Listen instead
A patient diagnosed with chronic heart failure typically requires a regimen of four distinct classes of life-saving medications, a protocol known as guideline-directed medical therapy. To stabilize heart function, these drug doses must be carefully adjusted, or titrated, every two weeks based on renal function, blood pressure, and potassium levels. Yet, in nearly half of all United States counties, patients live without access to a practicing cardiologist to manage this clinical process. The burden of this highly repetitive, data-driven titration falls on primary care networks that are often too understaffed to perform the frequent laboratory checks and dose adjustments required. This care gap contributes to high rates of preventable hospitalizations and progressive cardiac decline.
To address this physician shortage, the Advanced Research Projects Agency for Health, or ARPA-H, has launched a clinical artificial intelligence initiative under its ADVOCATE program. The agency has committed up to 33.7 million dollars in first-year funding toward a four-year, 62.7 million dollar initiative. This program represents a coordinated federal effort to build and validate autonomous software systems capable of managing chronic heart failure patients directly, moving beyond passive clinical administrative tools.
Unlike current generative artificial intelligence systems in healthcare, which focus on transcribing doctor-patient conversations or summarizing medical charts, these new tools will have an active clinical scope. The selected performers, including Atman Health, Tempus AI, and Updoc, are tasked with developing autonomous software agents. These systems will assess clinical symptoms, order necessary laboratory tests, titrate guideline-directed medical therapy, and generate prescription orders. These actions will be executed under structured clinical safeguards to protect patient safety.
Building a multi-layered safety architecture
Transitioning from passive clinical documentation to autonomous prescribing requires rigorous clinical validation and multi-layered safety guardrails. To manage the risks associated with automated clinical decisions, the ADVOCATE program relies on a collaborative framework between academic institutions and established health systems.
Stanford University will construct a three-stage supervisory validation layer to oversee the AI agents. This safety architecture combines deterministic checks, predictive models, and independent software review to verify every medical decision before it reaches a patient or pharmacy. Deterministic checks ensure the software adheres strictly to established clinical guidelines, such as contraindications for specific drug classes. Predictive models assess the patient's likelihood of adverse drug reactions based on historical data. The final review layer provides an independent cross-reference to identify potential errors or anomalies in the treatment plan.
Once these safety layers are established, the systems will undergo prospective deployments in real-world clinical environments. Duke University Health System and Kaiser Permanente will lead these clinical trials across their respective healthcare networks. By testing the agents across diverse patient populations and distinct electronic health record systems, these institutions aim to gather the empirical evidence required to prove that autonomous software can manage complex medication regimens safely.
Bridging the specialist shortage
The clinical target for this program is guideline-directed medical therapy, which is the established standard of care for heart failure. Clinical trials have demonstrated that achieving target doses of these medications improves patient survival rates. Despite clear evidence that maximizing these doses saves lives, only a small percentage of heart failure patients are successfully titrated to these optimal levels. The titration process requires frequent blood tests and dose adjustments, a schedule that traditional outpatient clinics struggle to maintain due to persistent staffing constraints.
By automating the titration process, ARPA-H is testing whether algorithmic care can deliver specialist-level management to underserved populations. Software agents can monitor patient-reported symptoms, review lab results, and adjust prescriptions continuously. This approach could prevent avoidable hospitalizations at a broader population scale.
This initiative represents a shift in medical technology regulation. Under the terms of the ADVOCATE awards, all patient-facing software contractor teams must submit a completed Food and Drug Administration authorization package within 24 months of their award. The federal government is collaborating with regulatory bodies to establish a framework for autonomous clinical agents, setting the guidelines necessary to integrate these systems into the national healthcare infrastructure.
Source: ARPA-H
