The Advanced Research Projects Agency for Health (ARPA-H) has committed up to $33.7 million in the first year of a four-year, $62.7 million program to build autonomous AI agents that deliver cardiac care to patients regardless of where they live. The program targets a stark gap: nearly half of U.S. counties have no cardiology providers, and more than 200,000 Americans die each year from preventable cardiovascular disease.
The Agentic AI-Enabled Cardiovascular Care Transformation (ADVOCATE) program brings together three tech companies, two universities, and two health systems. Their goal is a patient-facing clinical AI system that can support heart failure patients between healthcare visits and escalate needs to a human care team when required. ARPA-H estimates the approach could generate $28 billion in annual cost savings through reduced preventable hospitalizations.
Three companies build the core AI agent system
Atman Health, Tempus AI, and UpDoc will work as a cohort to develop the clinical AI agent with technical interfaces, safety benchmarks, and evaluation protocols. Within 24 months of the award, the group must submit a first-of-its-kind FDA authorization package for the product. The U.S. Food and Drug Administration's Digital Health Center of Excellence is collaborating on the regulatory framework for this new class of autonomous, patient-facing clinical AI.
"Patient-facing agentic AI is one of the most consequential frontiers in digital health," said Dr. Rick Abramson, director of the FDA Digital Health Center of Excellence. "Getting the regulatory framework right requires close, iterative collaboration between developers, clinicians and the FDA."
Atman Health will build an evidence-based clinical decision engine using large language models and a voice-first interface. Tempus AI will develop continuous patient monitoring through its Olivia patient health app with deeper clinical analysis. UpDoc will design conversational intelligence based on a clinician-built rules system that validates every proposed action against approved protocols before execution. The approach draws on concepts familiar to professionals exploring AI Agent Courses, where autonomous decision-making and safety guardrails are core design challenges.
A supervisory AI watches for unsafe recommendations
Stanford University will develop a disease-agnostic supervisory AI agent that monitors the clinical agents after deployment. The system runs a three-stage pipeline - outlier filtering, rule-based screening, and deep-research auditing - to detect unsafe recommendations and out-of-distribution behavior in real time. It produces what ARPA-H calls "inspectable per-claim rationales," giving the FDA, health systems, and clinicians an ongoing assurance layer and a defined AI accountability mechanism.
Real-world validation across diverse care settings
Duke University will manage a multi-site validation platform testing the ADVOCATE agents across five health systems and rural sites, running on both Epic and Oracle electronic health records. The American Heart Association will help amplify the testing effort. Kaiser Permanente will embed the agents into EHR workflows for heart failure patients through an enterprise-scale deployment across 21 medical centers and more than 260 clinics, creating reusable deployment blueprints.
The Johns Hopkins University Applied Physics Laboratory will independently evaluate ADVOCATE's technical performance and clinical outcomes to validate results and strengthen regulatory confidence.
"This work will change how we practice medicine - and it would not happen without ARPA-H driving the vision for the future," said Dr. Haider Warraich, manager of the ADVOCATE program.
Why this matters for healthcare, science, and research professionals
ADVOCATE represents a concrete test case for deploying autonomous clinical AI at scale, with a regulatory framework built in parallel rather than retrofitted. For researchers and clinical informaticists, the program's architecture - supervised autonomy with inspectable rationales and independent evaluation - offers a template for safety-critical AI in medicine. The 24-month FDA submission timeline also sets a practical benchmark for how quickly AI for Healthcare Courses will need to evolve as agentic systems move from research to regulated products. The independent evaluation by Johns Hopkins APL, separate from both developers and deployers, establishes a model for building regulatory confidence that other clinical AI programs will likely follow.
Your membership also unlocks:
