Predicting mortality after transcatheter aortic valve replacement using preprocedural CT

Abstract

Transcatheter aortic valve replacement (TAVR) is a widely used intervention for patients with severe aortic stenosis. Identifying high-risk patients is crucial due to potential postprocedural complications. Currently, this involves manual clinical assessment and time-consuming radiological assessment of preprocedural computed tomography (CT) images by an expert radiologist. In this study, we introduce a probabilistic model that predicts post-TAVR mortality automatically using unprocessed, preprocedural CT and 25 baseline patient characteristics. The model utilizes CT volumes by automatically localizing and extracting a region of interest around the aortic root and ascending aorta. It then extracts task-specific features with a 3D deep neural network and integrates them with patient characteristics to perform outcome prediction. As missing measurements or even missing CT images are common in TAVR planning, the proposed model is designed with a probabilistic structure to allow for marginalization over such missing information. Our model demonstrates an AUROC of 0.725 for predicting all-cause mortality during postprocedure follow-up on a cohort of 1449 TAVR patients. This performance is on par with what can be achieved with lengthy radiological assessments performed by experts. Thus, these findings underscore the potential of the proposed model in automatically analyzing CT volumes and integrating them with patient characteristics for predicting mortality after TAVR.

Figure 1

(a) Example image slice with oblique orientation showing the locations of the five aortic landmarks. The image slice is defined by the plane containing landmarks 1, 3, and 5. The dots for landmarks 2 and 4 are projections on that plane. (b) 3D schematic of the region of interest (ROI) extraction from the CT volume. The landmarks (yellow dots) are interpolated with a smoothing spline (black curve), from which orthogonal image slices are extracted. For visualization purposes, only seven slices are shown in this figure. The extracted image slices are simply stacked to obtain the ROI. (c) Cross-sectional views through the center of the ROI.

Figure 2

Directed acyclic graph expressing the conditional dependence structure of the TAVR data. A are the tabular characteristics, $f(I;\omega )$ are automatically extracted image features, J are manual image measurements, and Y is the outcome. Arrows indicate a dependency, e.g., Y only depends on A and $f(I;\omega )$.

Figure 3

Overview of a conventional approach (left) and our automatic approach (right) for predicting TAVR outcome. In the conventional approach, radiologists manually extract measurements from the CT image, which are combined with the tabular characteristics A for risk assessment. Missing variables in A (shown in orange) are simply deleted. In our approach, task-specific image features $f(I;\omega )$ are automatically extracted from the CT volume through landmark localization, ROI extraction, and a 3D neural network. According to the hierarchy of our probabilistic model, the tabular patient factors A can also influence those image features $f(I;\omega )$. Missing values in A are marginalized. Both A and $f(I;\omega )$ are used to predict postprocedural mortality. In addition, our model leverages available manual image measurements J as auxiliary outputs to help guide the training of the network parameters $\omega$. The dashed arrow indicates that J is not required during inference.

Figure 4

Predictors sorted by their importance for our probabilistic model. Withholding the “CT image” input (both the unprocessed image I and the measurements J) during evaluation leads to the largest drop in AUROC, indicating high importance. The “Age” predictor is second-most important after “CT image”. Bars indicate the standard error of the mean estimate.