Key facts
- Parameters
- ≈1.5M params (vs PaPaGei 5.7M, NormWear 135.3M, Chronos 200M)
- Architecture
- bidirectional Mamba (selective SSM), temporal+frequency fusion, ~1.5M params
- Pretraining Data
- In-house: 100 participants × ~3 weeks Empatica E4 wrist BVP ≈ 50,000 h; 0.5–12 Hz band-pass, 32 s windows; masked reconstruction with multi-scale loss (global + masked + peak-amplitude)
- Venue
- IEEE ICASSP 2026
Summary
Self-supervised PPG FM used as a frozen feature extractor. Across 15 tasks on 8 public datasets it is best on 9/15 against Chronos, PaPaGei and NormWear (authors' re-runs). Strongest on HR (DaLiA MAE 7.27), BP (PPG-BP SBP 12.89) and AF (MIMIC 0.83 acc).
Reported benchmarks
AFib / arrhythmia
MIMIC PERform · Accuracy ↑
MIMIC 8 s AF, frozen features; accuracy only (no AUROC) — not comparable to AUROC rows; Chronos 0.81 / NormWear 0.78 / PaPaGei 0.66
Cuffless blood pressure
PPG-BP Database (Guilin / Liang 2018) · MAE (mmHg) ↓
PPG-BP (Liang 2018) 2 s segments, frozen features (paper Table 1); finger clinical PPG; DBP Chronos 8.37 better
Heart rate estimation
PPG-DaLiA · MAE (bpm) ↓
8 s windows, frozen features (paper Table 1); split unstated; same-paper re-runs NormWear 10.16 / PaPaGei 12.25 / Chronos 12.34
Mental workload / stress
WESAD · Accuracy ↑
Binary stress, 8 s, frozen features (paper Table 2); accuracy — not comparable to AUROC rows; PaPaGei 0.80 / Chronos 0.77 / NormWear 0.77