Glytec (Robby Booth)

I am co-founder and Chief Product Officer of Glytec, which Dr. Bruce Bode and I started together in 2006. That year our inpatient insulin dosing software became the first of its kind cleared by FDA for hospital use, and it is now in use in hospitals throughout the United States. Glytec has received multiple 510(k) clearances since, most recently in 2026 that included a Predetermined Change Control Plan (PCCP). In addition to its FDA-cleared SaMD, Glytec also develops clinical AI for hospitals.

Comments on Questions 1 and 4

The discussion paper used insulin dosing as an example of the high end of the consequences axis. Insulin is a high-alert medication and errors in either direction can cause harm. Too much can lead to hypoglycemia, and too little can lead to hyperglycemia. It’s worth noting that the same dosing function can carry different risk depending on the use environment. For instance, an insulin dose recommendation to a clinician in the hospital for a critically ill patient is not the same as an instruction to a patient adjusting basal insulin at home, even though both sit high on the consequences axis and FDA has cleared software for both uses. Within that framework, consequence should carry the most weight. Where software determines a medication dose using patient-specific inputs, FDA should expect the dose to come from a validated engine whose output can be reproduced from its inputs. GenAI can support that engine by explaining a recommendation or retrieving data. If a sponsor intends a GenAI model to determine the dose, the evidence standard should equal the one the engine was required to meet.

In hospitals, much of the insulin is managed by clinicians who are not diabetes specialists. FDA-cleared dosing software gives these clinicians specialist-level guidance they often cannot verify independently, or cannot verify quickly enough when a critically ill patient needs a dose right away. This is one of the additional dimensions Question 1 raises. When a clinician has little time to evaluate an output, time pressure should move a function higher on the consequences axis, and the evidence should reflect the intended users and use environment.

Comments on Question 18

Accepting greater premarket uncertainty in exchange for post-market monitoring may suit lower-risk functions. For those functions, it should depend on monitoring that can catch a problem before it spreads. The Agency should not extend it to functions that determine a high-risk medication dose. Postmarket monitoring identifies failures by observing them, and for a dosing function, each observed failure can be a patient harmed. Dose-determining software can be validated thoroughly before clearance, and later changes can be validated under protocols authorized in advance through a PCCP, so the uncertainty is avoidable.

Comments on Questions 22 - 24

PCCPs work best when each planned change is described narrowly enough for a reviewer to clearly see what will change and what will not. For GenAI devices, changes to the model itself should be treated separately from changes to how it is instructed or what information it draws on, with testing for each. Whether a model can be replaced under a PCCP should depend on the clinical consequence. For lower-risk functions such as documentation, a like-for-like model upgrade, tested against the same cases used before and held to limits set in advance, is reasonable. For functions that determine a high-risk medication dose, model replacement should not be included in a PCCP. A change made by a third-party model developer should be treated as a model replacement under the same rules.

Average accuracy can improve while rare, serious errors become more frequent. A GenAI PCCP should identify the worst errors the device could make, such as doses that lead to hypoglycemia, and set a separate limit for each. It should also include a tested plan for reversing a change.

On the topic of re-benchmarking, this can be suitable for testing a device on its own, but in practice safety depends on the device and the user (clinician) together. Experienced users learn where a device’s outputs deserve a closer look. An update that shifts where errors occur could score better and still be less safe, because the new errors land where no one is looking. For high-consequence functions, appropriate human factors testing of a change should include clinicians who used the prior version.

Thank you for the opportunity to comment.

Sincerely,
Robby Booth