A Kolmogorov-Arnold Network for 5-class heartbeat classification where every filter is a learnable wavelet that can be inspected directly rather than explained after the fact, currently under review at Artificial Intelligence in Medicine. Evaluated exclusively on a strict inter-patient split (the de Chazal protocol) across 20 seeds with Holm-Bonferroni-corrected significance testing, and validated zero-shot on an entirely unseen dataset (PTB-XL) rather than stopping at in-distribution numbers.

A self-supervised pretraining method for single-lead ECG that constrains its own augmentations to work around the QRS complex instead of distorting it, so downstream abnormal-beat detection needs far fewer expert-labeled examples. Combines physiology-aware augmentation with a hybrid contrastive + masked-reconstruction loss across a multi-dataset PhysioNet corpus.