AG-SCL Improves Long-Tailed ECG Arrhythmia Diagnosis

This research tackles the challenge of accurately diagnosing rare or "long-tailed" arrhythmias from electrocardiogram (ECG) data, a problem where deep learning models often struggle due to imbalanced class distributions and varied ECG morphologies. The proposed solution is Angular Gaussian Supervised Contrastive Learning (AG-SCL), a unified framework designed to enhance diagnostic reliability. AG-SCL incorporates three key components: an Angular Gaussian contrastive branch that models full-covariance class uncertainty on unit-normalized embeddings, allowing for better capture of direction-dependent morphological variations; Adaptive Logit Adjustment, which learns flexible, label-state-specific prior corrections instead of fixed frequency-based margins; and tail-aware augmentation, which generates morphology-preserving views while protecting critical QRS-dominant bands. Evaluated on the PTB-XL benchmark and a nocturnal ECG dataset, AG-SCL demonstrated superior macro-level performance, with significant gains observed particularly in rare or morphologically unstable rhythm classes, confirming the effectiveness of its integrated approach.