PulseBound: Future-Beat State Forecasting Under an Explicit Information Boundary

Anonymous (ICLR 2027 double-blind)

Summary

Points out that causal attention doesn't guarantee one-sided information access when normalization, nonlocal transforms or companion views see withheld samples. PulseBound enforces a stored-window cutoff (prefix-only normalization, suffix replacement before derived views, aligned masking) so the forecast context is provably invariant to the stored suffix, then predicts nine rhythm/morphology descriptors for up to four future beats (optional ECG-derived PAT supervision in training only). On held-out MIMIC / VitalDB groups: 28.06% / 22.22% lower nine-state MAE vs last-beat persistence; in a seven-model comparison across 13 downstream tasks, best mean on nine frozen linear-probe and seven full-fine-tune tasks.