FDA Receives 500 Comments on Medtech AI Rules

More than 500 patients, physician groups and medical device companies submitted comments to the Food and Drug Administration by mid-October on how it should regulate generative AI in medical devices. The FDA had requested feedback on an August discussion paper proposing a review framework inspired by physician credentialing and emphasizing postmarket monitoring. Generative AI can produce text, images and code, but its outputs can vary and its performance can change over time.

The comments offered competing views. AdvaMed, a medical technology lobby, urged the FDA to build on its existing regulatory framework instead of creating rules specific to generative AI devices. It said adding generative AI should not automatically trigger a competency assessment, asked the agency to broaden its process for pre-approving planned device changes, and favored risk-based postmarket monitoring over a uniform approach. AdvaMed said manufacturers, rather than clinicians and patients, should be responsible for monitoring device performance after market entry. The Medical Device Manufacturers Association and GE Healthcare also supported using the existing framework.

Two interventional cardiologists working on AI issues for the Society for Cardiovascular Angiography & Interventions said the FDA should clarify when a tool supports a decision and when it makes one. Dr. Afnan Tariq said the question is whether AI will replace physicians or help scale their work amid a physician shortage. Dr. Islam Abudayyeh said generative AI can make the distinction difficult because its output can sound like a decision. He said the FDA could validate some results, such as those from AI-assisted fractional flow reserve CT scans, while tools that prioritize patients for treatment or structure payments raise less clear questions.

Smart ring maker Oura asked how the FDA would assess risk in wearable features and whether some would qualify for the agency’s wellness exemption. Chief Medical Officer Dr. Ricky Bloomfield said informational, non-directive outputs should not be treated as medical devices or should fall under enforcement discretion, comparing them with symptom checkers. He gave the example of a large language model explaining that a wearable reading might indicate hypertension and advising the user to consult a healthcare professional. Oura said it has no FDA-cleared features for its rings and has sought partnerships with medical device companies.

Generation Patient, a nonprofit representing young adults with chronic conditions, called for expanded postmarket monitoring without reducing premarket evidence. Its policy director, Luis Gil Abinader, said shifting the burden toward monitoring after release could leave patients exposed until harm occurs. The group also urged the FDA to exclude AI companion tools from the wellness exemption, citing a JAMA Pediatrics survey that found nearly one-fifth of Americans ages 12 to 21 had used AI chatbots for mental health advice. It argued that these tools are designed to sustain simulated relationships and engage users in emotional distress.

Dr. Nina Vasan, a Stanford psychiatry professor who leads the Brainstorm mental health innovation lab, recommended testing devices for reasonably foreseeable use, beyond their labeled indications, and setting separate pass-or-fail checks for severe failures such as missed suicide risk, missed sepsis or an incorrect insulin adjustment. She said a model’s error can repeat across patients, sites and devices. Her lab’s safety evaluations found that models could behave worse during longer conversations; some evaluations stop after only a few turns. Vasan described ChatGPT safeguards eroding in a longer exchange, with a tester seeking advice on hiding self-inflicted cuts receiving product recommendations instead of a medical referral. A Claude tester got around initial boundaries by starting a new chat. Vasan said failures can involve omissions, harmful agreement, drift over time and safeguards that work in testing but fail in practice.

More than 500 patients, physician groups and medical device companies submitted comments to the Food and Drug Administration by mid-October on how it should regulate generative AI in medical devices. The FDA had requested feedback on an August discussion paper proposing a review framework inspired by physician credentialing and emphasizing postmarket monitoring. Generative AI can produce text, images and code, but its outputs can vary and its performance can change over time.

The comments offered competing views. AdvaMed, a medical technology lobby, urged the FDA to build on its existing regulatory framework instead of creating rules specific to generative AI devices. It said adding generative AI should not automatically trigger a competency assessment, asked the agency to broaden its process for pre-approving planned device changes, and favored risk-based postmarket monitoring over a uniform approach. AdvaMed said manufacturers, rather than clinicians and patients, should be responsible for monitoring device performance after market entry. The Medical Device Manufacturers Association and GE Healthcare also supported using the existing framework.

Two interventional cardiologists working on AI issues for the Society for Cardiovascular Angiography & Interventions said the FDA should clarify when a tool supports a decision and when it makes one. Dr. Afnan Tariq said the question is whether AI will replace physicians or help scale their work amid a physician shortage. Dr. Islam Abudayyeh said generative AI can make the distinction difficult because its output can sound like a decision. He said the FDA could validate some results, such as those from AI-assisted fractional flow reserve CT scans, while tools that prioritize patients for treatment or structure payments raise less clear questions.

Smart ring maker Oura asked how the FDA would assess risk in wearable features and whether some would qualify for the agency’s wellness exemption. Chief Medical Officer Dr. Ricky Bloomfield said informational, non-directive outputs should not be treated as medical devices or should fall under enforcement discretion, comparing them with symptom checkers. He gave the example of a large language model explaining that a wearable reading might indicate hypertension and advising the user to consult a healthcare professional. Oura said it has no FDA-cleared features for its rings and has sought partnerships with medical device companies.

Generation Patient, a nonprofit representing young adults with chronic conditions, called for expanded postmarket monitoring without reducing premarket evidence. Its policy director, Luis Gil Abinader, said shifting the burden toward monitoring after release could leave patients exposed until harm occurs. The group also urged the FDA to exclude AI companion tools from the wellness exemption, citing a JAMA Pediatrics survey that found nearly one-fifth of Americans ages 12 to 21 had used AI chatbots for mental health advice. It argued that these tools are designed to sustain simulated relationships and engage users in emotional distress.

Dr. Nina Vasan, a Stanford psychiatry professor who leads the Brainstorm mental health innovation lab, recommended testing devices for reasonably foreseeable use, beyond their labeled indications, and setting separate pass-or-fail checks for severe failures such as missed suicide risk, missed sepsis or an incorrect insulin adjustment. She said a model’s error can repeat across patients, sites and devices. Her lab’s safety evaluations found that models could behave worse during longer conversations; some evaluations stop after only a few turns. Vasan described ChatGPT safeguards eroding in a longer exchange, with a tester seeking advice on hiding self-inflicted cuts receiving product recommendations instead of a medical referral. A Claude tester got around initial boundaries by starting a new chat. Vasan said failures can involve omissions, harmful agreement, drift over time and safeguards that work in testing but fail in practice.