Regional wall motion abnormality (RWMA) assessment is time-consuming and subject to substantial interobserver variability. Few validated algorithms perform end-to-end 17-segment RWMA assessment integrating all six standard echocardiographic views.
To validate a deep learning algorithm for automated 17-segment RWMA assessment and prediction of significant coronary artery disease (CAD), benchmarked against expert echocardiographers and coronary angiography (CAG).
A 2D deep learning model based on a temporal shift module (TSM) architecture was trained on 22,527 proprietary echocardiographic studies to analyze radial wall thickness change across six standard views, outputting segment-level presence of RWMA and a study-level prediction of significant CAD.
Validation was performed on a retrospective cohort of 115 patients (1,955 segments) with index echocardiograms from September to October 2022 and paired coronary angiograms (CAGs) from a secondary hospital in Incheon, Korea. The per-segment RWMA reference standard was established by majority vote of three senior echocardiographers. Significant CAD on CAG was defined as ≥70% stenosis or fractional flow reserve <0.80. In a blinded reader study, three additional echocardiographers independently assessed segment-level RWMA and predicted significant CAD.
Prevalence of any RWMA on echocardiography was 41.7% (48/115); significant CAD on CAG was 52.2% (60/115).
For significant CAD prediction, the AI showed comparable sensitivity (0.55 [0.42–0.68] vs 0.56 [0.45–0.67], p=0.828) and specificity (0.56 [0.43–0.69] vs 0.50 [0.40–0.61], p=0.320) relative to expert echocardiographers.
The AI model achieved F1 score comparable to expert echocardiographers for RWMA detection and significant CAD prediction. These findings support the potential of AI as an automated screening tool for wall motion analysis, particularly in settings with limited access to experienced readers.