The diagnostic accuracy of left ventricular ejection fraction assessment between visual and artificial intelligence-based algorithms on bedside ultrasound.
Authors
Affiliations (8)
Affiliations (8)
- Division of Cardiology, CAPITAL Research Group, University of Ottawa Heart Institute, Ottawa, ON, Canada.
- Faculty of Medicine, University of Ottawa, Ottawa, ON, Canada.
- Division of Cardiology, University of Ottawa Heart Institute, Ottawa, ON, Canada.
- School of Epidemiology and Public Health, University of Ottawa, Ottawa, ON, Canada.
- Department of Cardiovascular Medicine, Mayo Clinic, Rochester, MN, USA.
- Division of Cardiology, Tufts Medical Center and Tufts University School of Medicine, Boston, MA, USA.
- Department of Medicine, University of Ottawa, Ottawa, ON, Canada.
- Division of Cardiology, Knight Cardiovascular Institute, Oregon Health & Sciences University, Portland, OR, USA. [email protected].
Abstract
Focused cardiac ultrasound (FoCUS) has become the standard of care for bedside assessments of cardiac function. With the integration of artificial intelligence (AI), there is limited evidence comparing it to bedside visual assessments by experienced users. In our prospective study conducted at Tufts Medical Center in Boston, Massachusetts from December 2020 to March 2022, patients ≥ 18 years requiring a TTE were recruited by convenience sampling. They each underwent FoCUS LVEF classification by AI and bedside visual assessment, with TTE as reference. LVEF was calculated by Simpson's biplane method of disks in AI-FoCUS and TTE, and visual global assessment by the bedside sonographer. Data analysis was completed in October 2025. Our 215 participants had a median age of 63 (IQR 49-73) years with 83 (38.6%) being female. AI-FoCUS assessments showed good agreement with TTE (intraclass correlation coefficient 0.84, 95% CI: 0.80-0.88), while visual assessments had stronger concordance (intraclass correlation coefficient 0.97, 95% CI: 0.96-0.98). Categorization of LV dysfunction severity showed excellent agreement with TTE for both AI-FoCUS (kappa 0.89, 95% CI: 0.84-0.94) and visual estimate (kappa 0.96, 95% CI: 0.92-0.99). For AI-FoCUS, the area under the curve (AUC) for identifying an abnormal LVEF (< 50%) was 0.9779 (95% CI: 0.9591-0.9967) with sensitivity 90.9% (95% CI: 88.1-100), specificity 95.6% (95% CI: 92.6-98.6), positive predictive value (PPV) 0.79 (95% CI: 0.66-0.92) and negative predictive value (NPV) 0.98 (95% CI: 0.96-1.00). For visual estimate, the AUC was 0.9961 (95% CI: 0.9915-1.000) with sensitivity 90.9% (95% CI: 88.1-100), specificity 97.8% (95% CI: 95.7-99.9), PPV 0.88 (95% CI: 0.77-0.99) and NPV 0.98 (95% CI: 0.96-1.00). AI-assisted FoCUS LVEF assessments provide accurate estimates for the presence and severity of LV dysfunction but are outperformed in the latter by experienced bedside users.