3D CT-based body composition imaging biomarkers improved prediction of all-cause and cardiovascular mortality after TAVI, raising c-index from 0.63 to 0.69 (P<0.001).
Does the addition of fully automated 3D CT-based body composition imaging biomarkers improve the prediction of cardiovascular death and all-cause mortality in patients undergoing TAVI?
Fully automated 3D CT-based body composition analysis provides imaging biomarkers that significantly improve long-term mortality prediction in TAVI patients beyond standard clinical risk scores.
Absolute Event Rate: 0% vs 0%
Abstract Background CT-based imaging biomarkers derived from automated artificial intelligence-based could offer promising opportunities for patient screening across different populations. Purpose In this study, we aimed to develop a fully automated whole-body CT-based body composition analysis tool to extract imaging biomarkers and investigate their added value and association with patient outcomes after transcatheter aortic valve implantation (TAVI). Methods The development of the 3D deep neural network (DNN) model was performed within our nnU-Zoo framework using 1,761 CT scans from two different open source dataset, with reference standard segmentations of various organs. The model enables fully automated segmentation of multiple anatomical structures, including organs, muscular systems (skeletal muscle, iliopsoas, autochthonous muscles), fat compartments (subcutaneous, torso, epicardial, and pericardial fat), and bones (vertebrae, ribs, scapula, sternum, hip, sacrum). Using thisinclude model, we performed automated whole-body composition analysis (volumetrics and intensity) in patients undergoing TAVI in a prospective TAVI registry in our internal centre to quantify key imaging biomarkers and evaluate their prognostic value for clinical outcomes after TAVI. The added value of the image biomarkers was assessed in cause-specific Cox proportional hazards models with and without the body composition measurements for predicting cardiovascular (CV) death and all-cause mortality. Models were adjusted for the following confounding variables: age, sex, body mass index (BMI), and Society of Thoracic Surgeons (STS) predicted risk of mortality score. To select the optimal features for outcome predictions, stepwise model selection with AIC criterion was applied. Added predictive value of the biomarkers was assessed checking the improvement in Harrell’s C-index. Results In the pre-TAVI population, 3064 patients underwent whole body CT scans. In the TAVI candidates with severe aortic stenosis, 3,064 patients underwent whole-body pre-TAVI CT before intervention were included in the analysis. Among them, 1,194 (39.0%) died, including 877 (28.6%) from cardiovascular causes, during a median follow-up of 497.5 days (IQR: 367–1,826). In the evaluation set, the DNN model achieved a mean Dice score of 0.90 across two datasets. The baseline Cox model demonstrated moderate discrimination, with Harrell’s c-index of 0.63 for all-cause mortality and 0.64 for cardiovascular death. The addition of DNN-derived imaging biomarkers significantly improved discrimination, with a c-index of 0.69 for both outcomes (P0.001). Conclusion Our developed open-source, deep learning-based segmentation and quantification model for full 3D body composition analysis to explore imaging biomarkers in CT images. The newly derived imaging biomarkers from fat and muscle volume, as well as intensity, provided added value and improved long term outcome prediction in TAVI patients.
Shiri et al. (Sat,) reported a other. 3D CT-based body composition imaging biomarkers improved prediction of all-cause and cardiovascular mortality after TAVI, raising c-index from 0.63 to 0.69 (P<0.001).