Ryden Solutions (Adam Foresman)

Comment from Ryden Solutions

Comment of Ryden Solutions — Docket No. FDA-2026-N-7874
Oct 7, 2026

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices — discussion paper
(PDF) · FDA request for feedback · Docket FDA-2026-N-7874 on regulations.gov · comments close
October 19, 2026

Interest of the commenter
Ryden Solutions ("Ryden") is a regulatory compliance technology company serving medical device
manufacturers. Its AI-enabled platform audits quality-system documentation against FDA, EU and ISO
requirements among others, and its team works daily with the quality and regulatory functions of
MedTech and pharmaceutical companies from startups to multi-facility enterprises, and with the QA/RA
consultants who support them. Ryden's advisory work includes dental AI software as a medical device:
FDA 510(k) strategy for adjunctive computer-assisted detection on dental radiographs (product codes
MYN, 21 CFR 892.2070, and QIH, 21 CFR 892.2050), clinical decision support determinations under the
Agency's January 2026 CDS guidance, and the quality systems behind both; spanning the manufacturers,
the practice-management platforms that integrate their outputs, and the dental practices that use these
functions chairside.

Ryden supports the discussion paper's central move: scaling evidence expectations to a function's level
of activity and the consequences of its failure, rather than to the technology inside it. This comment
proposes one refinement to the activity dimension and asks the Agency to recognize the risk-mitigating
weight of mandatory human authorization, reversibility, and time to harm; using dental imaging AI, the
specialty where these dynamics are clearest and the precedent base is mature, as a concrete worked
example.

Summary of requested changes
1. Add a distinct activity level between "non-directive information" and "action-directing":
action-directing information subject to mandatory, per-item human authorization before any
effect on the record or on care. A recommendation a clinician must individually review and
confirm, with the basis displayed and the source data one tap away, is categorically different
from a recommendation that acts, queues an order, or writes to the record on its own, and the
framework currently gives the two the same row.
2. Add reversibility of harm and time to harm as explicit scoring dimensions (or as defined
modifiers within the consequence dimension). A function whose worst-case failure is a
correctable record entry, caught at a review step that occurs before any treatment is scheduled
or consented, does not present the consequence profile of a function whose failure forecloses a
time-critical intervention; even when the two sit in the same clinical specialty.
3. Define the lower-risk cell these criteria create, with proportionate evidence expectations.
Where outputs are per-item authorized, fully verifiable at the point of use against source data
the clinician is already viewing, reversible, and slow-to-harm, premarket evidence should center

Page 1 of 4
 on bench performance and retrospective validation, not prospective clinical-outcome studies,
with the qualifying criteria verified by design review of the authorization gate itself.

Why dental imaging AI is the worked example
Dentistry offers the framework a clean test case, because the mitigations the proposed row depends on
are structural to the care setting rather than bolted on:

• The clinician verifies against the source data at the point of use. Adjunctive dental detection
software is labeled and cleared for use by a trained professional reviewing the radiograph or
other modality itself. Every AI output — a detection, a measurement, a historical treatment or
normal anatomy detection to setup a new patient, a suggested chart entry — is checked
chairside against the image it came from, by the professional who will treat the patient.
Verification is not an extra workflow step; it is the workflow.
• Time to harm is long by construction. A charting or treatment-planning output does not treat
anyone. Treatment follows examination, diagnosis, patient consent, insurance determination,
and separate scheduling; each a point where an erroneous suggestion is caught. Contrast timecritical settings (stroke triage, sepsis prediction) where an output influences care in minutes; the
framework currently offers both the same activity rows.
• Failure is reversible and detectable before harm. The worst case for a confirm-gated charting
suggestion is a wrong entry corrected at review. There is no irreversible physiological action
anywhere in the loop.
• The precedent base is mature. Multiple adjunctive dental AI devices have been cleared under
MYN and QIH with bench and standalone-performance evidence and, where warranted on
clinical-impacting detection types, reader studies. The proposed row would codify the risk logic
those clearances already embody; it does not lower any bar that exists today.

The same structural features recur across specialties — dermatology photo documentation, routine
radiology measurement review — so the row generalizes; dentistry is simply where every criterion is
met at once and the clinical stakes of the gated outputs are demonstrably bounded.

Proposed refinement
Ryden proposes the activity dimension read, in ascending order: (1) non-directive information; (2)
authorization-gated recommendations (new); (3) action-directing outputs; (4) clinician-supervised
action; (5) fully autonomous action. A function qualifies for level 2 only if it meets all of the following,
verifiable at design review:

Qualifying criterion What the Agency can verify

Per-item human authorization before any record No write path, order, or notification fires without
or care effect clinician acceptance; each output individually
editable

Basis displayed; source data at point of use Each recommendation names the finding(s) it
derives from, with the underlying data (e.g., the
radiograph) one interaction away

Page 2 of 4
Qualifying criterion What the Agency can verify

Explainability: independently reviewable logic The relationship between input and
recommendation is disclosed such that the
authorizing clinician can re-derive it

Reversible effect The authorized output is a record entry or plan
item correctable at any later review, not an
action delivered to the patient (e.g., drug dosing
or energy delivery). There should be an audit
trail history to verify.

Time to harm exceeds time to detection At least one independent downstream
checkpoint (consent, scheduling, treatment
appointment) occurs between authorization and
any patient effect

FDA activity axis as proposed vs. redline with the new authorization-gated level

The figure shows the paper's activity axis as proposed and the redline with the inserted level; the
paper's own Figure 1 (linked above) is the authoritative image, and the labels here paraphrase its text.

For functions in this row, proportionate premarket evidence is standalone/bench performance on
representative data plus retrospective validation of the recommendation logic; human-factors evidence
that the authorization gate functions as designed; and no requirement for prospective clinical-outcome
studies. Consequence severity still scales the evidence within the row; the row bounds activity, not
specialty.

Guardrails the row must carry
The row is defensible only if its boundaries are enforced, and Ryden asks the Agency to state them
explicitly. A function leaves the row, and takes its ordinary action-directing classification, when any of

Page 3 of 4
the following is present: any write to the patient record before authorization, including provisional or
"pending" entries visible in the record; composite clinical scores or rankings (quality, priority, or
confidence indices that adjudicate among options on clinical grounds), which reintroduce a non-rederivable clinical determination above the gated items; and learned models whose basis the
authorizing clinician cannot re-derive at the point of authorization. The dental AI manufacturers Ryden
works with apply these boundaries in their own products today, and the market runs the other way:
practice-management integration partners increasingly request unreviewed record population. A
framework row that rewards the authorization gate, and withdraws when the gate is hollowed out,
would align commercial incentives with the Agency's safety logic.

Conclusion
The discussion paper's activity-and-consequence structure is the right frame. Its current rows, however,
treat a confirm-gated recommendation the same as an output that acts — which overstates the risk of
the most common, most verifiable AI-assistance pattern in ambulatory care, and understates the
regulatory value of the human authorization gate the Agency's own commenters have asked it to
require before irreversible actions. Adding an authorization-gated recommendations row, scored with
reversibility and time to harm, would scale evidence to real risk, reward manufacturers who build the
gate, and give lower-risk specialties like dentistry, where verification against source data is the native
workflow, a classification that matches how care is actually delivered. Ryden appreciates the Agency's
consideration and is available to discuss the dental use cases in detail.

Contact

Adam Foresman

Co-Founder and CEO

Ryden Solutions

adam@ryden.ai

Page 4 of 4