Stanford researchers explain how an AI watchdog would check heart-care advice

Stanford researchers have explained how their proposed AI watchdog would check heart-care recommendations—even when clinical trial data is unavailable. In an October 22 Stanford Daily article, they described a system that would turn to past decisions and existing guidelines, then verify patient details and examine treatment reasoning. Called STEWARD, it remains in early development.

More capacity, with a supervisor

The catalyst is a capacity problem. Radiology professor Akshay Chaudhari told the Stanford Daily that healthcare demand exceeds physician capacity. Handling simpler cases autonomously or more efficiently could leave doctors more time for complex patients, he said.

Stanford Medicine’s September 30 announcement committed up to $14.9 million from the Advanced Research Projects Agency for Health, or ARPA-H, to develop STEWARD. Chaudhari co-leads the team with Roxana Daneshjou and Euan Ashley.

The project belongs to ARPA-H’s ADVOCATE program, which aims to develop an FDA-authorized clinical AI system for round-the-clock cardiovascular care. STEWARD’s role would be to monitor clinical agents and provide reports to care teams.

Three checks on a treatment recommendation

1. Check the evidence fit. STEWARD would examine whether a patient’s information and an agent’s recommendation are comparable to clinical trial data. Past decisions and existing guidelines would supply the fallback.
2. Verify the supporting information. The second stage would check details behind the recommendation, including the patient’s medical history, medication dosage and drug interactions.
3. Examine the decision. The third stage would evaluate the clinical agent’s reasoning and treatment choices, then provide feedback intended to improve accuracy over time.

Ashley offered changing drug doses as one possible application. Patients respond differently, he said, so several recommendations may be acceptable. STEWARD’s task would be to detect incorrect options rather than select one uniquely correct answer.

But our focus is on the smaller number of potential wrong answers and detecting those.

Euan Ashley, speaking to the Stanford Daily

Clinical testing comes next

Researchers describe STEWARD as an assistive tool, not a plan to automate the medical process. Before implementation, the team must conduct clinical studies to test reliability and work with the Food and Drug Administration.

A further milestone would be Medical Device Development Tool status, Chaudhari said. He described that designation as a route for STEWARD to become a generic framework for other monitoring systems, supporting broader use.