As artificial intelligence and automated technologies become more integrated into endocrine care, questions about liability increasingly extend beyond whether a device or algorithm performs as intended. For endocrinologists, patient selection, education, monitoring, documentation, and responses to known safety concerns can all influence how a physician’s use of these technologies is evaluated when something goes wrong. 

In part three of his conversation with AACE Endocrine AI Editor-in-Chief Johnson Thomas, MD, FSCE, FEAA, Steven Petak, MD, JD, MACE, FACP, explores the legal considerations surrounding the use of medical devices and artificial intelligence in endocrine practice. 

(Note: The information provided in this interview is for general informational and educational purposes only. It is not intended to constitute and should not be relied upon as legal advice. The following transcript has been edited for clarity and length.) 

Dr. Thomas: Now, we'll talk a little bit about malpractice suits and lawsuits that are happening. As you have alluded, there are several insulin pumps and closed-loop systems having major recalls and litigation. For a clinician who prescribes these systems, what's the real lesson from these cases?  

Dr. Petak: There have been several recalls which have been mainly based on product failure issues. These are product safety events, and they're not proof of the algorithm's negligence. But as far as the prescriber is concerned, the questions are still going to be there. Was the patient an appropriate candidate? Did the patient understand the alarms, the manual operation, testing for hypoglycemia and ketones, and whether a backup plan was present to use backup insulin if the pump was not operating properly? And most importantly, is the physician in the loop? Was there timely follow-up? Did the practice identify affected patients and communicate safety information to them?  

You're not expected as a physician to engineer the pump or guarantee against manufacturing defects. But you are expected to use the pump device reasonably, know its use and major warnings, educate the patient, and act on known safety information. The MAUDE (Manufacturer and User Facility Device Experience) database, for example, which lawyers in the FDA world know well, will identify these product claims (sometimes before a manufacturer lets the patient or the physician know). But note that a product claim against the manufacturer, as well as a negligence claim against the physician, can coexist simultaneously, so you really need to be aware of that. 

Dr. Thomas: Let's say one of these devices that I prescribed fails. How do physicians get affected? Can they blame the company that made the device? How can the physician's decision be brought into court and questioned? Can you give some examples? How is it practically affecting endocrinologists? 

Dr. Petak: Well, the defensible documentation should reflect what you've discussed with the patient. So, a risky note would say something like, “discussed the pump, start of the device, return in six months.” It doesn't show why the patient was selected, what was taught, or who owns the next step. So, it has to say something like, “We selected device X after reviewing hypoglycemia risk, renal function, cognition, dexterity, vision, response to alarms. Patient demonstrated bolus entry and understands these issues, and there is a backup plan with basal and rapid-acting insulin prescriptions and supplies, and that some follow-up will occur.” Download of the data, for example, in two weeks, with a follow-up in four to six weeks, or something like that. If material AI is involved, you need to also note the name of the tool and the version, if it is available, and why you used it, why you chose it for that particular patient. You need to show that you actively engaged with that patient, not just passive acceptance. 

Dr. Thomas: In our previous discussion, we talked about learned intermediary doctrine. Can you unpack what that means?  

Dr. Petak: In plain language, the learned intermediary doctrine in most jurisdictions—this is jurisdiction-dependent in part—the manufacturer of a prescription product warns the physician, and the physician translates those important risks and benefits for the patient. So, the patient stands at the end of the line, the physician stands between the manufacturer and the patient. AI really complicates that model, because a closed-loop system may continue making dose adjustments after the prescription is written, and an autonomous diagnostic tool may issue results without a specialist review. So there, physicians are not really mediating those decisions. It doesn't eliminate the doctrine, but it really raises much harder questions about who controls the decision and who's in the best position to warn the patient. For clinicians, a practical response is to understand the tool's limits, select the patients appropriately, teach the patient when and how to override or seek help, and document ongoing monitoring. The doctrine assumes the physician is in the middle of all of this, but autonomy makes the middle really hard to find.

Dr. Thomas: How would algorithmic bias show up in endocrine practice? Can you give us some examples? 

Dr. Petak: A lot of these tools are evolving, but let's hypothetically talk about a fracture risk model that was developed mainly in white postmenopausal women, which is kind of the history of some of these fracture risk models. It may perform well on average patients, but it may not be properly calibrated in men, younger patients, or some racial and ethnic groups that were underrepresented. If it underestimates a risk, a patient may miss treatment and later have a preventable fracture. If it overestimates risk, a patient may receive unnecessary medications, incur additional costs, and experience adverse effects. The legal exposure can run on two different tracks. There’s ordinary malpractice for individual patients and discrimination concerns if the system predictably disadvantages a protected group. 

The practical aspect of this is not to reject all algorithms; it's simply to know more about the validated population, check whether your patient resembles that population and apply additional scrutiny when they do not. And there are a lot of gray areas in fracture risk determination. For example, if a patient is from a particular ethnic group came to this country a year or two ago, do you use the specific data from their country of origin or have they already adapted to the clinical risks of their adopted country? And these are issues that are still of some concern when you determine whether your individual patient is really going to benefit from that fracture risk determination, using that as an example.