Artificial intelligence-guided detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a multicentre study.

Summary

A video-based, single-view-capable AI model applied to cardiac POCUS discriminated hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy with AUCs ~0.90–0.97 across two health systems and flagged cases a median of ~2 years before clinical diagnosis. In patients without known cardiomyopathy, higher AI scores independently predicted mortality over a median 2.8 years, supporting opportunistic screening with simple POCUS acquisitions.

Key Findings

• Single-view AI on POCUS discriminated HCM and ATTR cardiomyopathy with AUCs ~0.90–0.97 across independent health systems.
• AI-positive screens preceded clinical diagnosis by a median 2.1 years (HCM) and 1.9 years (ATTR).
• Among 25,261 individuals without known cardiomyopathy, top-quintile AI scores for HCM and ATTR associated with higher adjusted mortality (HR 1.17 and 1.39, respectively).
• Model used a multi-label video CNN with view-quality–weighted loss to adapt to POCUS variability.

Clinical Implications

Emergency and community settings could deploy AI-assisted POCUS to triage patients for confirmatory imaging, genetics or biopsy, initiate earlier disease-modifying therapy (e.g., ATTR therapies), and identify high-risk individuals for closer follow-up.

Limitations

• Retrospective design with potential selection bias and lack of prospective clinical impact trial
• Generalizability beyond two US health systems and to handheld devices requires testing

Future Directions

Prospective implementation trials to test clinical pathways, confirm diagnostic yield, outcomes impact, equity/fairness, and cost-effectiveness across diverse care settings.