Training and Scaling Compute-Optimal Physiological Waveform Foundation Models

Pingzhi Li, Jie Peng, Shuqing Luo, Zachary Plotkin, Tianlong Chen (Foci Labs; UNC Chapel Hill). Corresponding: Plotkin, Chen (tianlong@cs.unc.edu)

Summary

Aether: more than 100 masked-reconstruction FMs (20M–2.1B) on multichannel MIMIC-III waveforms at 125 Hz (ECG, ABP, pleth, resp), using 1-hour context through a Bi-Gated-DeltaNet/attention hybrid. The authors trained it on their own JAX stack (FoXus) across Trainium2, TPU v5p and B200. There are 8 held-out MIMIC-III clinical tasks with patient-level macro AUROC and frozen linear probes.