Welldoc, Inc.

Welldoc, Inc., a manufacturer of regulated Software as a Medical Device (SaMD), appreciates the opportunity to comment on Docket No. FDA-2026-N-7874.

Our comments touch on the following topics:

- Defining Small Language Models (SLMs) for edge devices.
- Incorporating probability of occurrence into the two-axis risk framework and using hybrid-AI architectures.
- Multi-dimensional benchmarking elements and risk-stratified benchmarks.
- Postmarket monitoring approaches.

Please see the full comment in the attached PDF.

Date: October 19, 2026

Division of Dockets Management (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

RE: Docket No. FDA-2026-N-7874 Considerations for the Regulation of Generative AIEnabled Medical Devices: Discussion Paper and Request for Feedback

Subject: Welldoc Comments on Considerations for the Regulation of Generative AIEnabled Medical Devices

Statement of Interest
Welldoc, Inc. appreciates the opportunity to submit comments to the Food and Drug
Administration (FDA) Center for Devices and Radiological Health (CDRH) regarding the
discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical
Devices. As a manufacturer of regulated Software as a Medical Device (SaMD), Welldoc
submits these recommendations to support DHCoE in addressing the unique risks,
evaluation criteria, and lifecycle monitoring approaches needed for GenAI technologies.
Research cited in these comments was conducted by Welldoc in collaboration with Johns
Hopkins University; Welldoc employees are co-authors.

Executive Summary
Generative AI-enabled medical devices hold transformative promise for clinical care, but
they require modernized regulatory frameworks that account for non-deterministic
behaviors, model architecture, and real-world performance tracking. In response to
DHCoE’s request for stakeholder feedback, Welldoc’s comments address critical topics
raised in the discussion paper, including:
• Section III (Definitions): Defining Small Language Models (SLMs) and their clinical
utility on edge devices.
• Section IV (Assessment of Risk): Incorporating probability of occurrence into the
two-axis framework and leveraging hybrid-AI architectures (Questions 1, 2, 5).
• Section V (Competency-Based Premarket Evaluation): Recommending additional
benchmarking elements and risk-stratified benchmarks (Questions 8, 9).
• Section VI (Postmarket Monitoring): Supporting the Agency’s proposed postmarket
monitoring approaches (Question 19).

Regarding Section III: Definitions
Welldoc suggests the Agency define the term Small Language Model (SLM), which are
“lightweight” artificial intelligence algorithms that generally exhibit fewer parameters
(typically ranging from a few million to around 10 billion) and a simpler neural architecture
than those in large language models. SLMs can be trained or fine-tuned on knowledge that
PHONE +1 (443) 692-3100
FAX +1 (443) 692-3099

10221 Wincopin Circle, Suite 150
Columbia, Maryland 21044 WWW.WELLDOC.COM
is pertinent to the intended use of the model (e.g., a model to guide protein and strength
training for use with a GLP-1 need not be able to answer questions about Shakespeare,
etc.). They are computationally more efficient, require less power, and can run locally on
“edge” devices such as smartphones. These models have tremendous use in regulated
medical devices where they can perform specific clinical functions that will be easier to
measure, benchmark, and assess for quality, safety, accessibility, and accuracy. Although
not a language model, Welldoc’s research illustrates that compact, purpose-built
generative models can perform well on focused clinical tasks: a generative transformer
model trained on CGM data reduced 1-hour glucose forecasting error by 48.5% compared
with baseline methods [1].

Regarding Section IV: Assessment of Risk 2-Axis Framework
Welldoc generally agrees with the two-axis framework, provided that the vertical (“y”) axis
is the product of the severity of the risk (as shown in the document) AND the probability that
the risk should occur (consistent with ISO 14971, the FDA-recognized standard for medical
device risk management, and the traditional FMEA approach used within it) (Question 1).

Welldoc believes that the framework’s discrete risk categories are an acceptable point to
begin the dialog. As model complexity, accuracy, testability, and traceability mature, a
dialog on a higher-resolution view (i.e., continuous) of risk can take place, consistent with
the Agency’s recognition that directiveness may exist along a continuum (Question 2).

Welldoc believes in the use of “hybrid-AI”: that is, a combination of GenAI with more
deterministic forms of AI such as support vector machines or rules engines, that can
complement the use of a generative AI solution when risk is inherently higher. Such a hybrid
architecture is also valuable when a series of decisions in a progressive query or
progressive dialog presents additional risks that the GenAI model may not be trained to
handle – at such a point, the decision is escalated to a more traditional rules-based
system, thus allowing complex decisions to be made autonomously but more confidently
against evidence-based medicine guidelines used by HCPs today. Although conducted on
glucose forecasting rather than dialog, Welldoc’s research found that a generative model’s
prediction error rose as its perplexity increased, suggesting that model-internal uncertainty
can serve as a measurable escalation trigger [2] (Question 5).

Regarding Section V: Competency-based Frameworks
Welldoc generally agrees with the need for a multi-dimensional framework that is used to
measure GenAI models. These dimensions can include, but are not limited to:
• Clinical accuracy
• User-friendliness or user-interpretation (e.g., readability, understandability)
• Model accuracy
• Social accuracy (e.g., bias, fairness, toxicity)
• Traceability and reproducibility
• Time-based appropriateness (e.g., drift)
PHONE +1 (443) 692-3100
FAX +1 (443) 692-3099

10221 Wincopin Circle, Suite 150
Columbia, Maryland 21044 WWW.WELLDOC.COM
These dimensions align closely with the Agency’s proposed benchmarking elements;
Welldoc recommends adding traceability and time-based appropriateness, which are not
currently included (Question 9). Once benchmarking elements are finalized, reasonable
benchmarks or targets should be put forth. These benchmarks can vary for different levels
of risk – that is, a GenAI model to detect A-Fib may need to meet a higher clinical accuracy
requirement than a GenAI model that suggests a fitness plan to lose weight (Question 8).

Regarding Section VI: Postmarket Monitoring
Welldoc agrees with the Agency’s depiction of approaches to postmarket monitoring
(Question 19).

Conclusion
Welldoc commends DHCoE and CDRH for initiating this timely dialogue to advance
regulatory approaches for GenAI-enabled medical devices. We respectfully submit these
targeted responses under Docket No. FDA-2026-N-7874. Welldoc would be pleased to
participate in future dialog related to the evolution and implementation of GenAI in
regulated medical devices.

References
1. Luo J, Kumbara A, Shomali M, Han R, Iyer A, Aleppo G, Agarwal R, Gao G. A large
sensor foundation model pretrained on continuous glucose monitor data for
diabetes management. npj Health Systems. 2025;2:35. doi:10.1038/s44401-02500039-y
2. Luo J, Kumbara A, Iyer AK, Shomali ME, Gao G. Evaluating Perplexity and Glucose
Level Impact on State-Of-The-Art Generative Pre-trained Transformer (GPT) Model
to Predict Glucose Values at Different Time Intervals. Diabetes Technology Meeting
2024, Burlingame, CA, USA.

Respectfully submitted,
Welldoc, Inc.
www.welldoc.com

PHONE +1 (443) 692-3100
FAX +1 (443) 692-3099

10221 Wincopin Circle, Suite 150
Columbia, Maryland 21044 WWW.WELLDOC.COM