essential additional consideration when required to substantiate the
device's clinical claims.
The evidence required should be proportional to intended use, clinical consequences, and the authority
granted to the system. The ability of the intended user to independently assess the output and the
availability of effective safeguards should also inform evidence requirements. Evidence requirements
 FDA-2026-N-7874 | Public Comment

should reflect both the severity and reversibility of potential harm and the degree of clinical decision
authority exercised by the device. These requirements should be reassessed when changes in
intended use, system capabilities, or deployment context materially alter either dimension.
The level of clinical authority granted to a system should remain connected to the evidence supporting
its use. Changes in information availability, clinical context, or system behavior may require
reassessment of that authority.
This establishes a direct relationship between scientific evaluation and governance. Evaluation
establishes the conditions under which a device can safely and effectively perform its intended function.
Governance ensures that those conditions remain satisfied and provides the capacity to intervene when
they do not.

Respectfully submitted,

Tina Hernandez-Boussard, PhD, MPH, MS, FACMI
Stanford University School of Medicine
Submitted in an individual academic capacity.

Selected Supporting References
1. Lee H, Handler R, Mungle T, Hernandez-Boussard T. Building safer artificial intelligence
mental health chatbots: a framework for transparency, evaluation, and shared
accountability. J Am Med Inform Assoc. 2026 Aug 1;33(8):1538-1553. doi:
10.1093/jamia/ocag078. PMID: 42161859; PMCID: PMC13385991.
2. Handler R, Sharma S, Hernandez-Boussard T. The fragile intelligence of GPT-5 in medicine.
Nature Medicine. 2025. doi:10.1038/s41591-025-04008-8.
3. Hernandez-Boussard T, Macklin P, Greenspan EJ, et al. Digital twins for predictive oncology
will be a paradigm shift for precision cancer care. Nature Medicine. 2021.
doi:10.1038/s41591-021-01558-5.
4. Hernandez-Boussard T, et al. Promoting transparency in AI for biomedical and behavioral
research. Nature Medicine. 2025. doi:10.1038/s41591-025-03680-0.
5. Gorijavolu, R., Jaiswal, N., Kesselheim, A.S. et al. What Utah’s clinical AI sandbox reveals
about independent oversight. Nat Med 32, 2330–2332 (2026).
https://doi.org/10.1038/s41591-026-04418-2.
6. Gu, Y. and Topol, E.J., 2026. Decision authority in health AI. Nature Health, pp.1-3.