AI-assisted handheld echocardiography for hospital bedside cardiac triage: The prospective OPTIMUST implementation study

BACKGROUND: Timely echocardiography remains difficult because expertise is concentrated within cardiology laboratories. Artificial intelligence (AI)-assisted handheld ultrasound (HUD) could expand bedside cardiac imaging, but prospective implementation data integrating AI, structured training, digital workflows, and expert governance are lacking. METHODS: OPTIMUST was a prospective, single-center implementation study evaluating AI-assisted HUD across cardiology and non-cardiology wards. Physicians without formal echocardiography certification completed a structured two-month curriculum and performed focused examinations using Caption AI-enabled HUD integrated into institutional archiving and electronic medical records. Co-primary endpoints were implementation feasibility (analyzable examinations) and interpretive agreement for left ventricular ejection fraction (LVEF) and filling-pressure categories between ward operators and centralized expert review of the same handheld image sets. Secondary endpoints included image quality, physician-reported clinical impact, downstream referral, and workflow integration. RESULTS: Among 287 attempted examinations, 206 (71.8%) were analyzable; 50 (17.4%) were non-analyzable and 31 (10.8%) lacked a complete report. Among analyzable studies, image quality was excellent in 21%, sufficient in 34%, and suboptimal but interpretable in 45%. Operator assessment correlated with expert review of the same handheld images for LVEF (r = 0.84); Bland-Altman limits of agreement were approximately -21 to +19 percentage points. Filling-pressure category agreement yielded a quadratic weighted κ of 0.659. Physicians reported that HUD findings changed or confirmed management in 93.7% of examinations; this outcome was not independently adjudicated. Comprehensive echocardiography was performed within one month after HUD in 32.8%. Digital integration enabled centralized archiving, structured reporting, remote review, and quality assurance. CONCLUSIONS: A governed AI-enabled HUD pathway was feasible in a tertiary hospital and provided interpretable studies in 71.8% of attempts. The study demonstrates implementation feasibility and agreement between operator and expert interpretation of the same handheld images; it does not establish safety, diagnostic accuracy versus comprehensive echocardiography, or an independent effect of AI. Multicenter studies with reference-standard imaging, safety endpoints, objective clinical outcomes, and non-AI comparators are required.