Heart fat detection has emerged as a crucial aspect of cardiovascular health assessment. Accurate identification and quantification of fat deposits around the heart play a significant role in diagnosing various cardiac diseases. Through this manuscript, a proposal of novel methodology for heart fat detection by integrating multiple features extracted from medical imaging data using advanced data mining techniques is given. Specifically, we combine the powerful ensemble learning algorithm XGBoost with the cutting-edge vision transformer model CaiT (classifier-agnostic image transformers) to enhance the detection accuracy and robustness. Our approach enhances the important features of both algorithms to improve the performance of heart fat detection systems. On conducting comprehensive experiments using diverse dataset to validate effectiveness of proposed methodology, achieving promising results compared to state-of-the-art approaches, the proposed methodology resulted at an 94.2% in accuracy, 92.6% in sensitivity, 95.8% in specificity, and F1-score of 93.9, surpassing baseline methods vision transformers (ViT), support vector regressor (SVR), and algorithms constituting to the state-of-the-art such as convolutional neural network (CNN) which evaluated its capability to accurately detect heart fat deposits.