Transcatheter aortic valve implantation (TAVI) is a minimally invasive procedure for treating aortic valve stenosis. It is associated with potential complications, including paravalvular leak (PVL) and conduction disturbances, such as atrioventricular block (AVB). We present a fully automated pipeline to predict these complications using radiomics features automatically extracted from convolutional neural network (CNN) segmented computed tomography (CT) images. Radiomics features contribute to interpretability by quantifying shape, intensity, and texture patterns in anatomically defined regions. Our approach focuses on the aortic root, left ventricular outflow tract (LVOT) and proximal ascending aorta (pAAo). Using a gradient-boosted decision tree algorithm (XGBoost), we calculate patient-specific risk scores for overall complications as well as AVB and PVL. We compare our radiomics-based model to (i) a model trained on conventional image-based biomarkers, (ii) a DenseNet-based CNN trained on CT cropped to a bounding box around the aortic root, and (iii) a dual-channel DenseNet using the cropped CT and corresponding segmentation. We also evaluate whether combining radiomics with conventional parameters enhances predictive performance. Model performance varied across complication types and models, with AUCs on the test set ranging from 0.42 (DenseNet_CT+Seg) to 0.54 (combined) for any complication, 0.40 (radiomics) to 0.59 (DenseNet_CT) for AVB, and 0.35 (DenseNet_CT) to 0.63 (combined) for PVL. Although predicting TAVI complications remains challenging, our approach mitigates operator-dependent variability in image segmentation and feature extraction, promoting fully automated, reproducible, and interpretable risk stratification for TAVI-related complications.