VitaSignal LLC (Alexis M. Collier)

Please see the attached comment from Dr. Alexis M. Collier, founder of VitaSignal, responding to FDA-2026-N-7874. It offers conceptual recommendations on clinical AI evidence, human authority, monitoring and reassessment. It does not report completed product validation, clinical benefit or FDA endorsement.

Comment on FDA-2026-N-7874
Alexis M. Collier, Founder, VitaSignal LLC
Prepared October 2, 2026

I am the founder of VitaSignal LLC, a research-stage company developing
clinical AI evaluation and governance approaches. I offer the following
recommendations as conceptual input, not as findings from a completed
institutional validation, evidence of clinical benefit, or a request for
endorsement. My interest is in evidence that remains interpretable when an
AI system operates within a local clinical documentation and decision
pathway.

### Question 1 Risk assessment

I recommend assessing whether a safeguard can actually interrupt the
pathway to harm. Naming a reviewer is not enough. Evaluation should
establish what information the reviewer receives, how much time is
available, whether the reviewer has relevant expertise and authority, and
what happens when review cannot occur. A hypothetical chart-synthesis
system could display accurate source text yet omit the observation time.
Review of that summary would not necessarily reveal that its information
was unavailable at the decision point. Tests should distinguish source
access from usable decision-time evidence.

### Questions 12 and 13 Evidence and synthetic data

I recommend keeping generated challenge cases and empirical performance
estimates separate unless a justified statistical design supports
combining them. Generated cases can test whether a system blocks an
unsupported action or retains an unresolved conflict. Passing those tests
does not establish the frequency of such failures in practice.

Reports should distinguish a missing field among represented patients from
an absent population. Creating plausible records for an unobserved group
does not establish coverage of that group's actual circumstances.
Evaluation should identify which observation opportunities, documentation
processes and reference outcomes are missing. Where a subgroup estimate is
not supportable, report that limitation rather than supply a synthetic
estimate as though it were observed.

### Question 19 Monitoring and reassessment

I recommend combining scheduled review with event-triggered review.
Relevant triggers should include changed input definitions, local
mappings, documentation timing, user roles, retrieval sources and workflow
conditions, as well as model changes. The responsible parties should
define which evidence remains applicable after each type of change.

Monitoring records should retain the source and version supporting a
conclusion, the analysis state, the limitation and the decision taken.
Blocked, null, adverse and conflicting findings should survive storage and
export. A fluent generated summary should not convert an unresolved
finding into reassurance. Monitoring should also measure failure to detect
a problem, inappropriate interruption and delayed escalation separately. A
low alert count alone does not establish reliability.

### Question 20 Supervisory agents

I recommend evaluating a supervisory agent as a separate system with its
own failure modes. Test whether it can detect relevant failures when it
shares models, retrieval sources or assumptions with the system it
monitors. Include missing inputs, stale evidence, unavailable monitoring,
misleading tool outputs and disagreement between monitors.

The supervisor's recommendation should remain distinct from an authorized
decision. Where an automated block is part of the approved design, tests
should verify that it is enforced even when generated text recommends
proceeding. Where human approval is required, test that neither the
evaluated agent nor its supervisor can grant itself new permissions.
Preserve identifiable AI activity and the reason for any human override.
These are proposed test objectives, not claims that an existing product
meets them.

### Question 21 Responsibility

I recommend written allocation of operational tasks without treating that
allocation as a transfer of manufacturer accountability. Identify who
supplies monitoring information, investigates signals, authorizes local
restrictions and communicates changes. Establish design authority before
execution and identify the authorized reviewer at each consequential
action. Preserve reviewer disagreement as evidence about ambiguity or
operating conditions, not merely as a failed acceptance metric. Review
capacity, staffing availability and fallback should be evaluated
independently of whether a nominal owner is listed.

### Questions 22 and 24 Changes and foundation models

I recommend retaining a versioned evaluation baseline and testing the
consequences of a change against the actual intended workflow. Compare
relevant inputs, outputs, permissions, tools and record transformations
before and after the change. Specify what requires broader reassessment
and when the system should be restricted pending evidence. Contractual
notification, technical version detection and fallback are complementary
controls; none alone establishes that a changed system remains suitable.
Classification of a change for regulatory purposes should follow the
applicable framework rather than a universal label proposed here.

### Question 26 Agentic systems

I recommend testing complete action sequences, including partial execution
and recovery. A useful challenge should attempt a blocked action, correct
the missing or conflicting evidence, obtain authorized review, export the
resulting record and reconstruct why the decision occurred. Replay should
distinguish the original evidence from later amendments. Tests should also
cover revoked permissions, repeated actions, failed tools and interruption
between steps. Report whether rollback is possible and which consequences
cannot be reversed.
Technical performance, clinical benefit and economic effects require
distinct endpoints. Successful execution or user acceptance should not
substitute for evidence of patient benefit. These recommendations seek
auditable evaluation and proportionate safeguards without prescribing
proprietary algorithms, numerical thresholds or a single commercial
implementation.