ADVOCATE: The Round-the-Clock AI Cardiologist the FDA Is Backing

In January 2026, a federal agency still little known to the public, ARPA-H, launched a program with an almost earnest name: ADVOCATE, short for Agentic AI-Enabled Cardiovascular Care Transformation. Its ambition fits in a sentence: to win FDA authorization, within three years, for the first artificial intelligence agent able to deliver continuous cardiac follow-up, around the clock. The selected teams are due to be named this month, in June 2026, with a first cull a year later.

Cardiovascular disease remains the leading cause of death in the United States, close to seven hundred thousand a year. A cardiologist, meanwhile, sleeps, takes weekends off and runs weeks-long waiting lists. ADVOCATE targets exactly that gap: not to replace the specialist, but to stand watch when none is on call. The real question is not whether a machine can read an electrocardiogram, it already can, but what it does to a patient to have a doctor who never closes.

A watch that never pauses

The program breaks into three parts. The first is a patient-facing clinical agent: it talks, monitors signals, adjusts treatment within set bounds, raises a flag when a number drifts. The second is a supervisory agent whose only job is to watch the first and guarantee that it stays safe and effective. The third is a plan to fold the tool into care organizations, so it does not stay a lab demo.

The architecture says something about the intent. The AI is not asked to deliver one isolated diagnosis and vanish; it is asked to accompany over time, between appointments, where the current system leaves a void. A chronic cardiac patient sees a specialist a handful of times a year. The rest of the time, they manage their own symptoms, doses and doubts alone.

That silence is what ADVOCATE wants to fill. Between the March visit and the September one, the agent stays reachable, watches the trends nobody is tracking, and answers at three in the morning when worry rises and no office is open.

What the patient gets back

The most immediate benefit is time. Time saved on trips not taken, on appointments no longer booked for a question settled in two sentences, on the wait spared before a symptom turns into an emergency. For a chronic illness, this daily texture matters as much as the big appointments: it is between them that everything plays out.

Then comes a form of autonomy. A patient who understands their numbers, who sees their blood pressure or rhythm explained in real time, regains a grip on a body that had slipped away. Heart disease has long been lived as a dependence on the specialist's calendar. Continuous follow-up shifts the center of gravity toward the patient, who stops being an occasional visitor to their own health.

Last is the less measurable comfort of availability. Knowing an answer exists at night, on Sunday, while traveling, changes one's relationship to worry. For many chronic patients, the fear of the unexpected episode weighs as much as the episode itself. A permanent watch, even a software one, defuses part of that mental load.

The doctor scores 90 percent, but not always

That leaves reliability, and here the enthusiasm has to slow. The consumer self-assessment tools already on offer post low diagnostic accuracy, on the order of nineteen to thirty-eight percent across study reviews. Urgency triage does better, but with a variability that forbids blind trust.

Recent models blur the picture rather than clarify it. A study by Ada Health with researchers from Brown and University College London compared eight apps with seven general practitioners: none beat the doctors, who held a mean around eighty-two percent. Conversely, on written clinical cases, GPT-4 reached some ninety percent where doctors using the same tool topped out at seventy-six. The machine shines on the clean case and stumbles on the real one, which is rarely clean.

On top of that sits the risk specific to large models, hallucination: a wrong answer delivered with the poise of a right one. In cardiology, that kind of error is not a typo but a delay in care, or worse. ADVOCATE's supervisory agent is precisely ARPA-H's answer to this danger: a second AI to watch the first. It remains to be shown that a software guardrail is enough where a human colleague would hesitate.

The price of availability

A doctor who never closes is also a sensor that never switches off. To keep watch continuously, the agent has to receive an unbroken stream of intimate data: rhythm, pressure, sleep, activity, sometimes location. The comfort of being watched and the unease of being watched are the same thing, seen from two sides. The question is no longer only who treats you, but who holds the minute-by-minute log of your heart, and what they can do with it.

Dependence is the flip side of the promised autonomy. As we hand the agent the job of reading every signal, we unlearn how to listen to our own body, and the practitioner too can lose the feel for gestures a machine performs in their place. A tool that reassures too well ends up making both anxiety and judgment something outsourced.

The regulatory frame adds its own unknown. The FDA is piloting predetermined change control plans that will let model weights be updated without a full resubmission: handy for improving the tool, dizzying once you realize that tomorrow's doctor will not quite be the one authorized today. And availability has a cost, which will decide who gets their own permanent watch and who still falls back on the waiting room.

ADVOCATE is not a promise of a cardiologist in your pocket. It is a state bet on a precise idea: that following a chronic disease is worth as much through its constancy as through its peak of expertise, and that a patient machine can hold that constancy where humans wear out. The bet is serious, the timeline tight, three years to a first authorization.

What will need measuring is not the agent's performance on a textbook case, but what it does to a patient's relationship with their own body. Giving back time and autonomy, or installing a gentle surveillance there is no leaving. The line between the two does not turn on the quality of the algorithm. It turns on what we decide to expect from it.