Recent ECG foundation models provide transferable representations, but most downstream pipelines still compress the waveform into a global embedding and apply classifiers that ignore waveform structure and the clinical hierarchy of ECG abnormalities. We propose a segment-aware framework for hierarchical ECG classification that combines robust automatic delineation, pretrained ECG foundation model embeddings, and a lightweight hierarchy-aware classification head. The method is designed for large-scale 12-lead datasets without manual annotations and is evaluated on CODE-15, a cohort of 345,779 ECGs from 233,770 patients.
ECG delineation remains an open problem in real-world recordings due to noise, morphology variability, lead-dependent visibility of waveform components, and scarcity of large annotated corpora. To address this, we adopt a two-stage delineation pipeline that first localizes QRS complexes and then estimates P-wave, QRS, and T-wave boundaries, enabling scalable physiological decomposition without relying on manual segmentation. Each delineated segment is independently encoded by a pretrained ECG foundation model, preserving morphology-specific information before classification.
Instead of heavily adapting the full encoder, we learn only a lightweight hierarchy-aware classifier on top of the segment representations. This parameter-efficient design retains the generalization ability of the pretrained model while structuring predictions across broad diagnostic groups and fine-grained abnormalities. It also yields a clinically intelligible decision process: rather than producing diffuse saliency over the entire signal, the model decomposes diagnostic evidence across physiologically meaningful waveform components, making it possible to inspect whether a prediction is primarily supported by atrial activity, ventricular depolarization, or repolarization patterns.
On CODE-15, the proposed framework achieved performance comparable to a strong, well-established ResNet baseline previously reported for hierarchical ECG classification (macro-F1 0.760), while providing richer physiological explanations of how each diagnosis is supported. These findings suggest that robust automatic delineation, combined with foundation-model representations and lightweight hierarchical prediction, is a practical path toward scalable and interpretable ECG AI.