Deep learning for predicting transvalvular gradient outcomes for patients undergoing transcatheter aortic valve replacement for native aortic stenosis.

Objectives: To develop a deep learning-based predictive model for preprocedurally predicting post-transcatheter aortic valve replacement (TAVR) gradient waveform using pre-TAVR echocardiographic information only for candidates for both balloon-expandable and self-expandable THV in a TAVR procedure. Methods: A total of 69 patients (mean age 81 ± 9.04 years, 58% female) receiving Edwards SAPIEN 3 and 77 patients (mean age 85 ± 8.56 years, 43% female) receiving Medtronic Evolut were included for pressure gradient collection. Two deep machine learning models were trained on the cohorts, respectively, each using the paired pre/postprocedural gradient waveform. Results: The SAPIEN model demonstrated an average accuracy of 84% on point-by-point agreement between predicted and clinically measured post-TAVR gradient waveform and an average prediction error of 2.1 and 5 mm Hg in mean and peak gradient, specifically. This group of metrics were reported as 87%, 1.4 mm Hg, and 3.1 mm Hg for the Evolut model. Bland-Altman analyses on both cohorts showed >90% agreement between clinical measurement and prediction for both mean and peak gradients. The SAPIEN model was additionally validated on a prospective cohort (n = 33) with average prediction error of 2.2 and 4.8 mm Hg for mean and peak gradient. Conclusions: A deep machine learning rationale was introduced to infer the full post-TAVR pressure gradient pattern directly from the preprocedural one obtainable through Doppler echocardiogram. It helps guide decision-making for the prevention of various post-TAVR complications. Further studies are necessary to investigate the gradient change of other valve types under specific deployment scenarios in a lifetime timespan.