Automated survival prediction based on body composition analysis of Pre-TAVI-Procedural CT scans

Abstract Background Computed tomography (CT) is an integral part of planning a transcatheter aortic valve implantation (TAVI). We developed an automated image analysis pipeline to analyze CT scans and identify anatomical features significantly associated with patient survival outcomes. Methods We have analyzed consecutive 299 TAVI patients (2016-2024), all of whom underwent pre-procedural CT scans including the heart and femoral arteries prior to TAVI. The CT scans were automatically processed from our hospital’s PACS by the Body and Organ Analysis framework (UMEssen), which integrates TotalSegementator and Body Composition Analysis. We assessed correlations between the radiomics features and survival outcomes using the Bonferroni method with Benjamini-Hochberg correction. Based on the derived features, we trained a machine learning model to predict short (12months), intermediate (12m 36m) and long term (36m) survival. Results In our cohort, the median age was 82 (IQR: 79 - 85) years, 155 (51.8%) were women, and the median BMI was 26.3 kg/m2. Survival rates were 90% at 6 months, 86% at 12 months, 79% at 24 months and 70% at 36 months post-TAVI. Based solely on the radiomics data, our machine learning model reached an AUROC 0.727 for short 0.374 for intermediate and 0.620 for long term survival. The F1 score indicates that the model is more accurate in predicting short- and long-term survival than intermediate-term, with scores of 0.833 (long-term), 0.133 (intermediate-term), and 0.608 (short-term). We found key predictors for patients survival including pulmonary fat, vertebral bone volume, intermuscular adipose tissue, subcutaneous adipose tissue, and organ volumes (kidney, spleen, pancreas, esophagus, adrenal glands). Conclusion We have developed an unbiased analysis pipeline for existing pre-TAVI CT scans, deriving valuable data that would otherwise remain unutilized. Our approach corroborated established prognostic factors, such as bone volume and intermuscular adipose tissue, while also identifying novel survival predictors, including pulmonary fat. Following validation in a randomized controlled trial (RCT), the data obtained through this method could inform clinical decision-making, particularly in optimizing personalized follow-up care.