AI

ECRI says it will work with the FDA to track how AI medical devices perform in hospitals

Marcus Schabacker said the nonprofit is seeking U.S. hospitals to help develop a framework for evaluating AI devices after deployment. He urged buyers to settle how they will detect failures and who can take a device offline.

Author
HDM Staff
Save
Share
Illuminated HLTH lettering and event entrance arch
Event branding at the entrance to HLTH USA 2026 in Las Vegas. Photo credit: HLTH USA 2026

    ECRI has received a contract from the Food and Drug Administration to help develop a way to evaluate AI-enabled devices in real-world use, the nonprofit’s president and CEO, Marcus Schabacker, said Wednesday. He said ECRI is looking for U.S. hospitals to help build the framework, with the aim of measuring device performance over time in the settings where clinicians actually use the tools.

    Schabacker described the work during the opening keynote and Q&A at The Futurist Physical AI Summit, part of HLTH USA 2026 in Las Vegas. Summit host and curator Ian Khan introduced him and led the Q&A. Schabacker gave no timetable, contract value or participating hospital names during the session. He said ECRI had also extended its existing medical-device problem reporting network to AI and had begun receiving reports.

    The contract and the reporting network address related problems. One is tracking how a device behaves across hospitals and over time. The other is giving clinicians and hospitals a way to flag an incident when it happens. For an operating room or another setting where software can guide a physical action, a hospital also needs a plan for what staff do between noticing something wrong and getting a device safely back into use.

    What buyers should ask before deployment

    Schabacker began with the proposed use. A purchasing committee should ask whether a device solves the problem staff are encountering and whether it fits their working conditions. A surgeon who wants to use a new system can explain what it might make possible in the procedure. The staff who set it up and the anesthesiologist who needs working space around the patient can identify constraints that the same product demonstration may miss.

    He also asked buyers to examine who was represented in the data used to develop and test the device. “There can’t be any black boxes,” Schabacker said. He compared the disclosure he wants from suppliers to a nutrition label. A hospital could use that information to judge which evidence applies to its patient population and where a local evaluation may be needed. The comparison also gives committees a concrete document to ask for before committing to a product.

    During Khan’s Q&A, Schabacker gave an example from other medical technologies. Performance can vary with factors such as skin color, and a hospital needs to understand whether evidence from one population travels to another. He advised buyers to ask who else uses the device, consider whether those settings are comparable and seek an independent assessment where possible. A reference site with similar patients, clinical staffing and workflows can help a committee decide which questions to carry into local testing.

    That approach is consistent with ECRI’s position paper, which calls for organizations to define the outcome they expect from AI, assess the risk in the working environment and monitor performance after implementation. The details of the new FDA contract came from Schabacker’s remarks. The agency has separately outlined its work on real-world performance and sought input on how organizations evaluate devices after deployment.

    Assigning the authority to intervene

    Schabacker questioned how much protection a nominal human review provides. His example involved AI-assisted interpretation of an X-ray. An organization can require a clinician to look at the image, but a brief glance cannot tell the organization how often reviewers catch consequential errors. The hospital needs to know what the reviewer is supposed to check and how it will audit whether that check works.

    “Who has the authority?” he asked, describing what health systems should settle before introducing a tool. In practice, that question reaches beyond the clinician using it. Somebody needs responsibility for an immediate override, the decision to suspend use across a unit, and the decision that the system is safe to bring online again. A protocol can define who takes each action, while drills and incident reviews can test whether that protocol works when staff are under pressure.

    Those decisions carry added weight with robotics. A faulty recommendation may enter a clinical workflow as information that someone can call into question. When software directs an instrument already inside a patient, staff may have only a short window to recognize that an action is wrong and intervene. Schabacker used the distinction to explain why a hospital should calibrate its reliance on a tool to its purpose, how it is deployed and the consequences of failure.

    What hospitals can learn after launch

    Schabacker told Khan that some problems take time to surface because individual medical devices may be used too infrequently for a premarket study to reveal them. He asked health system leaders to press vendors on what they measure after launch, how they record adverse events and what they do with the reports. A device can also encounter a changing patient mix or different conditions over time, making continued evaluation part of the job even when its initial local testing goes well.

    The FDA’s draft lifecycle guidance includes recommendations on the information developers provide in marketing submissions for AI-enabled device software functions. Its public device list describes U.S. marketing authorizations and explains how the agency identifies entries. Schabacker focused on what a hospital can observe in its own deployment, where staff use devices under conditions that can differ from those represented in a submission.

    Shared monitoring also gives hospitals a reason to decide in advance what they will record locally. A device malfunction, a clinician override and an uneventful procedure answer different questions. A useful evaluation needs a way to relate events to how often, where and on whom the device was used. The framework Schabacker described is still being developed, leaving those measurement choices open.

    A hospital participating in a shared evaluation could give investigators evidence about a device’s performance across more sites and uses than one institution could see. The first decisions, however, sit with the hospital. Schabacker’s questions for buyers point to a practical sequence: check which patients and settings the vendor tested, decide who can intervene during a failure, and agree on what evidence the hospital will collect once clinicians start using the device. That record can help the next team reviewing a purchase ask more precise questions before bringing the same tool into care.