Abstract
Background
A key mechanism and major predictor in the progression of aortic stenosis (AS) is the aortic valve calcification (AVC). Computed tomography (CT) aortic valve Agatston calcium score (AVCS) remains the gold standard for quantifying AVC.
Purpose
Develop deep learning models to quantify AVC in echocardiography (TTE).
Methods
From the randomised clinical trial BICATOR [1] and a local study, 439 TTE and their corresponding CT AVCS from 234 patients were included. Data were partitioned at patient level into 165 TTEs (996 videos of parasternal long and short axis and 3-chamber views, from 7 centres) for training and cross-validation, and 274 TTEs (1488 videos, 9 centres) for testing, with 89 studies (434 videos) from two centres isolated from the training set. To evaluate performance on an independent dataset, SALTIRE2 [2] was used, comprising 255 TTEs (2159 videos) from 154 patients with mild or moderate AS. Models were fine-tuned on PanEcho [3] as the backbone, fusing task-specific heads for binary classification and AVCS quantification. Five models per task were obtained via five-fold cross-validation. After outlier removal, the averaged predictions were compared against CT AVCS.
Results
The cross-validation cohort had an age of 53 years (44-72), with 32 females (29%) and 65 patients (59%) diagnosed with bicuspid aortic valve (BAV). AVC was absent in 37 patients (33%), while those with calcification showed an AVCS of 1371 AU (687-3112). The time interval between TTE and CT was 40 days (7-134). In the testing cohort, age was 67 years (48-74), with 61 females (22%) and 126 patients (46%) with BAV. AVC was exhibited in 215 patients (79%), with an AVCS of 1103 AU (530-2067). The interval TTE-CT was 0 days (0-32). Excellent performance was achieved in distinguishing patient with and without AVC, with AUROC values of 0.980 (0.961-0.993) in internal cross-validation (Fig 1), 0.974 (0.62-0.983) in the testing set, and 0.983 (0.972-0.992) in external multicentre validation. The model accurately quantified AVCS in relation to CT AVCS, yielding R=0.89 with an error of 116 AU (0-551) during cross-validation, and retained robust performance in the testing cohort (R=0.64; error: 308 AU [3–977]) and external validation (R=0.47; error: 669 AU [184–1191]). Analysing baseline AVCS by TTE, a correlation was observed with the progression of AS severity measured by changes in maximum AV velocity, pressure gradient and CT AVCS (Spearman test, all p<0.001). Predictions also demonstrated significant discrimination for future valve replacement (AVR), both in terciles stratification and after the identification of sex-specific cut-offs (Youden index: 1052 AU females, 1343 AU males; Fine and Gray model with mortality as competing risk, p<0.001) (Fig 2).
Conclusions
A deep learning model accurately estimated AVCS from TTE with strong generalizability. TTE-based AVCS predicted aortic stenosis progression and valve replacement, supporting its role in risk stratification.
Fig 1.
AUROC AVCS>0 by Echocardiography
Fig 2.
Cumulative incidence of AVR.