US Health Agency Launches $62.7M AI Heart Care Initiative

The US government has awarded contracts to build a new class of artificial intelligence designed to act as an autonomous member of clinical care teams. The Advanced Research Projects Agency for Health (ARPA-H) is funding the ADVOCATE program with a total budget of $62.7 million over four years. The primary goal is to create a reliable system that supports patients with heart failure between doctor visits, potentially reducing preventable hospitalizations.

Proponents estimate that successful deployment could generate $28 billion in annual savings within the heart failure population alone. However, this ambitious target comes with a strict condition: the system must achieve FDA authorization as a medical device within 24 months. This requirement highlights the regulatory hurdles facing autonomous AI in critical healthcare settings.

Patient-Facing AI Agents

Three teams have been selected to develop the patient-facing component of the system. Atman Health is building a voice-first interface that adjusts diagnostic questions in real-time based on patient responses. Tempus AI is extending its existing health app to include continuous monitoring and deeper clinical analysis when it detects significant health changes.

UpDoc, a collaboration involving OpenAI, Microsoft, and NVIDIA, is developing a system where clinical actions are validated against approved protocols before execution. This approach separates conversational intelligence from clinical authority. UpDoc’s selection is partly due to its existing FDA clearance for a software medical device that uses large language models within electronic health record workflows.

Oversight and Safety Monitoring

To address the risks of autonomous decision-making, Stanford University has been tasked with building a supervisory AI agent. This system will continuously monitor the patient-facing agents after deployment. It uses a multi-stage pipeline to filter outliers and screen for unsafe recommendations, providing a layer of assurance that clinicians and regulators can trust.

This supervisory layer is critical because the FDA requires rigorous safety benchmarks for such systems. The Stanford team’s approach involves generating inspectable rationales for claims, ensuring that the AI’s decisions are not just accurate but also explainable to humans reviewing the data.

Real-World Health System Integration

Duke University and Kaiser Permanente will lead the implementation phase, focusing on diverse care settings including rural and underserved communities. Duke will validate the technology across five health systems using different electronic health record platforms. Kaiser Permanente will deploy the agents across 21 medical centers and over 260 clinics, embedding them into existing workflows.

According to GN technics/ai (en-US), this broad testing strategy is designed to create reusable deployment blueprints. By using shadow-mode deployments and pragmatic randomized clinical trials, these institutions aim to ensure the technology works effectively in the complex, varied environments of real-world healthcare.