Automated Detection of Apical Rocking in Echocardiographic Videos Using Deep Learning

Apical rocking (AR) is a dynamic echocardiographic manifestation of left ventricular (LV) mechanical dyssynchrony, but routine assessment remains qualitative and reader-dependent. This retrospective two-center proof-of-concept study developed and externally tested a video-based deep-learning pipeline for automated AR detection in a selected cohort of 351 patients with heart failure undergoing cardiac resynchronization therapy (CRT) evaluation. The framework first applied a U-Net-based segmentation model with a shape-prior module to delineate the LV endocardium and suppress non-anatomic background information. Segmented video sequences were then classified using a convolutional neural network (CNN) with long short-term memory (LSTM), with a CNN+Transformer model evaluated as a comparator. The segmentation model achieved a mean Dice coefficient of 0.956 in the held-out test set. In the 56-patient external test cohort from the second center, the CNN+LSTM model achieved an accuracy of 0.91, sensitivity of 0.96, specificity of 0.87, and area under the receiver-operating characteristic curve of 0.95. These findings support the technical feasibility of segmentation-constrained spatiotemporal learning for detecting expert-defined visual AR. Because the classification cohort was enriched for AR and excluded poor-quality examinations, the results should not be interpreted as performance in an unselected screening population. The study did not evaluate CRT response or clinical outcomes. Prospective multicenter validation in consecutive, multi-vendor cohorts that include suboptimal-quality examinations and longitudinal outcomes is required before clinical implementation. Keywords Apical rocking; cardiac resynchronization therapy; deep learning; echocardiography; left ventricular dyssynchrony; video analysis