This paper presents ProtoECGNet, an interpretable multi-label deep learning model for electrocardiogram (ECG) classification. ProtoECGNet employs a structured multi-branch architecture that reflects clinical interpretation workflows. It performs rhythm classification using a 1D CNN and global prototypes, morphology-based inference using a 2D CNN and time-localized prototypes, and diffusive abnormality classification using a 2D CNN and global prototypes. Each branch is trained with a prototype loss function designed for multi-label learning and incorporates a novel contrastive loss that promotes clustering, separability, diversity, and proper separation between prototypes of unrelated classes. We evaluate ProtoECGNet on 71 diagnostic labels from the PTB-XL dataset, demonstrating competitive performance against state-of-the-art black-box models and providing structured, case-based explanations. Clinician review of the prototypes confirmed their representativeness and clarity. ProtoECGNet demonstrates that prototype learning can effectively scale to complex multi-label time series classification, providing a practical path toward transparent and reliable deep learning models for clinical decision support.
