Research map / Label-efficient and robust learning / Self-supervised and contrastive
PaPaGei: Open Foundation Models for Optical Physiological Signals
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces PaPaGei, an open PPG foundation model pre-trained on ~57K hours of public PPG data with a morphology-aware contrastive objective, evaluated on 20 downstream tasks against time-series FMs and SSL baselines.
Problem
PPG signal models are task-specific and generalize poorly; there are no open, generalizable pre-trained PPG foundation models and out-of-domain evaluation is limited.
Approach
Pre-trains a ResNet-style PPG encoder with a morphology-aware contrastive objective (PaPaGei-S) that pairs segments by discretized sVRI bins and adds mixture-of-expert heads predicting IPA and SQI, alongside a patient-contrastive variant (PaPaGei-P).
Key results
- PaPaGei-S achieves the highest average classification AUROC (0.67) and lowest average regression MAE (10.12) across 20 tasks, versus 0.62-0.64 AUROC for Chronos/Moment/TF-C.
- PaPaGei-S with 5M parameters outperforms models up to 70x larger (Chronos 200M, Moment 385M) on many tasks.
- PaPaGei-S improves ICU admission AUROC to 0.79 and SDB AUROC to 0.70 versus the best baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use morphology-aware PaPaGei-S instead of the participant-aware PaPaGei-P objective | PaPaGei · Downstream classification tasks (average) | AUROC: 0.63 ± 0.08 -> 0.67 ± 0.09 | Helped |
| Use morphology-aware PaPaGei-S instead of the participant-aware PaPaGei-P objective | PaPaGei · Downstream regression tasks (average) | MAE: 10.92 ± 4.25 -> 10.12 ± 3.47 | Helped |
| Train with all three SSL objectives rather than sVRI only (ablation) | PaPaGei-S · Downstream classification tasks (average) | AUROC: 0.64 -> 0.67 | Helped |
| Train with all three SSL objectives rather than sVRI only (ablation) | PaPaGei-S · Downstream regression tasks (average) | MAE: 10.35 -> 10.12 | Helped |
| Add IPA and SQI objectives to sVRI (ablation) | PaPaGei-S · Downstream classification tasks (average) | AUROC: 0.62 -> 0.67 | Helped |
| Add IPA and SQI objectives to sVRI (ablation) | PaPaGei-S · Downstream regression tasks (average) | MAE: 10.80 -> 10.12 | Helped |
| Pre-train with more upstream data, using all three datasets rather than fewer datasets (ablation) | PaPaGei-S · Downstream tasks | Performance | Helped |
| Increase pre-training model size from 5M to 35M or 139M parameters (ablation) | PaPaGei-S · Downstream tasks | Performance | Hurt |
| Increase embedding size from 512 to 768 or 1024 (ablation) | PaPaGei · Downstream tasks | Performance | No clear effect |
| Change training batch size from 128 to 64 or 256 (ablation) | PaPaGei · Downstream tasks | Performance | No clear effect |
| Include demographic features alongside PaPaGei-S (ablation) | PaPaGei-S · Downstream tasks | Performance | Helped |
| Use PaPaGei-S rather than demographic features plus PPG (ablation) | PaPaGei-S · Downstream tasks | Performance: 14 out of 18 tasks | Mixed |
Surprising findings
- Chronos performs well on mortality, pregnancy stage, and smoking, suggesting general-purpose time-series models suffice for slowly varying outcomes.
- SimCLR and BYOL embeddings show clustering of participants (mild mode collapse), while PaPaGei-S gives the widest inter-participant dispersion.
Limitations
- Pre-training uses only three public datasets; diversity of devices and populations is limited.
- Performance on darker skin tones remains weaker; no single model is best across all skin tones.
- Scaling analysis is non-monotonic and the reported gains are modest on several tasks; hyperparameter selection picked best of multiple pre-training runs per task.
Experimental setup
Datasets: VitalDB MIMIC-III waveform matched subset MESA nuMoM2B VitalVideos PPG-BP SDB ECSMP WESAD PPG-DaLiA
Benchmarks: VitalDB MIMIC-III MESA nuMoM2B VitalVideos PPG-BP SDB ECSMP WESAD PPG-DaLiA
Metrics: AUROC MAE accuracy F1-score R^2 sMAPE
Baselines: Chronos MOMENT REGLE SimCLR BYOL TF-C Statistical features
Builds on: SimCLR BYOL TF-C NT-Xent
Closest papers by meaning
- FactorizePhys: Matrix Factorization for Multidimensional Attention in Remote Physiological Sensing (NeurIPS 2024 · similarity 0.86)
- PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement (AAAI 2025 · similarity 0.86)
- Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.86)
- FacePhys: State of the Heart Learning (NeurIPS 2026 · similarity 0.86)
- WildPPG: A Real-World PPG Dataset of Long Continuous Recordings (NeurIPS 2024 · similarity 0.85)
- Constraint Latent Space Matters: An Anti-anomalous Waveform Transformation Solution from Photoplethysmography to Arterial Blood Pressure (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.