Scaling Representation Learning from Ubiquitous ECG with State-Space Models