Sovereign Execution: Vigilant, Gated and Verifiable Agentic Care Orchestration - Title: Sovereign Execution: Vigilant, Gated and Verifiable Agentic Care Orchestration Area 1: Accelerate the use of agentic artificial intelligence solutions in clinical care and/or clinical trials Applicant Name: MedStar Health Research Institute Physical Address: 10980 Grantchester Way, Columbia, MD 21044 Contact Name: Kristen E. Miller, DrPH Contact Phone: 574-527-6685 E-Mail Address: Kristen.e.miller@medstar.net Web Address: https://www.medstarhealth.org/innovation-and-research/national-center-for-human-factors-in-healthcare The most dangerous moment in clinical care is often the interval between a provider's recommendation and the patient's follow-up. Virtually every clinical encounter generates follow-up instructions: see a specialist, complete a lab test, obtain imaging. Yet follow-through fails at catastrophic rates. Adherence to Lung-RADS-recommended follow-up windows averages only 57 percent nationally, meaning 43 percent of recommended follow-ups are missed, and one large integrated system reported follow-up failure as high as 72.4 percent in lung cancer screening. Roughly 80 percent of healthcare data is unstructured, and follow-up instructions typically live in that text, invisible to traditional reminder systems where care continuity can be lost- particularly when crossing organizational boundaries. This project delivers the Sovereign Execution Bridge, a first-of-its-kind, fully open-source middleware harness that safely orchestrates complex clinical follow-ups by unifying three foundational capabilities: Durable Execution (persistent, vigilant, long-running care processes that survive the waits, retries, and handoffs that break traditional workflow engines), Sovereign Data (patient-gated data pods built on the Solid protocol created by the inventor of the World Wide Web), and Sovereign Lineage (cryptographically verifiable artificial intelligence (AI) routing, so every agentic decision produces a portable, tamper-evident certificate of what was done and why). The work proceeds in three phases. First, the American College of Emergency Physicians (ACEP) will derive a national taxonomy of clinical follow-ups from 1.3 million encounters processed for the Centers for Medicare & Medicaid Services (CMS) Merit-based Incentive Payment System (MIPS) program, drawn from emergency departments nationwide, with follow-up recommendations already pre-extracted by production large language model pipelines and 1.4 million additional records arriving during the project. Second, MedStar Health's National Center for Human Factors in Healthcare will conduct shadow-mode "Training Wheels" evaluations, riding along with the agent through simulated journeys and measuring incomplete actions, routing disruptions, permission-gating accuracy, and transformation deviations against non-AI baselines. Third, a supervised real-world pilot of 40 to 60 patient journeys will run each follow-up as a three-party ride: the patient, a human factors coordinator, and the AI agent, with patients gating every action through simple micro-authorizations. The project spans the data formats from multiple major electronic health record platforms (Epic, Oracle Health, Meditech, and others) sampled from 800+ emergency department systems that send data to ACEP to prove the architecture generalizes across institutions. Anticipated products include 100% open-source middleware with complete documentation, a national follow-up taxonomy, a published risk-management assessment report, a national educational webinar within 12 months, and at least one peer-reviewed publication. A technical expert panel of approximately ten stakeholders, including patients, practicing clinicians, and health information technology developers, advises throughout. If successful, this project moves the healthcare ecosystem beyond passive patient data access and into vigilant, gated, and verifiable agentic care orchestration, giving patients sovereign control over the AI executing their care.
