CURL-AID: Automated Echocardiographic Motion Analysis for Quantitative Assessment of Posterior Systolic Curling

Posterior systolic curling (PSC) is a morphofunctional abnormality of the posterior mitral annulus associated with malignant ventricular arrhythmias and sudden cardiac death. Current diagnosis is qualitative and operator-dependent, limiting reproducibility, objectivity and standardization. This study introduces CURL-AID (Curling Ultrasound-based Recognition and Labeling-Automated Intelligence-driven Diagnosis), a fully automated echocardiographic framework for PSC detection through quantitative analysis of posterior annular and left ventricular wall hypermobility.

Parasternal long-axis transthoracic echocardiograms from 100 patients (47 female; median age 58 years) were retrospectively analyzed. Three independent clinicians classified subjects as PSC (n = 40) or noPSC (n = 60) based on dynamic visual assessment. CURL-AID integrates convolutional neural network segmentation of the posterior wall and annulus, point-wise tissue tracking, extraction of interpretable kinematic and dynamic features and supervised machine learning classification. Feature selection combined bootstrapped Elastic Net and Sequential Backward Selection, identifying nine motion descriptors. Model performance was evaluated using repeated, stratified nested cross-validation.

Manual annotations showed high inter-operator consistency. Automated segmentation achieved a median Dice of 0.88 and Jaccard index of 0.79 for the posterior wall, with median centroid differences of 2.0 mm (validation) and 1.6 mm (test) for the annulus. Selected features reflected increased tangential and radial displacement and irregular systolic acceleration in PSC. The linear support vector machine achieved the best balance between accuracy and interpretability (area under the curve 0.865 [interquartile range 0.855-0.875], accuracy 0.82, sensitivity 0.84, specificity 0.82, F1-score 0.79).

CURL-AID provides a quantitative and reproducible assessment of PSC and may support automated identification of a phenotype associated with arrhythmic risk.