Danaher Corporation (Ronald D. Dunn)

See attached file(s)

October 19, 2026

VIA Electronic Submission to www.regulations.gov

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback (FDA-2026-N-7874)

On behalf of Danaher Corporation (Danaher), we appreciate the opportunity to comment on the
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback (FDA-2026-N-7874) and respectfully submit the summary and
responses below. As a manufacturer of in vitro diagnostic (IVD) instruments and assays, digital
pathology systems, and laboratory informatics software used by clinical laboratories and
hospitals worldwide, Danaher has direct experience developing, validating, and maintaining AIenabled software within FDA’s existing regulatory framework.

Summary of Key Recommendations

 One risk framework for all device software. FDA should establish a single risk-based
framework for software functions, consistent with IMDRF SaMD principles and the
existing device classification system, and address GenAI-specific considerations (such
as conversational drift) within that framework rather than through a separate regime.
 Clarify the device/non-device boundary. FDA should explain how the continuum from
“non-directive” to “action-directing” informational functions relates to the non-device
clinical decision support criteria in section 520(o)(1)(E) of the Federal Food, Drug, and
Cosmetic Act (FD&C Act) and FDA’s clinical decision support guidance.
 Build on existing pathways and policy. Competency-based evaluation should be
implemented through existing premarket pathways, FDA-recognized consensus
standards, the Q-Submission program, and FDA’s existing AI policies, including its
PCCP guidance and the guiding principles developed with Health Canada and the
MHRA, and should reflect lessons learned from the Software Precertification Pilot
Program.
 Treat benchmarks as supporting evidence. Benchmarking is valuable but should not,
on its own, establish a reasonable assurance of safety and effectiveness. Evidence
expectations, including independence safeguards, should be proportionate to risk.
 Leverage established IVD and laboratory practice. Established IVD methods for
precision, analytical and clinical validity, reference standards, and adjudication offer
tested models for evaluating GenAI output variability and benchmark validity.
 Enable predictable lifecycle change. FDA should scale modification evidence to risk,
permit lower-risk changes to be managed within the quality management system, and
adapt PCCPs to accommodate bounded GenAI and foundation-model changes.
  Harmonize internationally. FDA should align its approach with IMDRF and coordinate
with other regulators, including in light of the EU AI Act and the EU MDR and IVDR, to
avoid divergent requirements for globally marketed devices.

Responses to Questions

1) Does the two-axis risk framework, organized around device activity and the
consequence of relying on an incorrect output, appropriately capture the dimensions
most relevant to the risk of a GenAI-enabled software function? If there are additional
dimensions—such as the reversibility of a resulting action, the availability of
downstream safeguards, the time pressure of the deployment setting, or the traceability
of the output (i.e., to primary source materials)—that should be represented in a risk
framework, please describe and provide examples of how they should be represented.

2) CDRH seeks input on how to account for the spectrum of GenAI-enabled informational
functions that vary in the degree to which they direct a user to a particular action (i.e.,
between “non-directive” and “action-directing”). What characteristics of an output—such
as its wording, specificity, personalization, or context—could be considered as modifiers
of the risk associated with an informational function, after accounting for the device’s
overall functionality and intended use? What additional information would provide
manufacturers with sufficient clarity and predictability around risk assessment for
informational functions while recognizing that directiveness may exist along a
continuum rather than as a binary distinction?

3) CDRH seeks input on whether and when a GenAI-enabled function that results in
delivery of clinical information to patients, as opposed to HCPs, could present different
or higher risks, while also recognizing the potential benefits associated with improved
patient empowerment, engagement, and access to clinical information. What device
characteristics, output features, or safeguards might mitigate risks that could arise when
a user lacks the domain knowledge to independently evaluate an output, without
unnecessarily underestimating patient capability?

4) CDRH seeks input on whether and how the distinction between generalist and
specialist physicians might inform the assessment of risk for HCP-facing GenAI-enabled
functions. Under what circumstances, if any, might risk be affected when an HCP who
lacks the relevant clinical specialist knowledge receives information that falls within a
specialist area of practice? What device characteristics or safeguards might mitigate
such risks?

5) For multi-turn conversational GenAI-enabled devices that may migrate from providing
“non-directive” information to “action-directing” information over the course of an
exchange, how should risk be assessed across realistic conversational trajectories?
How could the intended use of such a device be characterized when its behavior is
emergent across a conversation?

6) For GenAI-enabled care escalation functions, CDRH is considering how the evaluation
may account for both under-escalation and over-escalation. How could manufacturers
characterize and weigh these two directions of error, given that they may not be
commensurable and that acceptable trade-offs may vary by clinical context?

Q1–Q6 (General Comments): The two-axis framework FDA has proposed in Figure 1 of the
Discussion Paper is closely aligned with the framework put forth by the International Medical
Device Regulators Forum (IMDRF) in 2014 and published in “Software as a Medical Device:
Possible Framework for Risk Categorization and Corresponding Considerations” (IMDRF/SaMD
WG/N12FINAL:2014). This framework is provided below for reference (Figures A and B) and
was also leveraged significantly in IMDRF’s “Software as a Medical Device (SaMD): Clinical
Evaluation” guidance (IMDRF/SaMD WG/N41FINAL:2017).
Figure A: Risk Categorization Framework presented in IMDRF Risk Categorization
Guidance.

Figure B: Example of Risk Categorization Framework Implemented in IMDRF SaMD
Clinical Evaluation Guidance.

Many of the considerations raised in Q1–Q4 and Q6 were also examined in developing and
implementing the IMDRF SaMD framework and during the Agency’s Software Precertification
Pilot Program. We encourage FDA to build on that body of work, including the associated
guidance, working-group records, and pilot findings, as it refines its approach for GenAI-enabled
devices.

Since 2014, however, the principles described in this IMDRF framework have not been fully
translated into regulators’ medical device classification schemes, including FDA’s. For example,
no guidance currently exists from FDA regarding how to translate this framework into a Class I,
II, or III classification for a given device. FDA’s 2017 SaMD Clinical Evaluation guidance
containing this framework was withdrawn in January 2026. Clear, consistently applied
classification rules for SaMD would benefit patients, FDA, and manufacturers alike. Thus,
although Danaher does not oppose the framework described in Figure 1 of the Discussion
Paper, we encourage FDA to address this implementation gap so that any risk framework for
GenAI-enabled device software translates predictably into classification, submission pathway,
and evidence expectations.

Further, it should also be noted that GenAI-enabled devices, while having unique
characteristics, are a subset of SaMD. The framework FDA has proposed in Figure 1 of the
Discussion Paper is not unique to GenAI-enabled device software – it could be applied to any
SaMD (software medical device function). The unique aspect of GenAI-enabled device
software, in relation to Figure 1, is that it may evolve across the proposed axes, as described in
Question 5.

Danaher recommends against establishing different risk-based frameworks for non-AI enabled
software medical device functions (or non-AI enabled software as a medical device, SaMD) and
GenAI-enabled device software, and we recommend that the Agency not create a unique risk
categorization approach only for GenAI-enabled device software. Rather, we encourage FDA to
focus its efforts on developing a singular, risk-based classification framework that fits within the
existing medical device classification scheme and that can be applied to software medical
device functions (i.e., SaMD) and GenAI-enabled device software alike. Any nuances related to
GenAI-enabled device software functions, such as the situation described in Question 5, can be
considered within the overall context of such a SaMD risk categorization framework.

Within this unified framework, FDA should clarify the distinction between the Activity and
Consequence axes. The Consequence axis should primarily reflect the severity of harm if an
output is incorrect, including the seriousness and urgency of the clinical condition, the
reversibility of harm, and the consequences of false-positive and false-negative outputs.

The Activity axis should principally reflect how likely the GenAI output is to be followed or acted
upon. Relevant considerations include the degree of automation, the expertise and availability of
human review, the effectiveness of downstream safeguards, the user’s ability to independently
evaluate the output, and the extent to which the output directs a specific action. This approach
would better account for differences among autonomous, patient-facing, generalist-facing,
specialist-reviewed, and laboratory-reviewed functions.

FDA should also recognize that directiveness exists along a continuum. Although wording,
specificity, personalization, and expressed confidence may affect reliance, the availability of
effective guardrails and qualified human review may be more consequential than wording alone.
FDA should provide examples illustrating how these factors affect placement on the Activity
axis.
For care-escalation functions, under-escalation and over-escalation should be evaluated
separately because false-negative and false-positive determinations present distinct clinical
risks across indications. These errors should be assessed primarily on the Consequence axis,
with context-specific performance targets based on clinical acuity, reversibility of harm,
confirmatory testing, and downstream safeguards.

Question-Specific Comments (Q1–Q5)

Q1 – Additional risk dimensions. Of the additional dimensions identified in Question 1,
reversibility of harm is best reflected on the Consequence axis and the availability of
downstream safeguards on the Activity axis, as described above. Time pressure in the
deployment setting should be treated as a modifier on the Activity axis, because it reduces the
opportunity for independent verification before an output is acted upon. For example, an output
used in emergency department triage warrants different treatment than the same output
reviewed during routine laboratory sign-out. Traceability of the output to primary source
materials should also be recognized as a risk-mitigating characteristic. Where a GenAI-enabled
function cites verifiable primary sources (e.g., the specific laboratory result, image region,
clinical guideline, or labeling on which it relies) and the intended user can efficiently check those
sources, the likelihood of an undetected erroneous output is reduced. FDA should clarify that
validated traceability and source-grounding controls may support a lower placement on the
Activity axis, provided their effectiveness is demonstrated with representative users.

Q2 – Directiveness and the device/non-device boundary. The continuum between “nondirective” and “action-directing” informational functions overlaps substantially with the criteria for
non-device clinical decision support (CDS) software under section 520(o)(1)(E) of the FD&C Act
and FDA’s CDS guidance. Some GenAI informational functions intended for HCPs will fall
outside the device definition, while others will not. To provide clarity and predictability, FDA
should (1) explain how the factors it identifies (wording, specificity, personalization, and context)
relate to the statutory CDS criteria, including the criterion that an HCP be able to independently
review the basis for a recommendation; (2) provide paired examples of GenAI functions on
either side of the device boundary and at different points along the directiveness continuum; and
(3) confirm that the risk framework applies only after a function has been determined to be a
device. Examples drawn from laboratory medicine, such as GenAI-generated interpretive
comments on test results, would be particularly helpful to IVD manufacturers and clinical
laboratories.

Q3 – Patient-facing functions. Delivery of clinical information directly to patients can present
different risks, primarily because a patient may be less able to independently evaluate an output
and may act on it without intervening review by an HCP. However, patient-facing delivery should
be assessed as one input to the Activity axis rather than as an automatic elevation in risk.
Relevant mitigations include constraining outputs to the patient’s own validated results and to
FDA-authorized labeling; plain-language explanations with clear statements of uncertainty and
limitations; consistent prompts to consult an HCP before acting on clinically significant
information; reliable detection of, and escalation for, urgent findings; and usability evaluation
with representative lay users in accordance with IEC 62366-1. For IVD results in particular,
functions that explain a result to a patient should be distinguished from functions that
recommend treatment or changes in care, which present higher risk.

Q4 – Generalist, specialist, and laboratory users. Risk may be affected when an HCP
receives information outside his or her area of expertise, because the user’s ability to recognize
an erroneous output is reduced. FDA should treat the expected expertise of the intended user,
relative to the subject matter of the output, as a modifier on the Activity axis. FDA should also
recognize that many diagnostic outputs are first reviewed by laboratory professionals,
pathologists, and laboratory directors working under CLIA and accredited quality systems, and
that this review can be an effective downstream safeguard. Mitigations include labeling that
clearly defines intended users and their expected qualifications; output design that identifies
when information falls within a specialist domain and recommends specialist consultation; and,
where appropriate, human-device performance testing with representatives of each intended
user group.

Q5 – Multi-turn conversational devices. Question 5 identifies the characteristic that most
clearly distinguishes GenAI-enabled device software from other SaMD. Danaher recommends
that the intended use of a multi-turn conversational device be defined by the functional
boundaries the device is designed and validated to enforce (the topics, actions, and degree of
directiveness it is permitted to provide) rather than by any individual output. Risk should then be
assessed at the highest-risk state the device can reach across realistic conversational
trajectories, including trajectories in which users persist, rephrase, or attempt to elicit out-ofscope advice. Validation should include prespecified scripted and adversarial (“red-team”)
conversation sets representative of intended users, with performance metrics such as the rate
at which the device appropriately declines, redirects, or escalates out-of-scope requests. Where
scope-enforcement controls are shown to be effective, the device should be regulated according
to its bounded intended use; where they are not, the device should be assessed at the higherrisk function it can realistically reach. Postmarket monitoring of real conversational trajectories
should confirm that scope boundaries hold in actual use.

7) Is the competency-based approach described above, i.e., device benchmarking
followed by clinical confirmation, a useful and appropriate framework for evaluating
GenAI-enabled devices?

Danaher supports FDA’s objective of developing risk-proportionate approaches for evaluating
GenAI-enabled device software. We note, however, that the proposed competency-based
framework incorporates several concepts previously explored through the Software
Precertification Pilot Program, including broad capability assessment, risk-tailored premarket
review, reliance on structured objective evidence, and continued assurance through real-world
performance monitoring. In its September 2022 final report, FDA concluded that the Pre-Cert
Working Model was not practical to implement under its existing statutory and regulatory
authorities and identified substantial difficulties involving standardized appraisal criteria,
repeatable methodologies, KPI selection, streamlined review, real-world performance
measurement, and integration with existing device-classification pathways.

The Discussion Paper does not yet explain how the proposed GenAI competency-based
approach differs materially from those prior concepts, which lessons from the Pre-Cert Pilot
informed the current proposal, or how the legal and operational barriers documented in the final
Pre-Cert report would be overcome. To give stakeholders confidence in a competency-based
framework for GenAI-enabled devices, we recommend that FDA clarify: (1) how the approach
would operate within existing statutory authorities and premarket pathways; (2) how
competency assessment would affect classification, submission pathway, and product-specific
evidence requirements; (3) the criteria and validated measures FDA would use to ensure
consistent and repeatable assessment; (4) how postmarket evidence would be collected,
standardized, and linked to regulatory action; and (5) why a GenAI-specific structure is
preferable to a coherent framework applicable to SaMD and software medical device functions
more generally.

Absent this clarity, there is a risk that the proposed framework will reproduce the conceptual and
implementation challenges encountered in Pre-Cert without providing manufacturers with a
practical, predictable, or least-burdensome pathway. Danaher therefore recommends that FDA
explicitly map the proposed competency-based framework to the findings and limitations of the
Pre-Cert Pilot before developing future policy.

Danaher believes many of the benefits FDA seeks can be realized within the existing
framework. Competency-based evidence could be incorporated into 510(k), De Novo, and PMA
submissions through FDA-recognized consensus standards, the Q-Submission program, and
PCCPs, and could be refined through a voluntary pilot with defined success criteria. This
incremental approach would allow FDA and industry to learn from practical experience while
avoiding the implementation barriers identified in the Pre-Cert final report.

8) How might the two-axis risk framework described in Section IV be considered within
the competency-based approach described above to help determine the level of
evidence needed for a premarket submission?

Previously, IMDRF’s “Software as a Medical Device (SaMD): Clinical Evaluation” guidance
contemplated the level of independent review needed, based on the IMDRF SaMD Risk
Categorization. This approach is illustrated in Figure B above. A similar model could be applied
to the two-axis framework FDA has proposed in this discussion paper, with increasing evidence
required as the device risk increases.

Danaher recommends that FDA publish an illustrative matrix mapping each combination of
Activity and Consequence to expected evidence elements. For example, lower-risk
combinations might be supported by benchmarking, analytical validation, and retrospective
evaluation on representative data, with independent involvement limited to execution of
sequestered tests. Higher-risk combinations would add clinical confirmation using real-world or
prospective data, human-device performance testing, and greater independence in evaluation.
Such a matrix, accompanied by worked examples, would give sponsors the predictability
needed to plan development programs and Pre-Submissions.

9) Would a benchmarking structure such as the one described above be likely to provide
adequate evidence of clinical knowledge, analytic capabilities, safety behavior,
communication, and generalizability to support a reasonable assurance of safety and
effectiveness? Are there elements that are missing, redundant, or inappropriately
categorized? Are there externally developed standards that could be leveraged?

Danaher agrees that structured benchmarking may be a useful component of the premarket
evaluation of a GenAI-enabled device. However, a benchmarking structure such as the one
described would not, standing alone, be likely to provide adequate evidence to support a
reasonable assurance of safety and effectiveness. Benchmark results should be considered
supporting evidence within an intended-use-specific validation framework that includes
appropriate analytical validation, clinical validation, usability evaluation, risk management, and,
where warranted, clinical confirmation.

High performance on a prespecified benchmark should not, by itself, be presumed to
demonstrate generalizable clinical capability. Unlike competency assessment of human
professionals, strong performance by a GenAI model may reflect optimization or overfitting to a
particular evaluation framework rather than reliable performance across novel, uncertain, or
previously unseen clinical scenarios.

The following elements should be considered more significantly within such an approach:

 Intended Use Fidelity: The framework should assess whether the device consistently
operates within its intended use, including across multi-turn interactions. A device might
perform individual benchmark tasks correctly but allow the interaction to migrate into an
unvalidated clinical function.
 Reliability and Repeatability: Because GenAI outputs may vary for the same or similar
inputs, benchmarking should characterize the distribution of performance across
repeated trials, not merely score one response per test case.
 Human-Device Performance: As many devices are intended to support, rather than
replace, human judgment, benchmarking should assess whether representative users
can appropriately interpret, verify, accept, modify, or reject the output and whether use of
the device improves or degrades the performance of the combined human-device
system.
 Generalizability: The framework should assess performance using independent
datasets, previously unseen scenarios, diverse clinical conditions, and clinically relevant
 edge cases. For higher-risk functions, sponsors should not rely solely on evaluations
they designed and controlled to establish generalizability and reproducibility.

Where sponsors define evaluation methods, scoring rubrics, acceptance criteria,
generalizability, or reproducibility requirements, they should implement safeguards against
benchmark gaming that are proportionate to the device’s risk. Such safeguards may include
sequestered test sets, independent review of benchmark design, blinded execution, and
evaluation against datasets unavailable during model development.

With respect to externally developed standards, existing, FDA-recognized medical device
standards, such as ISO 14971, IEC 62304, IEC 62366-1, are applicable to GenAI-enabled
device functions and medical devices alike and should be leveraged in the development and
lifecycle management of these products. Other, cross-sector AI standards, such as ISO/IEC
23894 and ISO/IEC 42001 have also been established. We request that the Agency consider
established medical device and AI standards, as well as the guiding principles for good machine
learning practice, PCCPs, and transparency that FDA developed with Health Canada and the
MHRA.

Established IVD evaluation practice also offers tested models for several of the challenges FDA
identifies. Precision studies, such as those described in CLSI EP05 and EP12, characterize
variability across repeated runs, operators, and conditions, an approach readily adapted to the
output variability of GenAI models. The distinction between analytical validity and clinical validity
provides a framework for separating whether a benchmark measures what it purports to
measure from whether that performance predicts clinical benefit. IVD practices for establishing
reference standards, adjudicating discordant results, and evaluating interference provide
additional models. Danaher encourages FDA to draw on these approaches and on FDArecognized CLSI standards rather than developing GenAI-specific methods from first principles.

Benchmarking should be treated as one source of evidence within a risk-based, intended-usespecific validation strategy, not as a stand-alone demonstration of reasonable assurance of
safety and effectiveness. FDA should first define and validate the proposed competency
constructs, require clinically justified domain-level acceptance criteria, address missing
dimensions such as uncertainty, abstention, traceability, repeatability, human-device team
performance, robustness, and configuration control, and clearly state how benchmark evidence
affects established analytical and clinical evidence requirements. FDA should leverage existing
medical-device and AI standards where appropriate but should avoid creating a separate GenAI
benchmarking framework that is disconnected from established SaMD evaluation, risk
management, software lifecycle, usability, and clinical-validation principles.

The framework should reward transferable and generalizable clinical performance rather than
optimization to a fixed evaluation. Because FDA’s approach will influence how manufacturers
structure model development, testing, and preclinical validation, it should not inadvertently
incentivize development programs focused principally on maximizing a predetermined
benchmark score.

10) CDRH recognizes that publicly available benchmarking assets may be subject to data
contamination, saturation, and limited real-world representativeness. How should a
sponsor establish that performance on a given benchmark predicts safe and effective
real-world behavior for the device’s intended use? What evidence should support the
construct validity of a benchmark used to gate device evaluation, and what role should
sponsor-developed benchmarks play given potential concerns around independence
and optimization to the test?

A benchmark should be used to gate device evaluation only when its intended regulatory
purpose is clearly defined and evidence demonstrates that it validly measures a clinically
relevant construct and reliably predicts the aspect of real-world performance it is intended to
address. Public availability, widespread use, expert endorsement, or technical difficulty should
not, by themselves, establish validity. Public benchmarks affected by possible contamination,
repeated optimization, saturation, or limited clinical representativeness should receive limited
regulatory weight.

Benchmark performance should be evaluated not only by the score attained, but also by
whether performance transfers to independent datasets, previously unseen scenarios, diverse
patient populations, clinically representative workflows, and actual conditions of use. Where
feasible, validation should include independent and sequestered datasets that were unavailable
during model development. Clinically representative real-world evidence datasets may provide
stronger evidence of generalizability than performance on a public benchmark alone.

Sponsor-developed benchmarks should be permitted and may be necessary for intended-usespecific evaluation, provided that their development is prespecified, final test assets are
sequestered, clinical adjudication and scoring are appropriately independent, access is
controlled, and the benchmark is empirically linked to performance under clinically
representative conditions. FDA should focus on the independence and scientific integrity of the
evaluation process rather than requiring that all benchmarks originate outside the sponsor.

FDA should also apply the lessons from the Software Precertification Pilot Program before using
benchmark results to determine regulatory evidence requirements. The Agency identified 250
KPIs and metrics during Pre-Cert, grouped into nine categories, but the final report indicated
that appraisal outputs developed without prespecified criteria and key performance indicators
were less specific than necessary to inform regulatory decisions. FDA further identified the need
for clearly defined assessment elements, standardized structured data, consistent and
repeatable methods, appropriately trained assessment teams, and reliable outputs. The pilot
also demonstrated the difficulty of establishing that structured measures could replace or reduce
product-specific evidence.

Accordingly, FDA should not use a GenAI benchmark as a regulatory gate until the benchmark’s
constructs, scoring methodology, acceptance criteria, predictive relationship to clinical
performance, and regulatory consequences have been prospectively defined and validated. The
Agency should also clearly identify what subsequent evidence a successful benchmark would
replace, reduce, or merely supplement. Otherwise, benchmarking risks becoming an additional
evidentiary layer without improving regulatory predictability, efficiency, or assurance of safety
and effectiveness.

11) CDRH is considering that clinical confirmation for a GenAI-enabled device might not
require a prospective clinical study in every case, and has described above a range of
approaches of increasing rigor and patient exposure. How might a sponsor select and
justify a confirmation approach tailored to a device’s intended use and proportionate to
the device’s risk profile? Are there device types or risk profiles for which one or more of
these approaches would be insufficient or inappropriate? How might the anticipated
distribution of real-world inputs be taken into consideration? Are there other methods of
clinical confirmation that might help inform the evaluation of GenAI-enabled devices?

Danaher agrees that clinical confirmation should not be required for every GenAI-enabled
device. A sponsor should select the least burdensome confirmation method, or combination of
methods, capable of resolving the specific uncertainties remaining after nonclinical evaluation.
The justification should consider intended use, clinical claims, severity and reversibility of harm,
detectability of erroneous output, effectiveness of human oversight, novelty, workflow
dependence, anticipated real-world input distribution, and whether safety and effectiveness
depend on user behavior or human-device team performance.

Retrospective evaluation may be sufficient for a bounded function with representative realpatient inputs, a reliable reference standard, limited workflow dependence, and adequately
controlled residual risk. Shadow deployment may establish performance under live input and
integration conditions, and standardized-patient and simulated-use methods may be particularly
useful for communication, multi-turn interaction, escalation, and rare hazardous scenarios.
Prospective clinical evidence is more likely to be necessary for autonomous or agentic
functions, high-consequence decisions, claims dependent on changes in user behavior, and
situations where errors cannot be independently detected or reversed.

In order to ensure a robust and useful framework, it will be important to clearly define the
scientific questions answered by each confirmation method, the criteria for progression between
methods, what evidence a selected approach replaces or supplements, and which uncertainties
may appropriately be deferred to postmarket monitoring. Otherwise, the proposed range of
approaches may produce case-by-case evidentiary expectations rather than a predictable,
repeatable, and least-burdensome pathway.
12) How should sponsors achieve statistically meaningful performance measurement for
GenAI-enabled devices? Where synthetically generated inputs supplement real patient
data, how should sponsors account for differences between the synthetic and real-world
distributions when estimating performance, and under what conditions, if any, is it
appropriate to combine benchmarking evidence and clinical confirmation evidence to
support a single performance estimate?

Statistically meaningful performance measurement for a GenAI-enabled device should begin
with a prospectively defined intended use, unit of analysis, clinically justified acceptance criteria,
and sampling and analysis plan. The design should address variable outputs, repeated
interactions, clinically distinct error types, relevant subgroups, rare but serious failure modes,
and the anticipated real-world input distribution.

Synthetic inputs may appropriately supplement real patient data for challenge testing, rare-event
enrichment, controlled evaluation of specific variables, robustness assessment, and exploration
of hazardous failure modes. Their principal value may be in expanding test coverage, rather
than estimating the frequency of success or failure in clinical practice. If synthetic data
contribute to a population performance estimate, the sponsor should demonstrate that the
generating process preserves the clinically relevant joint distribution and relationships required
for that estimate, account for differences from the target-use population, and characterize
uncertainty introduced by transport or weighting.

Performance estimates derived from synthetic inputs and real patient data should initially be
reported separately. Benchmarking and clinical-confirmation evidence should support a single
estimate only when the evidence sources address comparable clinical conditions, performance
is consistent across those sources, appropriate statistical methods account for distributional
differences, and pooling does not obscure clinically meaningful variation or subgroup disparities.

FDA’s consideration of synthetic data and simulation may also affect how evidence is generated
during research, development, and preclinical validation. FDA should clarify when synthetic
molecular data and simulation-based evidence may supplement traditional testing, particularly
where biological or mechanistic relationships can be modeled with high confidence. These
considerations are consistent with the broader development of scientifically validated in silico
methods and are not unique to GenAI-enabled devices.

Synthetic data and simulation may also contribute to device evaluation beyond GenAI. For
example, synthetic or augmented digital pathology images may help evaluate robustness to
staining, scanner, and artifact variation, and simulation may supplement certain analytical
studies where the underlying relationships are well characterized. FDA should clarify when such
methods may supplement traditional testing. Because these considerations are not unique to
GenAI-enabled devices, they would benefit from a consistent, cross-cutting policy.
13) For which clinical domains, device functions, or subpopulations is synthetic data
particularly well-suited, or particularly inadequate, as a supplement to real-world
evidence? What safeguards would mitigate the risk that synthetic data generated by
models of the same class as the device under evaluation reproduces the very
performance gaps the evaluation is intended to detect, particularly for underrepresented
subgroups?

Synthetic data are particularly well suited for rare-event enrichment, expanding test coverage,
safety challenge sets, and exploration of known hazardous failure modes. In these uses,
synthetic data can expand test coverage and improve failure detection without being interpreted
as an estimate of the frequency of successful or unsuccessful performance in clinical practice.

Synthetic data may be particularly valuable for mechanistic and molecular applications,
including small-molecule and protein-structure prediction, molecular-interaction modeling, and
biological systems in which the relevant relationships are well characterized and amenable to
simulation. Advances in these areas may support algorithm development, testing, and
preclinical validation.

Synthetic data are less suitable for evaluating behavior that depends materially on real users
and workflows, or assessing novel or poorly characterized conditions.

Danaher is concerned that synthetic patient populations and virtual patient avatars may be
limited when generated from historically biased or unrepresentative clinical datasets. For
example, where source clinical datasets underrepresent particular demographic groups, disease
subtypes, or specimen types, those gaps may be reproduced or amplified in synthetic
populations rather than corrected. As a result, patient-avatar approaches may be less reliable
for detecting performance differences across demographic subgroups than synthetic data used
for targeted challenge testing.

Sponsors should validate synthetic data against independent real-world clinical evidence,
assess performance separately for relevant demographic and clinical subgroups, and
demonstrate that synthetic data do not mask, dilute, or artificially reduce performance
disparities. When synthetic data are generated using the same or a similar model class as the
evaluated device, independent testing should specifically assess shared assumptions, biases,
and failure modes.

Synthetic data should supplement, rather than replace, real-world evidence when evaluating
patient-level or demographically driven performance. FDA should distinguish these uses from
targeted challenge testing and well-characterized technical variation (e.g., image acquisition
variability), where synthetic data may offer a more mature and reliable evidentiary contribution.

Danaher encourages FDA to coordinate with other federal agencies and stakeholders regarding
validated uses of synthetic data and simulation in other sectors, while ensuring that lessons
from those sectors are appropriately adapted to the clinical and regulatory context of medical
devices.

14) For open-ended device outputs, how should performance comparators and
acceptance criteria be selected? When a panel of qualified clinicians serves as the
comparator, how should the applicable standard (for example, the standard of care
versus the performance of a median clinician in practice) be defined and justified?
Should generalist or specialist physicians be used as a performance standard? When
should human-AI team performance, rather than the device operating alone, serve as the
basis for evaluation?

For open-ended outputs, the comparator and acceptance criteria should be selected based on
the device’s intended use, clinical claims, users, use environment, and consequences of error.
Acceptance criteria should be clinically justified and risk-based, with safety-critical criteria
evaluated separately rather than combined into a compensatory aggregate score.

The appropriate comparator should reflect the level of performance necessary for safe and
effective use, rather than defaulting to the performance of a median clinician. Median
performance may normalize common but unacceptable practice variation and may not represent
the standard of care. Conversely, requiring performance equivalent to an expert specialist may
be unnecessarily burdensome where the device is intended to support a generalist performing a
generalist-level function. Sponsors should prospectively define the applicable clinical standard
and justify it based on the intended function, intended users, clinical setting, available
authoritative guidance, and consequences of disagreement or error.

Clinician panels may be appropriate where no objective reference standard exists. Sponsors
should justify panel composition, adjudicator qualifications, information available to reviewers,
scoring criteria, handling of multiple acceptable answers, disagreement-resolution procedures,
and inter-rater reliability. Specialists should be used when the intended function requires
specialist knowledge or the reference determination ordinarily requires a specialist. Generalists
may be appropriate when the function and intended users are generalist in nature. Mixed panels
may be useful where the device bridges levels of care, provided that the role of each reviewer
group is defined and results are not combined in a manner that obscures clinically meaningful
disagreement.

Human-AI team performance should be an important consideration when the intended use
depends on a user interpreting, verifying, or acting upon the device output, with the extent of
such testing proportionate to risk. Where a device’s intended use and claimed benefit depend
on improving the performance of the healthcare professional using the device, human-device
team performance should be evaluated in a manner proportionate to risk and consistent with the
marketed use. For many lower-risk functions, usability evaluation and device-only performance
may be sufficient; reader or user studies should be reserved for functions likely to materially
change clinical decisions. Where such studies are warranted, device-only performance does not
fully characterize safety and effectiveness because the device may improve or degrade user
decisions through clarification, automation bias, inappropriate confidence, or changes in
workflow. Device-only performance should still be evaluated to characterize intrinsic failure
modes, but the primary clinical assessment should reflect the marketed use. For autonomous or
action-taking functions, the complete device should be evaluated independently, including
action selection, execution, traceability, reversibility, and failure handling.

15) Are there ways in which performance might be assessed relative to the care,
technology, or course of action likely to occur in the absence of the device, rather than
to the comparators described in this section? What approaches might be used to identify
and justify a comparator such as unaided clinical judgment, delayed specialist review, or
no intervention?

For many GenAI-enabled devices, performance may be more appropriately assessed relative to
the care, technology, or course of action likely to occur in the absence of the device, rather than
against a clinical reference. The benefit-risk profile of a device depends not only on absolute
performance, but also on whether it improves upon the real-world alternative available to the
intended user in the intended use setting.

The appropriate comparator should therefore be selected prospectively based on the intended
use, intended users, clinical setting, and current standard of practice. Potential comparators
may include unaided clinical judgment, review by a generalist or specialist clinician, delayed
specialist consultation, an existing software tool, current workflow processes, or no intervention
where that reflects actual practice. The comparator should be justified using objective evidence
that it represents the course of action reasonably expected to occur in the absence of the
device.

16) Should independent third parties be involved in some or all aspects of a competencybased approach, including device benchmarking and clinical confirmation? If so, in what
ways might qualified, independent third-party participation contribute to this approach,
and what qualifications and independence criteria should apply? Are there aspects of
the assessment for which third-party involvement would be impractical or inadvisable?
What safeguards or program design features would be critical to prevent third-party
participation from limiting competition or preventing innovation?

Independent third parties may contribute to certain aspects of a competency-based framework,
particularly where greater confidence in benchmark integrity, clinical adjudication, or evaluation
independence is needed. However, third-party involvement should not be presumed necessary
for all devices or all stages of evaluation. FDA should first establish that a given assessment
method is scientifically valid and reproducible; third-party participation alone does not establish
validity.
Third parties may be most useful for activities such as development and maintenance of
benchmark assets, oversight of sequestered test datasets, independent execution of predefined
evaluations, clinical adjudication of open-ended outputs, and verification that evaluation
protocols were conducted as specified. In these roles, third parties may help strengthen
transparency, consistency, confidence in the independence of evaluation results, and
comparability across submissions.

Independent third parties may be particularly valuable as stewards of clinically representative
real-world datasets, sequestered benchmark resources, and external evaluation programs
designed to assess device performance under realistic conditions. These arrangements may
provide stronger evidence that benchmark results reflect generalizable clinical capability rather
than optimization to a sponsor-controlled test.

Any third party involved in device evaluation should possess appropriate technical, clinical,
statistical, and regulatory expertise relevant to the device under review. Independence criteria
should be tailored to the third party’s role. A third party that executes testing or adjudicates
results should not have participated in developing the evaluated device, its training data, or its
commercialization. A third party that develops or maintains benchmark assets should operate
under transparent governance and should not participate in developing or commercializing the
devices evaluated against those assets. Methodologies, governance, and potential conflicts of
interest should be disclosed.

Third-party involvement may be impractical or inadvisable when evaluation depends on
proprietary data, rapidly evolving device configurations, specialized domain expertise not
broadly available, or highly device-specific intended uses that a centralized framework cannot
reasonably assess. Sponsors should retain primary responsibility for demonstrating safety and
effectiveness, and third-party participation should not be used to transfer regulatory
accountability.

FDA should preserve multiple qualified evaluation pathways and should not allow any single
third party, benchmark provider, or dataset curator to become a required gatekeeper for market
access.

17) CDRH is exploring whether a competency-based approach could be applied to
devices that incorporate a variety of underlying model architectures, such as multimodal
vision-language models and generative or predictive world models. Are there aspects of
the described approach that might be ineffective or inapplicable to these devices? What
additional or distinct considerations might need to be addressed?

A competency-based approach may be applied across different model architectures, but the
competency domains, test methods, and acceptance criteria should be tailored to the intended
use and the complete user-facing device rather than applied uniformly across architectures. A
common high-level framework may be useful, but architecture-specific testing will be necessary
where different technologies introduce distinct capabilities, dependencies, and failure modes.

For multimodal models, evaluation should address performance within each modality,
integration across modalities, conflicting or missing inputs, data alignment, and whether an error
in one modality is propagated or amplified in the final output. For generative or predictive world
models, evaluation should assess whether modeled scenarios remain clinically plausible,
whether uncertainty increases appropriately as predictions extend beyond observed information,
and whether the device safely handles conditions outside the validated operating range.

FDA should avoid assuming that a common set of competency scores will have the same
meaning across model architectures. Rather, an architecture-neutral framework should
promote consistency without imposing uniform tests that may be ineffective, irrelevant, or
misleading for a particular technology.

18) CDRH is considering whether it may be appropriate to accept greater premarket
uncertainty regarding a GenAI-enabled device’s benefit-risk profile through greater
reliance on postmarket monitoring. Under what conditions might such an approach be
appropriate, and what characteristics of a monitoring program would need to be in place
to justify reduced premarket evidence? Are there device types or risk profiles for which
this approach would not be appropriate?

Greater reliance on postmarket evidence may be appropriate where (1) the residual premarket
uncertainty is clearly identified and bounded; (2) potential harms are detectable and reversible
and the device operates with effective human review; (3) the manufacturer has a prespecified
monitoring plan with defined performance metrics, data sources, sample sizes, and action
thresholds; and (4) the manufacturer can promptly restrict, roll back, or correct the device if
thresholds are crossed.

A monitoring program that supports reduced premarket evidence should be described in the
marketing submission, use clinically justified acceptance criteria, rely on representative realworld data, and include defined timelines for reporting to FDA. FDA should also identify the
regulatory mechanism, such as special controls, conditions of approval, or postmarket study
requirements, so that expectations are transparent and consistently applied.

This approach would generally not be appropriate for autonomous or agentic functions that act
without human review, for outputs that drive irreversible or time-critical clinical decisions, or
where errors are unlikely to be detected in routine use.

19) Please comment on the potential approaches to postmarket performance evaluation,
including periodic re-benchmarking, sample-based clinician review, and performance
degradation monitoring. What additional approaches should CDRH consider, and how
should the cadence and triggering events for reassessment be determined?
Periodic re-benchmarking, sample-based clinician review, and performance degradation
monitoring are useful and complementary. Their selection and frequency should be risk-based
and tied to the hazards identified in the device’s risk management file. Additional approaches
FDA should consider include monitoring for input-distribution drift (e.g., changes in patient
population, specimen types, instrumentation, or user prompting behavior); tracking user
overrides, corrections, and complaint trends; and, for IVD-associated functions, correlating
GenAI outputs with downstream confirmatory test results where such data are available.

Reassessment should occur on a prespecified cadence proportionate to risk and should also be
triggered by defined events, including changes to the underlying foundation model, significant
shifts in the input distribution, performance falling below prespecified thresholds, new safety
signals, or relevant changes in clinical practice or guidelines. These activities should be
integrated with existing Quality Management System Regulation, complaint handling, and
medical device reporting obligations rather than established as a separate regime.

20) Please comment on whether the proposed approaches to postmarket monitoring can
be facilitated by machine-based supervisory agents. What considerations, including the
evaluation and reliability of the supervisory agent itself, should CDRH take into account
for such an approach?

Machine-based supervisory agents may usefully extend the scale and frequency of postmarket
monitoring, for example by screening large volumes of outputs for potential errors, out-of-scope
behavior, or drift, and routing flagged cases for human review. However, a supervisory agent
should be treated as a monitoring tool whose own performance is validated for its intended
purpose, not as a substitute for human oversight or manufacturer accountability.

Relevant considerations include validation of the agent’s sensitivity and specificity for the failure
modes it is intended to detect; independence from the monitored device, including avoiding use
of the same or a closely related model that may share blind spots; periodic human audit of both
flagged and unflagged outputs; version control and change management for the agent itself;
and data privacy and security controls. Where a supervisory agent is used only within the
manufacturer’s quality system, it should be validated and controlled as quality system software
rather than reviewed as a medical device.

21) What roles might clinicians, healthcare institutions, professional societies,
standards-setting bodies, and other stakeholders appropriately play in postmarket
monitoring, and how can these roles be structured without diffusing manufacturer
accountability?

Clinicians, healthcare institutions, and clinical laboratories are well positioned to identify realworld performance issues, and professional societies and standards-setting bodies can
contribute clinical context, reference datasets, and consensus performance metrics. These roles
can be structured without diffusing accountability by keeping the manufacturer responsible for
the monitoring plan, signal evaluation, corrective action, and regulatory reporting, while defining
stakeholder contributions through mechanisms such as feedback channels, data-sharing
agreements, and participation in registries and standards development.

FDA should recognize that healthcare institutions and laboratories have finite resources.
Monitoring programs should not depend on uncompensated or unreasonably burdensome site
activities, and any expectations of deploying institutions should be clearly described in labeling.
Voluntary real-world performance collaborations may also help FDA and manufacturers detect
class-wide issues affecting devices built on common foundation models.

22) CDRH envisions that a manufacturer’s premarket competency-based assessment
might serve as a baseline against which post-deployment modifications could be reevaluated. How might the extent of re-benchmarking or other evidence be scaled to the
nature and expected impact of a given modification? For example, are there categories
of change that might not significantly affect safety or effectiveness of a GenAI-enabled
device, or might be appropriately managed within a sponsor’s quality management
system versus requiring FDA premarket review and authorization before
implementation? Are there categories of changes that might be appropriate for inclusion
in a PCCP?

Danaher supports using the premarket competency-based assessment as a baseline for
evaluating modifications, with evidence scaled to the nature and expected impact of the change.
Consistent with FDA’s existing guidance on deciding when to submit a 510(k) for a software
change to an existing device, changes that do not affect intended use, clinical performance, or
risk controls should be managed within the manufacturer’s quality management system with
documented verification. Examples may include user-interface and formatting changes that do
not alter clinical content; infrastructure, security, and performance-optimization changes; and
minor prompt or guardrail refinements verified against the baseline regression set.

Changes appropriate for inclusion in a PCCP may include foundation-model version updates
within the same model family; retraining or fine-tuning with additional data under a prespecified
protocol; prompt, retrieval, and guardrail modifications; updates to reference content within a
validated retrieval architecture; and expansion to additional validated input types, each with
predefined re-benchmarking methods and acceptance criteria. Changes that alter intended use,
expand to new patient populations or clinical indications, increase autonomy, or move a function
to a higher-risk position on either axis should continue to require FDA review.

23) How might PCCP concepts or other change-control approaches be adapted for
GenAI-enabled devices when the nature or scope of future modifications cannot be fully
prespecified?

Where future modifications cannot be fully prespecified, PCCPs could be adapted by focusing
on the boundaries and methods of change rather than enumerating each specific modification.
A GenAI PCCP might define (1) the performance envelope, meaning the clinical functions, user
groups, and risk positions within which changes may occur; (2) a standardized regression and
re-benchmarking protocol, including sequestered test sets and acceptance criteria tied to the
baseline; (3) objective triggers that would place a change outside the PCCP; and (4)
transparency obligations, such as updated labeling and change logs.

FDA should also clarify how a PCCP may address upstream foundation-model changes that the
manufacturer did not initiate (see Q24). Danaher encourages FDA to implement these concepts
through its existing PCCP guidance and the PCCP guiding principles developed with Health
Canada and the MHRA, and to work with industry on illustrative GenAI PCCP examples.

24) For devices built on third-party foundation models, changes to the underlying model
may be initiated by the third-party model developer rather than the device manufacturer.
How can a manufacturer detect, evaluate, and respond to such changes in a timely
manner? What mechanisms—for example, contractual, technical, or through a PCCP—
could provide reasonable assurance that third-party developer-initiated changes do not
compromise the safety or effectiveness of the device?

Manufacturers should remain responsible for the safety and effectiveness of the complete
device, including where the device relies on a third-party foundation model. Manufacturers
should maintain version identification, change detection, performance monitoring, and
regression testing sufficient to determine whether an upstream change affects the device’s
intended use, clinical performance, or risk controls.

Where feasible, contractual mechanisms should require advance notification of material
changes, continued access to validated model versions, information necessary for impact
assessment, and transition support. Technical controls should prevent the automatic
deployment of an unreviewed model version into a regulated device.

Where the manufacturer maintains a validated, version-pinned instance of a foundation model
under its control, subsequent changes made by the third-party model developer that are not
incorporated into the deployed device should not themselves trigger additional regulatory
obligations. Regulatory assessment should be tied to changes implemented in the marketed
device, rather than changes that occur externally but are not adopted.

A Predetermined Change Control Plan may address anticipated categories of foundation-model
changes when the applicable assessment methods, acceptance criteria, and implementation
controls can be defined in advance. Changes outside those boundaries should undergo
appropriate quality-system and regulatory assessment before deployment.

Reliance on proprietary third-party models creates lifecycle risks similar to dependency on
unsupported or force-migrated software platforms. Models offering greater transparency,
version stability, reproducibility, and manufacturer control, including appropriately governed
open-weight models, may reduce dependency risk.

Manufacturers should maintain contingency plans for model discontinuation, forced migration, or
loss of access. Any replacement or “hot-swapped” model should undergo independent
validation before clinical use; technical interchangeability should not be presumed to establish
equivalent clinical performance.

25) Would voluntary Foundation Model MAFs be practical for foundation model
developers to provide and to keep current, and what content might they include? Given
that participation would be voluntary and that model developers may have limited
incentive to disclose safety-relevant information, what would make such a program
sufficiently useful for premarket review? Are there alternative mechanisms CDRH should
consider for obtaining information about underlying foundation models?

Voluntary Foundation Model Master Files may be useful if they contain standardized, current,
and sufficiently detailed information for FDA and device manufacturers to assess material risks.
Relevant content may include model architecture and intended limitations, training and
evaluation methods, known failure modes, version history, update practices, change-notification
procedures, security controls, bias and robustness evaluations, and other information necessary
to assess material downstream effects.

However, a voluntary program may not produce complete or consistently current information
where model developers lack regulatory obligations or commercial incentives to participate.
FDA should establish minimum content and maintenance expectations before relying on a
Master File during premarket review.

FDA should assess whether voluntary disclosure alone can provide adequate transparency,
version stability, reproducibility, and access to safety-relevant information. Where it cannot,
alternative evidence may include manufacturer-conducted evaluations, independent testing,
sequestered benchmark datasets, contractual access to material change information, and
clinically representative performance testing.

Greater model transparency and manufacturer control may affect the risks associated with a
particular foundation model. However, use of an open-weight or otherwise transparent model
should not reduce the manufacturer’s obligation to validate the configured medical device for its
intended clinical use.

A Foundation Model Master File should supplement, not replace, evaluation of the configured
device. Regardless of the underlying model, the device manufacturer should remain responsible
for safety, effectiveness, version control, and change management.
26) Are there additional considerations that inform the premarket and postmarket
evaluation of agentic GenAI-enabled devices, beyond those applicable to non-agentic
GenAI-enabled devices? How should the elevated risk associated with autonomous
multi-step action, tool use, and reduced opportunity for human review be reflected in
acceptance criteria and oversight?

Agentic GenAI-enabled devices raise additional considerations because they may execute
multi-step actions, invoke external tools or systems, and act with reduced opportunity for human
review. In addition to the considerations applicable to non-agentic devices, evaluation should
address the scope of actions the device is permitted to take and the technical controls that
enforce that scope; the correctness of each step, not only the final output, including tool
selection and execution; error propagation across steps; behavior when tools, data sources, or
connected systems fail or return unexpected results; the ability to reconstruct and audit each
action; and the reversibility of actions and availability of human intervention points.

Acceptance criteria should reflect this elevated risk, for example through separately evaluated
safety-critical criteria without compensatory scoring, stricter thresholds for irreversible actions,
and validated fail-safe behavior. Where an agentic function interacts with laboratory or hospital
information systems, cybersecurity and interoperability should be addressed consistent with
FDA’s existing cybersecurity guidance. Postmarket oversight should include action-level logging
and monitoring sufficient to detect unsafe or out-of-scope actions promptly. FDA should place
agentic functions toward the higher end of the Activity axis by default, while allowing sponsors to
justify a lower placement where effective human checkpoints are demonstrated.

Conclusion

Danaher encourages FDA to continue coordinating with IMDRF and other regulators as it
develops its approach. Many GenAI-enabled devices will be marketed globally and will also be
subject to the EU AI Act and the EU MDR and IVDR. Alignment on terminology, risk
categorization, evaluation methods, and change-control expectations would reduce duplicative
evidence generation and support timely patient access.

We welcome continued engagement with FDA on these important issues. If you have any
questions, please contact Ronald Dunn at ronald.dunn@leicabiosystems.com.

Sincerely,

Ronald D. Dunn
Senior Vice President, DHR Dx Global Regulatory Strategy
Danaher Corporation
Angela Stewart
Vice President, Government Affairs, Diagnostics Businesses
Danaher Corporation

About Danaher Corporation

Danaher is a leading global life sciences and diagnostics innovator, committed to accelerating
the power of science and technology to improve human health. Our businesses partner closely
with customers to solve many of the most important health challenges impacting patients
around the world. Danaher's advanced science and technology, and proven ability to innovate,
help enable faster, more accurate diagnoses and help reduce the time and cost needed to
sustainably discover, develop and deliver life-changing therapies. Focused on scientific
excellence, innovation and continuous improvement, our approximately 60,000 associates
worldwide help ensure that Danaher is improving quality of life for billions of people today, while
setting the foundation for a healthier, more sustainable tomorrow.

Our Diagnostics businesses build upon a foundation of quality and efficiency to give patients
access to high-quality, dependable diagnostic capabilities. With a data insight foundation for
established use-cases, we are a global leader in accelerating innovation. Our team of over
23,000 Diagnostics associates around the globe contribute to advancing life-changing science
and technology daily, so life-saving diagnostic healthcare is accessible to patients around the
world. Danaher’s diagnostic companies deploy tests in more than 20,000 hospitals, perform
1.6M tests per hour, and monitor more than 200 million patients each year to help clinicians
make critical healthcare decisions.