Artificial intelligence for the assessment of aortic valve calcification in echocardiography

Abstract Background Aortic valve calcification (AVC) is a fundamental mechanism and key predictor of aortic stenosis (AS) evolution. The gold standard for AVC quantification is computed tomography (CT) aortic valve Agatston calcium score (AVCS). Purpose To quantify AVC in echocardiography (TTE) using deep learning models. Methods 439 TTE and their corresponding CT AVCS from 234 patients included in the BICATOR 1 and a local study were partitioned at patient level into 165 TTEs (996 parasternal long and short axis and 3-chamber videos, from 7 centres) for training and cross-validation, and 274 (1488 videos, 9 centres) for testing, maintaining 89 studies (434 videos) from two centres exclusive for testing. Additionally, an external dataset, SALTIRE2 2 comprising 255 TTEs (2159 videos) from 154 patients with mild or moderate AS was used to test model performance on independent data. Models were fine-tuned from PanEcho 3, with task-specific heads for binary classification and AVCS quantification. Five-fold cross-validation on training produced five models per task. After outlier removal, their averaged predictions were compared to reference CT AVCS. Results In the training cohort, median age was 53 44; 72 years, with 32 (29%) females while 65 (59%) patients had a bicuspid aortic valve (BAV). AVC was absent in 37 (33%) patients, and those with calcification had a median AVCS of 1371 687; 3112 AU. The median TTE to CT was 40 7; 134 days. The testing cohort had a median age of 67 48; 74 years and included 61 (22%) females and 126 (46%) patients with a BAV. AVC was present in 215 (79%) patients with a median AVCS of 1103 530; 2067 AU. The cohort TTE-CT interval was 0 0; 32 days. The model achieved excellent accuracy in distinguishing patients with and without AVC in the internal cross-validation (AUROC of 0.980 0.961-0.993, figure 1), testing set (0.974 0.962-0.983) and in the external multicentre validation (0.983 0.972-0.992). AVCS quantification model performed well in cross-validation (R=0.89; error: 116 0-551 AU) and remained robust in the testing set (R=0.64; error: 308 3-977 AU) and external multicentre validation (R=0.47; error: 669 184-1191 AU). Baseline AVCS predictions were significantly correlated with future progression of AS severity, as measured by changes in AV maximum velocity, pressure gradient, and CT AVCS (all p0.001). Baseline AVCS by TTE demonstrated excellent discrimination for future valve replacement (AVR), either when predictions were stratified by terciles and after selecting sex-specific optimal thresholds (Youden index, 1343 AU for males and 1052 for females) (p0.001, Fine and Gray model with all-cause mortality as competing risk) (figure 2). Conclusions A deep learning model estimates AVCS from TTE with strong performance and generalizability. AVCS by TTE predicts aortic stenosis progression and need for valve replacement, supporting its potential for risk assessment.Fig 1.AUROC AVCS0 by Echocardiography Fig 2.Cumulative incidence of AVR.