Aortic stenosis (AS) is the most prevalent cardiac valvular disease, characterized by calcification or restricted motion of the aortic valve (AV) leaflets. Machine learning-driven approaches have emerged as promising tools for automated AS severity classification, with prototypical neural networks offering an inherently interpretable alternative to black-box models. However, existing prototype-based methods do not account for the inherent hierarchy in AS-related features, which range from localized spatio-temporal patterns—focusing on the AV region—to broader features that span larger spatial areas or longer temporal scales. To address this, we propose HiProtoNet, a hyperbolic part prototype-based framework that systematically learns hierarchical feature representations in a Lorentzian hyperboloid space. By leveraging hyperbolic distance, HiProtoNet explicitly models both local and broad prototypes, ensuring that localized diagnostic features are mapped near the root of the hyperboloid while broader spatio-temporal patterns reside farther out. Evaluations on a private real-world dataset demonstrate that HiProtoNet outperforms state-of-the-art baselines in AS classification accuracy. Furthermore, by structuring the prototype space hierarchically, HiProtoNet provides a more principled way of learning localized and broad features, which enhances interpretability. Our source code is available at: https://github.com/DeepRCL/HiProtoNet .