ECG-PPG JEPA (Bridging ECG and PPG; unnamed in paper)
Key facts
Summary
Multimodal Joint-Embedding Predictive Architecture for PPG + ECG, built on LeJEPA. Separate ViT1D encoders predict in a shared latent space with SIGReg (isotropic-Gaussian) regularisation and stochastic modality dropout, so the PPG encoder works alone at inference. "ECG-PPG JEPA" is the PPG Atlas label; the paper leaves it unnamed.
Reported benchmarks
AFib / arrhythmia
Stanford wrist PPG AF (DeepBeat) · Macro-F1 ↑
Predefined split, frozen encoder; F1 — not comparable to AUROC rows; 11M 0.791
Cuffless blood pressure
pulsedb-vital · MAE (mmHg) ↓
Vital-BP predefined split from PulseDB; PulseDB also in pretraining (overlap not discussed); clinical finger
Cuffless blood pressure
uci-bp · MAE (mmHg) ↓
Nested 5-fold CV (AnyPPG protocol); clinical finger
Heart rate estimation
Brno University of Technology Smartphone PPG Database (BUT PPG) · MAE (bpm) ↓
Nested 5-fold CV (AnyPPG protocol); smartphone PPG
Heart rate estimation
Gyro-Acc-PPG (BAMI-I, Wonkwang) · MAE (bpm) ↓
Nested 5-fold CV (AnyPPG protocol); treadmill wrist PPG
Heart rate estimation
IEEE Signal Processing Cup 2015 (TROIKA) · MAE (bpm) ↓
LOSO, frozen
Heart rate estimation
PPG-DaLiA · MAE (bpm) ↓
LOSO, frozen; PPG-only input
Heart rate estimation
WESAD · MAE (bpm) ↓
LOSO, frozen encoder (KDD'26 Table 2)
Datasets used
Source paper
Bridging ECG and PPG: Latent-Space Prediction for Robust Physiological Analysis
Related Research
- PulseDB v2.0
cleaned MIMIC-III + VitalDB cuffless-BP benchmark (MIMIC part ODbL; VitalDB part CC BY-NC-SA)
- WESAD
Wearable Stress and Affect Detection
- PPG-DaLiA
PPG dataset for motion compensation / HR
- Bridging ECG and PPG: Latent-Space Prediction for Robust Physiological Analysis
Zhaoliang Chen, Saurabh Kataria, Minxiao Wang, Xiao Hu. Nell Hodgson Woodruff School of Nursing, Emo
- Aether
compute-optimal physiological waveform FM family (20M–2.1B)
- AnyPPG
Foundation model