A Physiology-Aware Hierarchical Foundation Model for Long-Context Waveform Representation Learning (Physio-HNet)

Anonymous (ICLR 2027 double-blind)

Summary

Causal hierarchical waveform FM replacing fixed patches with physiology-aware dynamic chunking (landmark / cardiac-cycle guided sub-beat and beat-level units), with hierarchical compression and recurrent state carrying context across windows. After SSL pretraining, routine vital signs (HR, RR, SpO₂, BP) supervise higher-level representations along each recording ("clinically anchored continued pretraining"). 1.2 M parameters; SSL encoder comparable to existing waveform FMs; continued pretraining gives highest R² and lowest MAE in 11 of 12 BP comparisons across two external clinical cohorts and two probing protocols (all six operating-room comparisons).