Neural Network-Based Prediction of Residual Paravalvular Leak in Bicuspid Aortic Valve TAVI Using CT-Derived Anatomical Features

Figure 1. Study Flowchart.

Figure 2. Architecture of the multi-modal deep learning model (Model B) for predicting residual paravalvular leak. Q, query; K, key; V, value (inputs to the cross-attention module); MLP, multilayer perceptron. Arrows indicate the direction of data flow.

Figure 3. Receiver operating characteristic curves of Model A and Model B for predicting residual ≥moderate PVL after BAV-TAVI. The blue line represents Model A and the orange-red line represents Model B. The diagonal dashed line represents the reference line for a no-skill classifier (AUC = 0.5). The shaded bands around each ROC curve denote ±1 standard deviation across the 5 cross-validation folds.

Figure 4. Calibration curves of Model A and Model B for predicting residual ≥moderate PVL after BAV-TAVI. Calibration was assessed on pooled out-of-fold predictions from 5-fold cross-validation. Model A (blue) yielded a Brier score of 0.141 ± 0.015, close to the no-information benchmark of 0.139 expected for a model predicting only the cohort event rate (16.7%). Model B (red) yielded a Brier score of 0.106 ± 0.015, indicating better-calibrated probability estimates. BAV, bicuspid aortic valve; PVL, paravalvular leak; TAVI, transcatheter aortic valve implantation.

Figure 5. Grad-CAM for a representative bicuspid aortic valve patient with residual ≥moderate paravalvular leak predicted by Model B. Left panel: coronal reference view of the aortic root, with horizontal lines indicating the three axial levels displayed on the right. For each level, the original CT slice (left), the corresponding Grad-CAM heatmap derived from the final convolutional layer of the 3D ResNet encoder (middle), and the heatmap overlaid on the CT image (right) are shown. Warmer colors (red) denote voxels with greater contribution to the model’s prediction of residual ≥moderate PVL, whereas cooler colors (blue) denote minimal contribution. (A) Ascending aortic level: model attention is concentrated on the lumen and wall of the dilated ascending aorta. (B) Aortic annular level: attention localizes to the annulus and adjacent leaflet calcification, the principal anatomical interface for prosthesis sealing. (C) Left ventricular outflow tract (LVOT) level: attention extends into the LVOT and sub-annular region, where calcium protrusion and outflow geometry are known to compromise skirt apposition. Collectively, the spatially coherent activation across the ascending…

Figure 6. Decision curve analysis of Model A and Model B for predicting residual ≥moderate PVL after BAV-TAVI. Model A (blue line; conventional approach with Random Forest classifier) demonstrated a positive net benefit at threshold probabilities below approximately 0.27, falling below the “Treat None” line at higher thresholds. Model B (red line; multi-modal deep learning) maintained a positive net benefit across a broad range of threshold probabilities (0.01–0.80), consistently exceeding both reference strategies. At threshold probabilities above approximately 0.13, Model B provided a greater net benefit than Model A, with the advantage becoming increasingly pronounced at higher thresholds. The grey dashed line represents the “Treat All” strategy (intervene on all patients), and the black solid line represents the “Treat None” strategy (intervene on no patients; net benefit = 0). BAV, bicuspid aortic valve; PVL, paravalvular leak; TAVI, transcatheter aortic valve implantation.