Society of Diagnostic Medical Sonography

The Society of Diagnostic Medical Sonography (SDMS) submits the attached comments in response to FDA’s discussion paper, “Considerations for the Regulation of Generative AI-Enabled Medical Devices.” Please see the attached PDF for SDMS’ complete responses and recommendations concerning diagnostic ultrasound acquisition, examination completeness, user competence, and patient safety.

FDA-2026-N-7874-0001
Society of Diagnostic Medical Sonography, diagnostic medical sonographers, and
generative AI-enabled medical devices
The Society of Diagnostic Medical Sonography (SDMS) appreciates the opportunity to respond to
FDA’s discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical
Devices. SDMS provides advocacy and education for diagnostic medical sonographers
(sonographers). Our comments address Questions 1, 2, 4, 9, 11, 14, 15, 18, 19, and 21 and focus
on the implications of generative artificial intelligence (GenAI)-enabled devices for diagnostic
ultrasound acquisition, examination completeness, clinical communication, and patient safety. [1]
SDMS supports the development of GenAI-enabled medical devices that improve diagnostic quality,
support sonographers, reduce unnecessary workload, and expand access to appropriate care.
Evaluation should account for how a device changes both the examination and the performance of
its intended users. A clinically useful output on an existing image does not establish that a GenAIassisted ultrasound examination will acquire all the information needed to answer the clinical
question.
Sonography is highly operator-dependent. Sonographers assess relevant clinical information,
physically manipulate the ultrasound transducer, adjust equipment settings, and acquire still images,
motion clips, measurements, and Doppler information as appropriate. They also exercise clinical
judgment to adapt examinations as findings emerge.
Sonographers provide the acquired images and data along with a technical report describing
findings and examination limitations to the interpreting physician. The sonographer’s report supports
physician interpretation and does not constitute a legal diagnosis. The sonographer’s responsibilities
and the expectations for education, training, competence, and relevant specialty certification are
described in the Scope of Practice and Clinical Standards for the Diagnostic Medical Sonographer.
[2] An acquisition error may prevent clinically necessary information from being collected, leaving
little evidence of the omission in the saved examination and potentially contributing to
misinterpretation or delayed care.
GenAI assistance should preserve the sonographer’s ability to exercise clinical judgment, adapt the
examination, and communicate findings and limitations in accordance with applicable requirements
and facility policies. Manufacturers should demonstrate that their devices meet clinically justified
requirements for examination quality and patient safety.
These comments concern GenAI-enabled medical device functions used in sonography and their
interactions with other functions in a product, whether integrated into ultrasound equipment or
provided through separate software. Automated measurements, image highlighting, and acquisition
guidance can also use conventional AI or other software. They are addressed here when generated,
directed, or materially influenced by a GenAI-enabled function.
The clinical scenarios described below illustrate foreseeable hazards for evaluation. They do not
describe documented adverse events or establish their frequency.
Question 1
SDMS supports FDA’s two-axis framework, which considers what the device does, how
independently it acts, and the potential harm from relying on an incorrect output. FDA should provide
sonography-specific examples showing how GenAI-enabled functions may influence image
acquisition, measurements, examination completeness, and the sonographer’s technical report
provided to the interpreting physician.

Page 1 of 14
The consequences of an error depend on the clinical situation and the decisions that follow. For
example, an incorrect output during a fetal or cardiac examination could contribute to a missed
finding, misleading reassurance, unnecessary further testing, or delayed care.
FDA should also consider whether the intended user has the training, experience, time, and access
to source images and data needed to recognize and correct an error before it affects care. A
complete ultrasound examination requires appropriate views, accurate measurements, and
adequate recording of relevant motion when indicated. If GenAI guidance causes necessary
information to go uncollected, a later review of the saved examination may not reveal the omission.
Risk may increase when users lack the relevant sonography expertise to check the guidance and
recognize examination limitations. Manufacturers should demonstrate that human review works
under actual scanning conditions.
We recommend that manufacturers of GenAI-enabled medical devices used in sonography address
the following questions for each function and each important way it could fail:
• Can the intended user recognize an incorrect or misleading output using the images, motion
clips, signals, and clinical information available during the examination?
• Can the intended user correct the GenAI-enabled function error before it affects care and while
the patient is still available? A recall or repeated examination can cause delay, extra cost, or
failure to return for needed imaging.
• Could the GenAI-enabled function cause a required image, measurement, Doppler assessment,
or further investigation to be skipped? A post-examination review cannot recover diagnostic
information that was not collected initially.
• Does the intended user have the relevant specialty expertise, time, and attention to check the
output during the examination? Manufacturers should demonstrate how reliably users recognize
and correct clinically important errors under actual scanning conditions.
• Are independent safety checks available when they are needed? Can the intended user trace
each clinically important output to the correct patient and relevant source images, signals,
measurements, or clinical references, as applicable, and identify content created by GenAI?
• How often could a harmful error occur across repeated uses, patient groups, and working
conditions? Even an uncommon error can affect many patients when a GenAI-enabled function
is widely used.
For example, an AI assistant might accept an inadequate fetal cardiac view, make anatomical
measurements, and mark the fetal examination complete. The sonographer or interpreting physician
might not recognize the error until after the patient has left. Additional imaging could be needed,
potentially delaying diagnosis or care. FDA's assessment should account for the missed acquisition,
the chance of detecting it, and the chance of correcting it in time.
SDMS recommends that FDA retain the two-axis diagram and require a structured assessment
alongside it for each device function and important way it could fail. That assessment should
address whether errors can be detected and corrected before affecting care, whether independent
safeguards are available, whether users have sufficient expertise and time to review outputs,
whether outputs can be traced to their sources, and how often harmful errors could occur.
These factors should inform the evidence and safeguards expected for the function. Errors that are
difficult to detect or correct, or that prevent necessary information from being collected, should
require stronger testing and safety measures proportionate to the potential harm. Manufacturers

Page 2 of 14
should provide evidence that claimed safeguards reduce risk under actual scanning conditions for
each intended user group, with the training and experience specified for the device.
Question 2

SDMS recommends assessing how a GenAI output is likely to influence the intended user. This
assessment should account for the device’s intended use and overall function. The wording matters,
but so do specificity, patient-specific content, screen display, timing, default choices, and how the
output enters the user’s workflow. An output can steer a decision without explicitly telling the user
what to do.
Examples include suggested next views, ranked choices, highlighted anatomy, a confidence display,
or an “exam complete” message. A suggested measurement or draft finding automatically entered
into a worksheet can become the easiest choice to accept. A warning that keeps reappearing, or an
output that takes several steps to reject, can also influence behavior. FDA’s transparency principles
support giving users information that fits their tasks and working conditions. [3]
For example, a GenAI-enabled function might enter “normal” into a draft worksheet for review even
though a structure has not been adequately visualized. That entry could discourage further scanning
or be accepted and communicated to the interpreting physician. A green completion indicator could
have a similar effect, even if a disclaimer says the user remains responsible.
Manufacturers should test whether the device clearly distinguishes among anatomy that is
adequately visualized with no abnormality identified, anatomy with limited visualization, and anatomy
that is not visualized. These distinctions should remain clear when information is transferred into
worksheets, draft reports, or other systems. Limited visualization or non-visualization should not be
presented as a normal finding.
FDA should provide manufacturers with sonography-specific examples showing a range of
influence, from general information to patient-specific recommendations and direct instructions. The
examples should illustrate how presentation and workflow can make an output more directive.
Manufacturers should explain:
• What action, including stopping or skipping further imaging, each clinically important output is
likely to encourage.
• How patient-specific wording, placement on the screen, claims of certainty, default choices,
and automatic transfer into other systems may affect reliance on the output.
• How easily the intended user can check, correct, reject, or bypass the output.
• Whether a conversation or series of outputs becomes more directive or influential as it
continues.
Manufacturers should test how these features affect the behavior of representative intended users,
with the training and experience specified for the device. Testing should include realistic scenarios
with incorrect or incomplete outputs and assess whether users recognize errors, continue necessary
imaging, and reject inappropriate guidance. FDA’s guidance on testing how people use devices
provides relevant methods. [4]
A confidence score should count as a safety measure only if testing shows that it meaningfully
reflects the reliability of the output for the intended task, that users understand its limits, and that its
presentation helps users respond appropriately. A reassuring display may otherwise increase
reliance on an incorrect output.

Page 3 of 14
FDA should assess these influences together with the potential harm from an incorrect output,
including when it encourages the user to stop collecting information. A disclaimer alone does not
demonstrate that risk has been reduced. Intended users need access to supporting information and
relevant limitations when making the decision.
Question 4
SDMS recommends that FDA include sonographers and other relevant healthcare professionals in
its discussion of generalist and specialist users. Risk depends on whether the intended user has the
knowledge and skills required for the specific task, including acquiring adequate images, assessing
examination completeness, reviewing AI outputs, and recognizing when additional imaging or
specialist help is needed. A professional title or experience in another specialty does not, by itself,
establish that competence.
Manufacturers of GenAI-enabled medical devices used in sonography should specify the education,
applicable credentials, specialty knowledge, clinical experience, device training, and task-specific
competence needed for each intended-user group. They should test whether representative
intended users can perform the task safely with the device, recognize relevant limitations, and
respond appropriately to important errors. Facilities should verify that individuals meet the applicable
qualifications and competency requirements for their assigned tasks. SDMS’ scope of practice and
clinical standards recognize the importance of individual competence and specialty preparation. [2]
Most U.S. states do not require sonographers to hold a state sonography license. [5] FDA should
distinguish licensure from sonography credentials and demonstrated task-specific competence. A
“licensed user” requirement alone could exclude qualified, credentialed sonographers in states that
do not license the profession. A license in another healthcare profession also does not, by itself,
establish competence to perform a particular ultrasound examination. [2]
For example, experience in abdominal sonography does not by itself establish competence to
perform pediatric echocardiography. A physician’s expertise in interpreting a condition also does not
necessarily establish the ability to acquire the ultrasound views needed to assess it.
Risk can increase when a user lacks the expertise to distinguish poor visualization from a normal
finding, recognize a measurement made in the wrong plane, identify findings that require further
assessment, or decide when specialist help is needed. A reassuring AI answer may discourage
additional scanning or referral when the user lacks the knowledge needed to challenge it. If
necessary information is not acquired, later review of saved images may not reveal the omission.
FDA should ask manufacturers to provide and test safeguards appropriate to the function and its
potential harms, including:
• Clear limits on the tasks, examinations, patient groups, equipment, and intended-user
qualifications covered by the device’s tested capabilities.
• Training followed by demonstration of task competence, including responses to poor images,
misleading AI outputs, and situations in which the AI fails.
• Access to the source images, measurements, signals, and other information needed to
check the output, as appropriate to the task.
• Clear messages that distinguish inadequate visualization from a normal finding, identify
missing information, and explain when the device cannot provide a reliable answer. Limited
assessments should state what they can and cannot determine.

Page 4 of 14
 • Clear instructions and practical arrangements for obtaining qualified assistance when the
task exceeds the user’s competence or the device’s tested capabilities. If safe use depends
on specialist review, that review should be available before the output affects care.
• A safe way to reject AI advice, stop the AI function, and document limitations. Plans for
device failure should account for whether the user is qualified to continue without AI or needs
assistance or alternative care.
• Where appropriate, controls that restrict specialized functions to authorized users whose
relevant qualifications and competence have been verified.
These safeguards should be tested with each intended-user group under realistic working
conditions, including time pressure, difficult examinations, and incorrect or incomplete AI outputs.
Testing should demonstrate that the safeguards reduce important errors. Clicking “reviewed” does
not establish that a user understood or independently checked an answer. When users cannot
independently assess an output, manufacturers should demonstrate that other safeguards can
prevent or detect consequential errors before they affect care.
For a narrowly defined ultrasound task, manufacturers should demonstrate safe performance with
the intended users and the training specified for the device. Intended users should have the
qualifications and procedure-specific competence required for that task and practice within their
applicable scope. AI assistance does not substitute for those qualifications, expand an individual’s
scope of practice, or establish competence to perform a broader diagnostic sonographic
examination.
Question 9
SDMS supports FDA’s broad benchmarking categories. For ultrasound, testing should address the
GenAI-enabled device function’s claimed role in image acquisition, examination completeness,
clinical communication, and patient safety. Manufacturers should evaluate the finished device as
intended to be used, including the AI model, instructions given to the AI, instructions provided to
users, reference material the AI consults, safety controls, screen displays, and connected tools or
equipment.
Manufacturers should document the quality, representativeness, and limitations of training data and
provide users with understandable summaries relevant to safe use. The volume of training data
alone does not establish clinical performance. Performance should be evaluated using test data
appropriately independent of the data used to train or tune the model. [6]

The following sonography-specific criteria should be applied according to the function’s intended use
and claims. Examination completeness and measurement accuracy, for example, involve both
clinical proficiency and safety. FDA should make these connections clear without treating the same
test result as independent evidence more than once.

What to test What the test should establish
Adequacy of Views are diagnostically adequate for the intended task, and poor
acquired views visualization is not treated as evidence of normal anatomy.
Examination The function recognizes missing information relevant to its claimed task
completeness and avoids falsely declaring the examination, or the portion it assesses,
complete.
Changes during Guidance appropriately accounts for the patient’s condition, limited
scanning acoustic windows, and findings that require adaptation of the examination
or additional help.

Page 5 of 14
 What to test What the test should establish
Continued Continued

Measurements Measurements use the correct anatomy, plane, units, caliper
placement, and acquisition conditions.
AI-created or altered Users can distinguish acquired information from content created or
content changed by AI.
User understanding and Users understand outputs, limitations, and uncertainty and can
decisions respond appropriately to incorrect or incomplete guidance.
Equipment or multi-step Actions remain within the device’s intended use, required human
actions, when applicable approval is respected, and failures of connected tools or
intermediate steps are handled safely.
Different patients and Performance meets clinically justified criteria across patient groups,
working conditions intended users, transducers, presets, ultrasound equipment, and
clinical settings covered by the device’s claims.

Manufacturers should define test methods, suitable reference assessments, and clinically justified
acceptance criteria before testing. Falsely declaring an examination complete and providing
reassurance without adequate patient-specific information should be assessed as distinct safety
failures. Strong performance in writing or image labeling should not obscure these failures. When
recognizing a particular finding falls outside the device’s intended use, the device should still avoid
unsupported reassurance and communicate its limitations.

Testing should include motion, artifacts, technically difficult examinations, incomplete or conflicting
histories, relevant uncommon findings, interruptions, and realistic time pressure. It should assess
interactions among the patient, operator, transducer, and device during an evolving examination.
Effects on actual scanning behavior and examination quality may also require clinical confirmation.

Manufacturers should repeat testing on the same cases and clinically equivalent descriptions to
assess whether changes in wording or information order produce unreliable conclusions about
visualization, examination completeness, or the need for additional help.

For functions that create or change diagnostic images, testing should assess whether important
anatomy or findings are invented, removed, or distorted compared with the acquired data. Users
should be able to review the patient’s original acquired images and relevant source data and identify
content created or modified by AI. A realistic-looking image alone does not establish accuracy.

Where relevant to the function, evaluation should assess patient comfort, examination duration,
acoustic exposure, and sonographer workload and ergonomics. Acoustic output, dwell time, and
scanning time should be kept as low as reasonably achievable (ALARA) while obtaining the required
diagnostic information. [2] Testing should assess whether reduced documentation time is offset by
repetitive scanning, sustained transducer pressure, awkward positioning, or longer examinations.
Ergonomic evaluation should be informed by the Industry Standards for the Prevention of Work
Related Musculoskeletal Disorders in Sonography. [7]

Useful resources include SDMS’ scope of practice and clinical standards, AIUM’s practice
parameters, FDA’s guidance on human factors and usability engineering, and the international Good
Machine Learning Practice principles. [2,4,6,8] These resources serve different purposes and do not
replace validation of the particular GenAI-enabled device and its intended use. Manufacturers
should identify the applicable standards and versions and explain how they support the evaluation.

Page 6 of 14
Question 11
SDMS recommends selecting and justifying clinical confirmation methods according to the device’s
intended use, intended users, and risk. Testing should demonstrate that the GenAI-enabled function
performs as claimed for the patients, equipment, and conditions covered by its intended use. More
serious errors, or errors that are harder to detect and correct, call for stronger evidence.
Past examinations may provide useful evidence for a limited function that works on images or data
already collected. For example, a tool that drafts documentation could be checked against the
complete source examination and relevant clinical information. Manufacturers should also test
whether the intended reviewers, with the required training and expertise, can identify and correct
important errors and omissions. The assessment should evaluate factual accuracy and
completeness, rather than simply whether the wording matches an existing report.
Saved-image testing alone cannot establish how a device affects image acquisition, protocol
changes, decisions about examination completeness, or the performance of users working with the
device. Saved images show what someone has already chosen to acquire. They cannot show what
intended users would have collected after seeing the AI’s guidance. When user review is claimed as
a safeguard, manufacturers should provide evidence that the intended users can perform that review
effectively.
Testing on live clinical inputs while recording AI outputs without showing them to clinicians or
patients, and without allowing them to affect care, is sometimes called “shadow deployment.” This
can reveal technical problems and differences among settings. It cannot establish how visible
outputs change scanning decisions or users’ reliance on the device. For those questions, FDA
should generally expect prospective studies in which intended users see and use the AI during
actual examinations, with appropriate patient protections. Any alternative approach should be
justified for the particular function, risk, and clinical conditions. Devices that act independently and
could cause serious harm need especially strong evidence, including testing of important failures
and how they are detected and managed.
Qualified, independent reviewers should assess whether the examination answers the clinical
question, includes the necessary diagnostic views, and communicates limitations correctly. They
should look for serious errors and examinations ended too soon. Where appropriate, researchers
could compare an AI-assisted examination with a separate examination by a sonographer with
relevant specialty expertise or compare care with and without AI assistance. Where feasible, the
reference assessment should be established without knowledge of the AI’s output.
Reviewing user–AI interactions may require access to the outputs but should use predefined
assessment criteria. Where feasible, time-stamped records of AI guidance and user responses
during scanning should help reviewers assess how the guidance influenced decisions to add,
repeat, or skip views or end the examination.
Manufacturers should describe the expected mix of patients, examination types, image quality,
clinical conditions, users, equipment, and sites. Studies should reflect routine use and deliberately
include difficult and uncommon cases that could reveal serious failures. Researchers should report
those additional cases separately or justify how they are combined with typical cases. Overall
averages should not hide poor performance in an important group.
Researchers should decide in advance which outcomes matter and include enough examinations to
measure them reliably. Depending on the function’s intended use, testing should separately assess
image adequacy, output accuracy, examination completeness, and the safety of decisions made
during scanning. Analysis should account for related observations from the same patient and
examinations performed by the same operator or at the same facility. Evaluation data should be

Page 7 of 14
appropriately independent of data used to train or tune the model, with potential overlap in patients,
examinations, and related data addressed. [6]
Ultrasound phantoms, simulations, and standardized scenarios can help test scanning mechanics,
communication, limits, and recovery from errors. They should not be substituted for patient-based
evidence where performance depends on variable anatomy, acoustic windows, evolving findings, or
patient tolerance.
Question 14

SDMS recommends choosing the comparison based on the task. For acquiring images and
completing an examination, the appropriate expert is often a sonographer with relevant specialty
skills. Physicians provide expertise in interpretation and diagnostic decisions. Physician
comparators, whether generalists or specialists, should have the knowledge and experience
required for the particular condition or examination. A review panel without relevant scanning
expertise may miss problems with technique or information that should have been acquired.
A device that covers acquisition and interpretation should be evaluated with input from sonographers
and interpreting physicians. Reference assessments should be established by qualified reviewers
independent of the device manufacturer and, where applicable, the underlying model developer.
Measurements should be compared with a suitable measurement reference or independent expert
assessment. Researchers should explain the limitations of the reference and decide in advance how
disagreements will be resolved.
The target should reflect acceptable care for the particular task and patient group. The performance
of a typical clinician may describe current practice, but it should not automatically set the target if
that practice falls below applicable quality expectations. Expert agreement also does not establish
correctness when the experts lack needed information or relevant skills.
For outputs that can be worded in several valid ways, scoring should allow those differences. It
should separately assess factual accuracy, completeness, communication of uncertainty and
limitations, and the appropriateness of any suggested action. Clinically important errors should be
distinguished from harmless wording differences, and critical safety errors should be reported
separately. A fluent, convincing answer can still be clinically wrong.
When the device is intended to assist a sonographer, the main clinical evaluation should examine
the sonographer and AI working together. Researchers should compare that performance with
sonographers of comparable training and experience performing the same task without AI, using
comparable cases and working conditions. Testing should include realistic time pressure, difficult
cases, and responses to incorrect AI outputs. It should establish whether users recognize and
correct important errors, rather than simply measuring agreement with the device. This team
evaluation should complement assessment of the device’s own outputs, errors, and limitations. The
international Good Machine Learning Practice principles support evaluating combined performance
in the intended use environment. [6]
For intended users who are qualified to perform the specific task but lack broader sonography
expertise, manufacturers should test representative users from that group and compare their
performance on the defined task with an appropriate expert assessment. If a manufacturer claims
that the device can perform the defined task on its own, testing should evaluate that use. Successful
performance with human assistance, review, or correction does not establish that the device can
perform the task safely and effectively without that support.
Before testing, researchers should define acceptable limits for important omissions, falsely
completed examinations, consequential measurement errors, and unsupported reassurance. They

Page 8 of 14
should explain why those limits protect patients and report the precision of the estimates, including
results for relevant patient and user groups. Any allowed difference from the comparison standard
should be defined before testing and justified by its effect on patient care.
The device should meet clinically justified safety and quality requirements. Gains in speed or
reduced workload should accompany adequate quality and safety and should not compensate for
failure on a critical safety measure.
Question 15

SDMS supports comparing a device with the care patients can actually obtain in the setting where it
will be used. Manufacturers of GenAI-enabled medical devices used in sonography should
demonstrate the claimed benefits in that setting while meeting clinically justified safety and
diagnostic quality requirements. Limited access should not lower clinically justified requirements for
safety, image quality, examination adequacy, or user competence. Comparisons should assess
clinically relevant outcomes, including patient outcomes where appropriate, as well as time to
appropriate imaging and time to an appropriate clinical decision.
These recommendations concern how manufacturers should evaluate device performance. They
should not be interpreted as SDMS’ endorsement of ultrasound use by individuals who lack the
qualifications or procedure-specific competence required for the task. Evidence of benefit from AI
assistance does not, by itself, establish an individual’s qualifications, expand their scope of practice,
or replace applicable education, credentialing, and competency requirements.
Manufacturers should justify the comparison using documented information about the care patients
would ordinarily receive without the device, including actual wait times, available imaging, and
referral options. A comparison with no intervention should reflect the actual setting and account for
reasonable alternatives. Needed care should not be withheld to make the device appear more
beneficial.
A study could compare the usual care pathway with the same pathway plus the device. A planned
study of actual use, or well-supported evidence from routine care, could assess time to an
appropriate clinical decision, image adequacy, completed referrals, delayed or missed findings,
unnecessary additional imaging, and patient burden. Comparisons should account for differences in
patient risk, clinical urgency, and availability of services. A limited examination that incorrectly
reassures the user could delay definitive care. The study should assess that possibility.
Narang and colleagues evaluated AI-guided, limited echocardiographic acquisition by nurses who
had no prior echocardiography experience but received study-specific training. Paired sonographer
examinations and blinded expert review provided a comparison for specified cardiac parameters. [9]
This conventional deep-learning study illustrates assessment of a defined acquisition task against
expert performance. It does not establish GenAI safety or demonstrate improved access or patient
outcomes relative to usual care. Such technical assessment can complement a usual-care
comparison when evaluating an access claim. The study is cited solely as a methodological
example, without endorsing the staffing model studied.
Even where care is delayed, the device should meet clinically justified requirements for safety and
diagnostic adequacy. Instructions, training, criteria for seeking help, and referral arrangements
should match the intended users and setting evaluated. Users and patients should understand what
the limited assessment can and cannot determine and when a more complete ultrasound
examination is needed.
Evidence that a device improves access to a narrowly defined ultrasound assessment should not
justify substituting that assessment for a clinically indicated comprehensive examination or

Page 9 of 14
bypassing the qualifications required of the person performing it. If broader use is proposed across
different settings, specialties, examination types, patient groups, or intended users, the evidence
and comparison should be reconsidered for those claims.
Question 18
SDMS recommends monitoring devices after they enter use, but that monitoring should not replace
the evidence needed beforehand to show that they are safe and effective for their claimed purpose.
It is particularly difficult to justify less evidence when a device can cause necessary information to go
uncollected or intended users cannot recognize its errors.
SDMS does not support relying on postmarket monitoring to resolve premarket uncertainty about
serious acquisition or examination-completeness failures when those failures cannot be reliably
detected and corrected before they affect patient care. If a problem is discovered only after the
patient leaves, obtaining a missed view or correcting an important measurement error may require
another examination. By then, misleading reassurance may already have affected clinical decisions
or delayed care. Some problems may never be discovered if the missing information is never
obtained.
A limited amount of remaining uncertainty might be acceptable for a narrow function whose errors
can be reliably checked and corrected before they affect care. An example could be editable draft
documentation that does not guide scanning. AI-generated draft documentation should be clearly
identified. Before it is treated as verified clinical findings or used to direct patient care, a reviewer
with the relevant training, experience, and task-specific competence should check its accuracy and
completeness against the source images and data. Documentation prepared for the sonographer’s
technical report to the interpreting physician should follow applicable reporting policies and review
requirements. Testing before release would still need to show that the review process finds and
corrects important errors and omissions.
If FDA relies more heavily on monitoring, the program should be ready before the device is
introduced. Manufacturers should provide:
• A clear explanation of what remains unknown, the safety and performance limits chosen in
advance, and evidence that monitoring can find important failures in time to limit harm.
• A reliable count of how many examinations or tasks use the device, a suitable sample of
clinical cases, independent expert review, and access to source data and follow-up when
needed.
• Records of the AI model and software versions, instructions given to the AI, instructions
provided to users, reference material, safety controls, connected tools, and screen displays,
including changes that could affect performance.
• Funding, named people responsible for the work, and agreements ensuring access to
needed information, including information held by an outside AI supplier.
• A plan with predefined triggers and response times for investigating problems, notifying
users, restricting or stopping use, and, where appropriate, returning to a version
demonstrated to remain safe for the intended use. The plan should also specify actions if
required monitoring or evidence collection cannot be completed.
• A plan and timetable for answering the remaining questions, with limits on use that match the
risk while evidence is collected.
Manufacturers should show that this monitoring is feasible. Reviewing the same saved images may
not reveal a view that was never acquired. Frequent user acceptance does not prove an answer is

Page 10 of 14
correct. If a serious failure cannot be found and addressed in time to protect patients, monitoring
cannot justify less evidence about that failure before release.
Question 19
SDMS supports repeat performance testing, independent clinical review of selected cases, and
monitoring for worsening performance. Repeat testing should assess the AI model and software
version actually in use, including the instructions given to the AI, reference material, safety controls,
screen display, and connected functions. Manufacturers should compare results with previously
established performance and clinically justified safety and performance limits set in advance.
Independent clinical reviewers with relevant specialty expertise should assess a sample of complete
examinations and important interactions between users and the device. When performance
changes, manufacturers should investigate contributing factors, including changes in the device,
patient population, users, equipment, or clinical practice.

What to monitor Why it matters
Nondiagnostic, failed, or abandoned These can reveal missing information, acquisition
examinations and examinations problems, or failure to recognize examination
incorrectly marked complete limitations.
Repeat examinations and recalls for These may reveal inadequate acquisition, avoidable
additional images work, and delays. Review should distinguish these
problems from clinically appropriate follow-up.
Measurement corrections, important These identify cases for review and help assess output
disagreements, and users rejecting accuracy, clinical usefulness, and whether users
AI advice recognize important errors.
Unexpected protocol changes, These may indicate design or workflow problems,
excessive alerts, and workarounds unclear instructions, training needs, or safeguards that
users bypass. Further review is needed to determine
the cause.
Changes over time in users’ reliance Independent case review and, when needed,
on AI and ability to recognize errors assessment of task performance can reveal increasing
acceptance of incorrect guidance or reduced ability to
recognize important errors.
Examination duration, repeated These can reveal workflow or ergonomic harms that
scanning, and changes in output accuracy and documentation speed alone may
sonographer workload or physical miss. Review should consider staffing, scheduling, and
strain case mix.
Later imaging, important findings Where follow-up information is available, these may
discovered later, and delayed reveal missed information, delayed care, or
requests for help unnecessary additional testing.
Complaints, near misses, and These can identify recurring problems that routine tests
adverse events recorded with the overlook and help determine whether particular
relevant AI model and software versions are affected.
version

Counts should include unsuccessful attempts. Rates should use a clearly defined, relevant total,
such as all AI-assisted examination attempts or all AI-generated measurements reviewed.
Manufacturers should explain when case selection or incomplete follow-up limits what the results
can show. Rarely rejecting AI advice may reflect accurate answers or errors that users do not
recognize. Independent clinical review should help distinguish between these explanations and
assess important disagreements.

Page 11 of 14
Results should be checked separately across the specialties, patient groups, user groups and levels
of experience, equipment, and sites covered by the device’s claims. Reviews should include
apparently normal cases and accepted answers, as well as difficult cases and known failures. When
a missing view is the concern, selected cases may need a separate examination or another reliable
clinical comparison, where clinically appropriate. Complaints and saved-image review alone may
miss these problems.
Manufacturers should set review intervals and triggers for additional review before deployment. The
schedule should reflect the severity of possible harm, how easily errors can be detected, how widely
the device is used, and the number of cases needed to detect a meaningful change. Useful deviceuse records may be collected continuously. Clinical review should be more frequent when a device
is first introduced and after important changes. Longer intervals may be reasonable once evidence
shows stable performance. Ongoing monitoring is part of the international Good Machine Learning
Practice principles. [6]
Additional review should follow serious safety complaints, results outside predefined limits, or
unexpected increases in repeat imaging, measurement corrections, or examinations incorrectly
marked complete. Changes to the AI model, instructions, reference sources, safety controls, screen
display, equipment and settings (including ultrasound systems, transducers, and presets), intended
users, or clinical setting may also require reassessment. Planned changes that could affect safety or
performance should be tested before implementation and monitored afterward. A serious safety
concern should prompt immediate assessment and, where needed, prompt user notification and
restrictions on use.
Reporting problems should be simple and fit routine work. Manufacturers should tell affected users,
including sonographers, about important findings, corrective actions, and any changes needed in
how the device is used. Data collection should minimize avoidable manual work and limit patient
information to what is needed for monitoring and investigation. Where practical, manufacturers
should use reliable automated collection of device-use and performance data and targeted audits to
reduce added work for sonographers and other users.
Monitoring arrangements should specify how patient information is protected when accessed,
retained, or shared with manufacturers or outside AI suppliers. Data collection should comply with
applicable privacy requirements and facility policies, with access limited to authorized purposes and
personnel. [10]
Question 21
SDMS recommends clearly defining stakeholder roles while keeping the manufacturer responsible
for its device. Manufacturers of GenAI-enabled medical devices used in sonography should provide
simple, accessible ways for sonographers, physicians, and other users to report suspected errors,
near misses, and unexpected device behavior, whether or not patient harm occurred. Manufacturers
should acknowledge reports, assess their urgency, investigate problems, identify recurring patterns,
and submit required reports to FDA. Users should also receive clear information about reporting
directly to FDA when appropriate.
Sonographers directly observe acquisition problems, inappropriate guidance, examination
limitations, and changes in device behavior. Their observations complement physicians’ assessment
of interpretation and subsequent clinical decisions.
Sonographers and other clinicians can report suspected problems, document examination
limitations, and help review cases within their expertise. Facilities can verify training and taskspecific competence, provide ways to obtain help or continue care safely without the device, support

Page 12 of 14
audits, and connect device concerns with patient safety processes. These activities require
adequate staff, time, and practical reporting tools.
Professional societies such as SDMS could bring specialty experts together, clarify terminology and
practice expectations, provide education about safe use and limitations, and help define useful
categories for reporting problems. Societies could also identify research needs, support independent
expert input, and flag recurring concerns raised by members. Member reports can identify potential
problems, but voluntary reporting alone cannot reliably establish how often those problems occur.
Standards organizations can develop consistent definitions, formats for records that different
systems can share, clinically useful testing methods, and ways to notify users about changes.
Patient representatives can help identify outcomes important to patients and explain the effects of
repeat examinations, misleading reassurance, delayed referrals, and difficulties returning for care.
Participants should disclose relevant conflicts of interest, and clinical review arrangements should
protect reviewers’ independent judgment.
The manufacturer should remain responsible for monitoring its device, obtaining needed data,
checking performance, investigating problems, taking corrective action, and communicating safetyrelated changes. It should also provide device-specific instructions and training materials. Using an
outside company’s AI model should not shift responsibility away from the manufacturer of the
finished device. Agreements with outside AI suppliers should ensure timely notice of changes and
access to information needed to evaluate performance and investigate problems.
Each monitoring arrangement should identify who collects and reviews information, who is
responsible for responding, who can restrict use, and how concerns reach the manufacturer and
FDA through the appropriate reporting channels. Required reports to FDA should be submitted
within applicable deadlines, even if an investigation is ongoing. [11]
Professional societies should not be made responsible for certifying an individual product’s safety,
funding monitoring that is the manufacturer’s responsibility, or fixing failures caused by a
manufacturer or AI supplier. Their role should match their expertise and capacity. Any formal
evaluation role should be clearly defined, adequately funded, and structured to protect independent
judgment. Clinicians and facilities retain their own practice and reporting responsibilities, while
manufacturers remain responsible for maintaining their devices’ safety and effectiveness.

References
[1] US Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled
Medical Devices Discussion Paper and Request for Feedback. August 2026. Sections IV through VI
and Appendix A. This paper seeks early input; it does not establish proposed or final FDA
requirements. FDA discussion paper and docket information

[2] Society of Diagnostic Medical Sonography and collaborating organizations. Scope of Practice
and Clinical Standards for the Diagnostic Medical Sonographer Revised 2023. Relevant sections
include terminology, personal scope of practice, Standards 1.3 through 1.6, quality improvement,
and competence. Scope of practice and clinical standards

[3] US Food and Drug Administration, Health Canada, and Medicines and Healthcare products
Regulatory Agency. Transparency for Machine Learning-Enabled Medical Devices Guiding
Principles. June 2024. Transparency principles

Page 13 of 14
[4] US Food and Drug Administration. Applying Human Factors and Usability Engineering to Medical
Devices. Final guidance, August 2026. This guidance addresses how intended users interact with
devices and how to reduce errors in use. FDA guidance on device use

[5] Society of Diagnostic Medical Sonography. State Licensure. States with licensure

[6] International Medical Device Regulators Forum. Good Machine Learning Practice for Medical
Device Development Guiding Principles. IMDRF/AIML WG/N88 FINAL 2025. Principles 3 through 5
and 7 through 10 are particularly relevant. Good Machine Learning Practice principles

[7] Society of Diagnostic Medical Sonography. Industry Standards for the Prevention of Work
Related Musculoskeletal Disorders in Sonography. 2017. These consensus standards address
ergonomic practices, workplace design, and equipment considerations intended to reduce
sonographers’ risk of work-related musculoskeletal disorders. Industry standards for WRMSD
prevention

[8] American Institute of Ultrasound in Medicine. Ultrasound Practice Parameters. Resources for
particular examinations and documentation. AIUM describes these as criteria for high-quality
examinations and notes that they are not intended to establish a legal standard of care. Ultrasound
practice parameters

[9] Narang A, Bae R, Hong H, et al. Utility of a Deep-Learning Algorithm to Guide Novices to Acquire
Echocardiograms for Limited Diagnostic Use. JAMA Cardiology. 2021;6(6):624-632.
doi:10.1001/jamacardio.2021.0185. This study tested conventional deep-learning guidance for
limited acquisition. It does not establish GenAI safety or support replacement of comprehensive
examinations. Original study

[10] Society of Diagnostic Medical Sonography. Code of Ethics for the Profession of Diagnostic
Medical Sonography. Effective September 24, 2024. Relevant principles address patient safety,
confidentiality, specialty-specific competence, professional accountability, and compliance with
applicable laws and facility policies. Code of ethics

[11] US Food and Drug Administration. Medical Device Reporting (MDR): How to Report Medical
Device Problems. Includes links to mandatory reporting requirements and manufacturer reporting
guidance. FDA reporting medical device problems

Online supporting resources accessed October 18, 2026.

Page 14 of 14