Real-world performance of AI-guided focused cardiac ultrasound: feasibility, interpretability, and detection of significant heart disease

Introduction

Limited sensitivity of physical examination and restricted access to echocardiography may delay recognition of structural heart disease. Artificial intelligence (AI)–guided focused cardiac ultrasound (FoCUS) may enable non-cardiology clinicians to perform structured bedside cardiac assessment with reduced operator dependency.

Purpose

To evaluate the real-world feasibility, interpretability, diagnostic yield, and learning curve of AI-guided FoCUS deployed at scale in routine hospital care.

Methods

We conducted a multicentre, retrospective, observational study (January 2024–November 2025) across emergency, internal medicine, anesthesiology, and oncology services. Consecutive adult patients undergoing bedside AI-guided FoCUS by non-cardiology clinicians were included. The AI system provided real-time acquisition guidance and automated analysis of six parameters (left and right ventricular function; mitral, tricuspid, and aortic regurgitation; aortic stenosis) from four core views.

The primary outcome was previously undocumented clinically significant structural heart disease, defined as left ventricular ejection fraction ≤40%, moderate-or-greater valvular disease, or moderate-or-greater right ventricular dysfunction. Diagnostic adequacy was defined as successful analysis of ≥5/6 parameters.

Results

A total of 2,767 patients (mean age 67±18 years; 56% men) were examined by 109 clinicians. Median examination time was 9 minutes. AI-generated output was available in 96% of examinations.

Clinically significant abnormalities were identified and suggested in 925 patients (33%), including 523 (19%) with previously undocumented disease. Significant findings included reduced left ventricular ejection fraction in 346 patients (13%), moderate-or-greater mitral regurgitation in 436 (16%), tricuspid regurgitation in 399 (14%), aortic stenosis in 182 (7%), and right ventricular dysfunction in 104 (4%).

Diagnostic adequacy was achieved in 69% of examinations and remained consistent across operator experience levels. After accounting for operator clustering, cumulative examination volume was not significantly associated with adequacy (odds ratio per doubling 1.09; 95% CI 0.97–1.22), indicating limited dependence on operator experience.

Conclusions

In routine clinical practice, AI-guided FoCUS performed by non-cardiology clinicians was feasible, frequently interpretable, and suggested clinically significant—often previously undocumented—structural heart disease in one-third of patients. Diagnostic performance was broadly consistent across experience levels, supporting the scalability of AI-guided FoCUS as an adjunct for early recognition and triage.