TEERAI-Pre: A Multiview Artificial Intelligence Model for Preoperative Assessment of Transcatheter Edge-to-Edge...

Transcatheter edge-to-edge mitral valve repair is a key therapeutic option for patients with severe symptomatic mitral regurgitation at high surgical risk. This prospective study aimed to develop a novel end-to-end deep learning model for preoperative artificial intelligence assessment in transcatheter edge-to-edge mitral valve repair (TEERAI-pre) candidates using multiview, multimodal echocardiography.

TEERAI-pre, a video vision transformer-based classification model, predicts morphological suitability for transcatheter edge-to-edge mitral valve repair from multiview, multimodal echocardiography. A transformer-based feature-level fusion module was designed in TEERAI-pre to integrate multiview, multimodal features for final prediction. An internal data set of 633 patients (7997 transthoracic echocardiographic videos; 766 pulsed-wave Doppler images) was split for 5-fold cross-validation. An external data set of 150 patients (1735 transthoracic echocardiographic videos; 169 pulsed-wave Doppler images) across 2 hospitals evaluated generalizability. Reference standards were provided by 2 experienced valvular cardiologists per international guidelines.

On the internal data set, TEERAI-pre achieved 75.0% accuracy (95% CI, 71.7%-78.4%) for classifying red (unsuitable), yellow (challenging), and green (ideal) zones, with 77.1% precision, 75.5% recall, and 76.2% F1 score. External validation yielded 73.3% accuracy, 74.0% precision, and 74.0% recall. Multiview multimodal integration improved performance. Binary classification (red versus green) showed TEERAI-pre matched senior experts and outperformed intermediate/junior echocardiologists. Feature-level fusion outperformed output-level fusion and single-view model. Backbone selection and calibration analysis confirmed robust performance.

TEERAI-pre demonstrates strong performance in transcatheter edge-to-edge mitral valve repair preoperative assessment using transthoracic echocardiographic videos and images, supporting more accurate patient selection and enhancing clinical workflow efficiency.

URL: clinicaltrials.gov; Unique Identifier: NCT05508438.