AG-SCL for Long-Tailed ECG Arrhythmia Diagnosis

Angular Gaussian Supervised Contrastive Learning (AG-SCL) addresses rare arrhythmia detection by modeling class uncertainty on unit-normalized embeddings. The method uses adaptive logit adjustment and morphology-preserving augmentations that protect the 7-25 Hz QRS band on the PTB-XL benchmark.