LLM-Assisted Cardiology: RCT Shows Improved Diagnosis & Management of Rare Cardiac Diseases

A randomized trial examined whether a large‑language model (LLM) named AMIE could augment general cardiologists’ assessments of patients with suspected inherited cardiomyopathies, conditions that usually require subspecialty care.

Study Design and Core Results

Researchers compiled an open‑source, de‑identified dataset of real‑world cases and created a 10‑domain rubric to evaluate diagnosis, triage, and management. Blinded subspecialists compared assessments made by general cardiologists with and without AMIE assistance.

Subspecialists preferred the AMIE‑assisted assessments, noting an 11.2% reduction in clinically significant errors and a 19.6% drop in missed important content, while preserving overall reasoning quality and avoiding extra erroneous information.

General cardiologists reported that AMIE helped in 57.0% of cases, missed clinically significant findings in only 6.5% of responses, and cut assessment time in 50.5% of cases.

Why This Matters

Inherited cardiomyopathies, such as hypertrophic cardiomyopathy, are a leading cause of sudden cardiac death in young adults and remain under‑diagnosed—over 60% of U.S. Patients with hypertrophic cardiomyopathy lack a diagnosis, with many states lacking dedicated centers of excellence.

The cardiology workforce crisis highlighted by the American College of Cardiology underscores the urgency of tools that can extend specialist expertise to generalists, potentially improving early detection and triage.

AMIE achieved these gains using feedback from just nine cases, demonstrating that specialized LLM performance can be attained with minimal data when iterative expert refinement and search capabilities are employed.

Clinical Implications and Cautions

General cardiologists showed high diagnostic precision, yet their management plans benefited most from AMIE, with fewer omission errors. The system also boosted efficiency and confidence, suggesting that LLM assistance can raise overall care quality.

Hallucinations—clinically significant inaccuracies—appeared in 6.5% of AMIE’s outputs, often mild (e.g., assuming patient sex or fabricating imaging findings). When prompted, AMIE corrected these errors, indicating that oversight by a clinician remains essential.

Limitations include reliance on textual reports rather than raw imaging, exclusion of physical examinations, a single‑center English‑only cohort, and lack of blinding for the participating cardiologists.

Future Directions

The open‑source dataset and validated rubric provide a foundation for further RCTs across other specialties. Researchers may explore AMIE‑like systems in settings with less complete diagnostic work‑ups or diverse patient populations to gauge broader applicability.

Regulatory scrutiny, training to mitigate automation bias, and patient‑centered studies will likely shape the next phase of AI‑augmented cardiology.

Frequently Asked Questions

What was the primary outcome of the trial?

Blinded subspecialists preferred AMIE‑assisted assessments, citing fewer clinically significant errors (‑11.2%) and reduced missed important content (‑19.6%) while maintaining reasoning quality.

How much did AMIE improve efficiency for general cardiologists?

General cardiologists reported reduced assessment time in 50.5% of cases when using AMIE.

Did the study find any safety concerns with AMIE?

Yes; 6.5% of AMIE’s responses contained clinically significant hallucinations, though most were mild and could be corrected when the clinician queried the system.

How might the integration of AI tools like AMIE reshape the future of cardiology practice?