News
HHS Funds MedStar Project to Put Patients in Control of Clinical AI Agents
An AI agent helping a patient complete a medical follow-up needs more than access to the chart. It also needs permission to act, and a way for people to check what it did.
A federally funded MedStar Health Research Institute project will test how to build those controls into software that coordinates clinical follow-ups. The work puts patient authorization at the center of a research effort to move AI beyond retrieving health information and into carrying out tasks.
The Office of the National Coordinator for Health Information Technology, part of the Department of Health and Human Services, announced the award October 6. It is one of two awards totaling $1.5 million under the Leading Edge Acceleration Projects in Health Information Technology program (LEAP in Health IT).
MedStar’s project, called Sovereign Execution, will develop an open-source software layer known as the Sovereign Execution Bridge. HHS describes it as a system for coordinating complex follow-ups using patient-controlled data access and tamper-evident records of AI routing.
Testing permission alongside performance
The
federal grant record lists MedStar’s award at $999,985. Its proposed work focuses on the interval between a clinician’s recommendation and the patient’s follow-up, such as a specialist visit or laboratory test.
The American College of Emergency Physicians will help derive a national classification of follow-up tasks from emergency department encounter records. MedStar’s National Center for Human Factors in Healthcare will then evaluate the system in simulated patient journeys, comparing its performance with approaches that do not use AI.
Those evaluations will examine incomplete actions and routing disruptions, including whether permission controls work as intended.
A subsequent supervised pilot will cover 40 to 60 patient journeys. Each will involve a patient, a human factors coordinator, and an AI agent. Patients are to authorize every action through what the proposal calls “micro-authorizations.”
The proposed workflow is designed to persist through delays and handoffs. It also calls for portable, tamper-evident certificates documenting what an agent did and why.
That documentation could help reviewers reconstruct an action. It would not, by itself, establish that the action was clinically appropriate.
Data access is part of the design
The
project overview specifies patient-gated data stores, known as pods, built on the Solid protocol.
The Solid specification defines mechanisms for applications to access externally stored data with permissions. Its design separates data storage from the applications that use it, allowing access controls to govern which resources an application can read or change.
For healthcare technology teams, that distinction raises a practical question: Can access to a patient’s information remain limited to the task the patient has authorized, even when a workflow involves multiple steps?
MedStar’s research plan treats that as something to evaluate. The public descriptions establish the intended controls and testing approach; they do not yet establish how reliably the system will operate in routine care.
HHS’s second award went to the Regenstrief Institute for a project addressing terminology and interoperability gaps in small, independent laboratories. That work will assess readiness and data quality, then test improvements in the use of standardized laboratory terminology.
The two projects approach different obstacles to using health data. Regenstrief will examine whether systems exchange information consistently. MedStar will test how software acts on information once it can access it.
For MedStar, the immediate measure of progress will be evidence from the planned evaluations: whether follow-up tasks are completed, whether authorization is respected, and whether people can understand and review the resulting actions. Improved patient outcomes remain a research question, not an established result of the award.