Fully automated artificial intelligence-based echocardiographic analysis substantially reduces workflow time while
Jonghee Sun1, Yeonyee E Yoon2,3, Jiyeon Lee1
1Cardiovascular Center and Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam, Republic of Korea.
Background:
Transthoracic echocardiography (TTE) requires time-intensive integration of quantitative measurements and qualitative visual assessment. Fully automated artificial intelligence (AI)-based analysis may reduce total analysis time while preserving accuracy, but systematic real-world validation remains limited.
Methods:
This prospective, single-center pilot study enrolled 40 TTE examinations. Identical deidentified DICOM datasets were independently provided to a trained cardiac sonographer and a fully automated AI system comprising quantitative and qualitative visual interpretation modules. All outputs were compared with a cardiologist-adjudicated reference standard. Primary endpoints were total analysis time and noninferiority of AI-derived left ventricular ejection fraction (LVEF) versus the reference standard, with a prespecified margin of 3 percentage points (one-sided α = 0.025).
Results:
Median analysis time was 94 s (interquartile range [IQR], 82-106 s) for the AI workflow versus 490 s (IQR, 438-626 s) for the human workflow (P < 0.001). AI-derived LVEF met the noninferiority criterion (mean difference, 0.00 percentage points; upper one-sided 95% confidence bound, 1.41 percentage points; P < 0.001), with an intraclass correlation coefficient (ICC) of 0.902 (95% confidence interval, 0.822-0.947). ICCs for secondary quantitative indices ranged from 0.625 to 0.989. For aortic regurgitation severity grading, AI's overall accuracy was 75.0% (quadratic weighted κ = 0.762), compared with 82.5% for human interpretation (κ = 0.812, McNemar P = 0.579).
Conclusions:
Fully automated AI-assisted TTE analysis substantially reduced total analysis time while maintaining noninferior LVEF accuracy and acceptable performance across secondary quantitative and qualitative indices. These findings support the use of AI as a practical workflow accelerator in routine echocardiography.
