← Back to feed

ARPA-H, the Advanced Research Projects Agency for Health, has committed $62.7 million to a program developing agentic AI for cardiovascular care — specifically heart failure, a condition affecting 6.7 million Americans that generates roughly one million hospitalizations per year.

ARPA-H has contracted six performer teams across three technical areas. Technical Area 1 (patient-facing agents) went to Atman Health, Tempus AI, and UpDoc — a clinical AI company founded by Stanford physicians with an existing FDA-cleared LLM medical device. Technical Area 2 (supervisory AI architecture) went to Stanford University. Technical Area 3 (real-world deployment) went to Duke University and Kaiser Permanente. The mandate for the TA1 teams: build a patient-facing agentic system managing heart failure in outpatient settings and submit an FDA pre-market authorization package within 24 months.

What the Program Requires

Unlike a pilot grant or research agreement, this is a performance contract with regulatory milestones. The agentic system must monitor patient vitals and clinical data continuously, initiate care pathways autonomously within defined protocols, communicate directly with patients in natural language, and submit evidence sufficient for FDA pre-market authorization.

Tempus AI brings its oncology-built data infrastructure. The company operates one of the largest structured clinical-genomic datasets in the U.S. and holds multiple prior FDA clearances for AI-enabled clinical software. UpDoc brings an existing FDA-cleared LLM-driven medical device and the clinical workflow infrastructure to host and deliver the deployed agent.

Why Heart Failure

Heart failure is unusually well-suited for agentic AI: the condition is chronic, highly protocol-driven, and has clear physiological signals — weight gain, fluid retention, elevated natriuretic peptide levels — that predict deterioration 2 to 5 days before a crisis. Current management relies on human check-ins that frequently miss the intervention window.

An agentic system that monitors in real time and intervenes — adjusting diuretic dosing recommendations, escalating to physicians, triggering home nurse visits — could reduce hospitalizations meaningfully. Average Medicare reimbursement for a heart failure hospitalization runs $11,000 to $15,000 per admission; the cumulative cost per patient across annual readmissions is substantially higher.

The Regulatory Bet

The FDA authorization requirement is the program’s most ambitious element. Previous ARPA-H AI programs produced research tools. This one targets cleared clinical use. The pre-market authorization pathway for AI-driven autonomous care coordination at this level of autonomy has no established precedent.

ARPA-H is effectively asking contractors to build the regulatory pathway as they build the product — a structure that has worked in defense procurement but has rarely been attempted in health AI.

The program runs 39 months in total. Up to $33.7 million is budgeted for the first year, with the TA1 teams required to reach an FDA pre-market authorization submission within 24 months. Subsequent phases cover real-world deployment and evaluation.

Significance for Agentic AI

Healthcare has moved slowly on agentic deployment. Liability exposure, regulatory uncertainty, and clinician resistance have kept most medical AI in a passive advisory role — surfacing information for a human to act on rather than acting autonomously.

The ARPA-H program is a direct challenge to that status quo. If the contractors deliver and the FDA clears the system, it would establish the first regulatory template for autonomous agentic decision-making in clinical care — a template the rest of the industry can follow.