Neko Health (Sander Riedberg)

See attached file(s)

DOCKET NO. FDA-2026-N-7874

PUBLIC COMMENT

Comment on Docket No. FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback

SUBMITTED BY Neko Health
CONTACT Sander Riedberg, Regulatory Affairs, sander.riedberg@nekohealth.com
DATE October 19, 2026

00 Introduction
Neko Health welcomes the opportunity to comment on this discussion paper. Neko Health is a
preventive health company. We operate clinics in Sweden, the United Kingdom, and the United
States. Most of our members have no symptoms. They come to us to understand their health and to
act early. We develop and operate medical devices and health software for this setting.

We support the direction of the discussion paper. The two-axis risk framework and the
competency-based approach give a useful structure for risk-proportionate oversight. We limit our
comments to two areas. In both areas, our experience in preventive care and in three jurisdictions
can add a specific perspective.

01 Risk and postmarket monitoring for people without symptoms
RELATES TO: RISK ASSESSMENT FRAMEWORK; QUESTION 18

Observation
The consequence axis describes the harm that can occur when a user relies on an incorrect output.
Most examples in the paper come from settings where the patient already has symptoms or a
diagnosis. In preventive care, most users are healthy. This changes the type of harm.

In this population, an incorrect output can cause harm in two directions. A false negative can give
false reassurance, and the person can then delay follow-up for a condition that is present but does
not yet present symptoms. A false positive can cause unnecessary tests, procedures, and anxiety.
Both harms are real. Both increase with the number of people who receive the output.

In preventive care, outcomes also occur late. A missed finding can take years to become clinically
visible, and the event rate is low. Postmarket monitoring that depends only on clinical outcomes
will therefore find problems too late.

Recommendations
– Include both false reassurance and unnecessary follow-up as explicit types of consequence on
the consequence axis. This is particularly important for functions that give outputs to people
without symptoms.

NEKO HEALTH // PUBLIC COMMENTPAGE 1 OF 3
 DOCKET NO. FDA-2026-N-7874

– Where FDA accepts more premarket uncertainty in exchange for postmarket monitoring
(Question 18), accept leading indicators as the primary monitoring signal. Examples are
periodic re-benchmarking, sample-based clinician review of real-world outputs, drift
monitoring, and the rate of recommended follow-up.
– Ask sponsors to define these indicators and their thresholds before market entry. This keeps
the balance between premarket and postmarket evidence predictable for FDA and for
sponsors.

02 Human review as a risk control, and international alignment
RELATES TO: RISK ASSESSMENT FRAMEWORK; QUESTION 7

Observation
We understand the FDA position that a disclaimer does not make an action-directing output less
directive. The same logic applies to human review. A clinician can review an output, but this alone
does not show that the review controls the risk.

Human review is less effective when the volume of outputs is high, when the output looks complete
and plausible, and when the reviewer cannot easily check the output against its source data. These
conditions are common for generative AI. The discussion paper does not yet define when human
review is an effective risk control.

We operate under three regulatory systems. Regulators in these systems do not agree on when
clinician review is sufficient to keep a generative AI function outside medical device regulation, or
to lower its risk. As a result, the same function can require different safety controls in different
markets. This does not benefit patients.

Recommendations
– Accept human review as a risk control only when the sponsor shows that the review is effective
in the intended setting. Evidence can include measured review and edit rates, review time per
output, and the rate at which reviewers find errors that the sponsor deliberately inserts into test
outputs during validation.
– Consider traceability to source data as a benchmark element when an output summarizes or
interprets patient data. A reviewer can verify an output more easily when its statements link to
the data that supports them.
– Work with other regulators, for example through the International Medical Device Regulators
Forum (IMDRF), toward common criteria for effective human oversight of generative AI.
Alignment with the human oversight requirements in Article 14 of the EU AI Act would also
help.

NEKO HEALTH // PUBLIC COMMENTPAGE 2 OF 3
 DOCKET NO. FDA-2026-N-7874

03 Closing
We thank FDA for the early and open engagement on this topic. We are available to discuss these
comments or to share more of our experience.

Sander Riedberg
Regulatory Affairs, Neko Health
sander.riedberg@nekohealth.com

NEKO HEALTH // PUBLIC COMMENTPAGE 3 OF 3