Preprocedural computed tomography angiography for prediction of transcatheter aortic valve implantation outcomes. A machine learning study

Abstract

Abstract Background Prediction of outcomes following transcatheter aortic valve implantation (TAVI) is challenging. While over the past decades multiple scores for risk stratification in TAVI patients have been proposed due to their limited accuracy none of them have been adopted clinically. Recently machine learning was leveraged for this purpose yet again, while new approaches showed promise in prediction of outcomes following TAVI this was only the case for models which included peri-procedural data. Purpose Considering that in aortic stenosis outcomes are governed by both valve degeneration and myocardial adverse remodelling which ensues, we aimed to employ the pre-procedural computed tomography (CT) to develop a machine-learning model for prediction of 1-year outcomes following TAVI. The model was developed on data of consecutive patients who underwent TAVI at a high-volume center between January 2016 and January 2022. Machine learning by extreme gradient boosting was trained and tested with repeated 10-fold hold-out testing using CT-derived volumetric measurements including myocardial mass and quantitative fibrocalcific valve volume and attenuation measured using standardized software. We compared our CT model based on 21 imaging variables, age and gender with the EuroSCORE II for prediction of 12-month all-cause mortality. External validation was performed on unseen data of consecutive patients who underwent TAVI in 2019-2020 at an independent high volume TAVI center. Results The derivation cohort included 631 consecutive patients (48% men, 80±8 years old, EuroSCORE II 6.5 [4.6-10.3] %). Our model was externally tested on data of 226 patients (46% men, 78±8 years old, EuroSCORE II 7.4 [4.9-10.8] %). Over the 12 months of follow-up 86 (13%) and 29 (13%) subjects died in the derivation and validation cohort respectively. In external validation the CT-based machine learning model had an area under the receiver operator curve (AUC) of 0.70 [0.63-0.76] which was superior to the EuroSCORE 0.59 [0.53-0.66] and age 0.58 [0.52-0.65], p<0.001 for a difference. The feature importance for the machine learning model are presented in Figure 1; the highest-ranked features included LV ejection fraction and mass, the noncalcific valve volume and the mean aortic valve calcification attenuation. Conclusions Machine-learning facilitates leveraging preprocedural CT angiography for prediction of TAVI outcomes. By combining CT derived cardiac assessment machine-learning provides superior prediction to EuroSCORE II.