Abstract
Background: It is currently unclear whether various measurements of the aortic valvular complex routinely performed on pre-operative computed tomography angiograms (CTA) can collectively predict the risk of conduction disturbances after transcatheter aortic valve replacement (TAVR) for aortic stenosis (AS). Here, we aimed to use unsupervised machine learning to analyze a multitude of CTA features and identify distinct patient sub-phenotypes with different risks for TAVR-related conduction disturbances. Methods: The pre-TAVR CTAs of 660 AS patients (330 males, 330 females) were analyzed using TeraRecon to extract 21 features that included aortic valve leaflet calcification loads, dimensions of the aortic root (annulus, sinus of Valsalva, sinotubular junction) and the ascending aorta, coronary ostial heights, and aortic angle. Agglomerative hierarchical clustering was performed separately on male and female datasets using R, with assessment of clusterability based on Hopkins statistics and agglomerative coefficient and the optimal number of clusters based on 30 previously validated indices (NbClust package). Multivariable logistic regression was conducted to assess the dependence of significant conduction disturbances (new left bundle branch block, advanced atrioventricular block) on cluster type. Results: Both male and female datasets were fairly clusterable (Hopkins statistics, 0.70 vs 0.68; agglomerative coefficient, 0.93 vs 0.93), with 3 and 2 as the optimal number of clusters, respectively. Of the 3 male clusters identified, the cluster with intermediate median values of aortic valve leaflet calcification loads and aortic root dimensions (M2) was associated with the greatest risk of conduction disturbances, compared to clusters with low (M1) and high (M3) median values of these variables (OR M2/M1 =9.24, p =0.03; OR M2/M3 =1.33, p =0.40; OR M3/M1 =6.97, p =0.07). The risk of the M1 cluster was not statistically different from those of the female clusters (F1, F2; OR M1/F1 =0.31, p =0.27; OR M1/F2 =0.25, p =0.21), which were not different from each other (OR F1/F2 =0.82, p =0.68). Conclusion: Machine learning of pre-TAVR CTAs can help identify subgroups of patients with similar or varying risks of TAVR-related conduction disturbances.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call