Title : ECHO-VIEWER: An AI-platform for real-time echocardiography data interpretation in experimental and clinical settings
Abstract:
Background: Transthoracic echocardiography (TTE) remains the mainstay non-invasive modality for cardiac assessment in clinical practice and experimental HF models, with over 7 million U.S. studies/year. However, interpretation is constrained by image quality, foreshortened or incomplete loops, suboptimal endocardial borders, and operator variability. In preclinical studies, small hearts, rapid rates, and subtle pre- and post-therapy remodeling create a diagnostic gap that conventional 2D-TTE cannot consistently resolve through real-time quality control and 3D interpretation.
Aim: To develop and externally validate ECHO-VIEWER, a quality-gated, human-in-the-loop hybrid AI platform for real-time TTE acquisition guidance, functional/structural interpretation, clinician feedback, and 3D visualization across clinical and preclinical workflows.
Methods/Approach: A CNN-based model was trained on 10,000 MIMIC-IV-ECHO/PhysioNet-derived 2D-TTE cine loops and externally validated on 1,276 Stanford echo clips. Before quantitative reporting, predefined quality and cycle-completeness criteria were applied, including LV/endocardial-border visibility, foreshortening, artifact/dropout/cropping, and ED/ES frame detectability. This adjudication identified 1,037 diagnostic-quality clips for primary analysis. Outputs included LVEF, EDV, ESV, fractional shortening, chamber quantification, septal wall-motion review, valvular-abnormality screening, and 3D reconstruction. A clinician-in-the-loop interface enabled expert accept/reject/edit decisions, with corrections archived for CNN refinement. Feasibility was explored in paired pre- and post-stem-cell murine studies (n=12) and in human cases. LVEF was compared with the reference clinical LVEF using MAE, RMSE, Pearson r, Bland-Altman analysis, and ±5- and ±10-point agreement.
Results: In the diagnostic-quality external validation cohort, LVEF estimation achieved an MAE of 2.72 points, an RMSE of 3.24 points, and a strong correlation with the reference LVEF (r=0.958). Bland-Altman analysis demonstrated near-zero mean bias (0.05 points), with 95% limits of agreement from -6.30 to +6.39. Estimates were within ±5 LVEF points in 86.11% and within ±10 points in 100%. The platform translated 2D echocardiographic data into expert-reviewable functional, chamber, septal, valvular-screening, and 3D outputs while mitigating the effects of unreliable reporting due to inadequate or incomplete loops.
Conclusion: In external validation, ECHO-VIEWER demonstrated strong agreement for LVEF estimation and extended 2DTTE into a quality-controlled, clinician-supervised, 3D-enabled workflow for patients and experimental models before and after stem-cell therapy interventions, with future expansion toward M-mode, Doppler hemodynamics, and 4D valve-focused analysis.
