In recent years, the use of deep learning models has established itself as a transformative approach in cardiology. In this context, disease categories often present hierarchical relationships, such as subtypes of arrhythmias. Moreover, Euclidean latent spaces struggle to encode hierarchical clinical structure, whereas hyperbolic geometry, with its negative curvature and exponential volume growth, naturally represents hierarchies with lower distortion, making it a promising basis for medical representation learning. In this work, we propose a model that integrates a Euclidean convolutional neural network backbone with a hyperbolic representation learning scheme. Guided by cardiology prior knowledge of class hierarchy, we introduce a hierarchy-aware loss that supervises superclasses, enforces parent–child consistency, and shapes embeddings geometry via radial/angular regularizers. Experimental results on multi-label ECG classification show consistent gains over a fully Euclidean baseline: macro F1 improves from 0.760 (ResNet) to 0.786 with a hyperbolic head (HypNet) and to 0.8509 with the proposed hierarchy-aware model (H-HypNet); macro recall rises from 0.662 to 0.7077 and to 0.787, respectively. These findings provide preliminary indications that hyperbolic representations, coupled with hierarchy-aware training, help model complex cardiovascular taxonomies in medical deep learning.