Benchmarking Open-Source PPG Foundation Models for Biological Age Prediction

Code and results for the preprint (2026). Work in progress.

What this is

Apple showed in October 2025 (Nature Communications) that PPG signals from a wrist sensor can predict biological age with 2.4-year MAE, and that this "PPG age gap" is associated with cardiovascular disease, diabetes, and mortality. A second group confirmed similar results on the UK Biobank (212K subjects). Both models are closed-source.

This repo benchmarks two open-source PPG foundation models (Pulse-PPG and PaPaGei-S) on the same task, using public clinical data (PulseDB/VitalDB, 906 subjects). No fine-tuning, just linear probing on frozen embeddings.

Results (906 VitalDB subjects, 5-fold stratified CV)

Model	MAE (years)	R2	Corr
Baseline (predict mean)	11.91
HR only	11.83
HR + HRV	11.49
Demographics (all)	10.04
HR/HRV + Demographics	9.59
AI-PPG Age (zero-shot)	11.50	0.043	0.313
PaPaGei-S (linear probe)	9.80 +/- 0.34	0.313	0.583
AI-PPG Age (linear probe)	9.72 +/- 0.37	0.310	0.563
Pulse-PPG (linear probe)	9.28 +/- 0.44	0.388	0.645
PaPaGei-S + Demographics	8.56 +/- 0.22	0.477	0.696
AI-PPG Age + Demographics	8.58 +/- 0.25	0.440	0.664
Pulse-PPG + Demographics	8.22 +/- 0.25	0.517	0.725

Three things stand out. Pulse-PPG embeddings alone (MAE 9.28) beat HR/HRV combined with all demographics (9.59), so the waveform morphology carries real aging information beyond heart rate. AI-PPG Age fails zero-shot on clinical data (predictions stuck at 38-67yr for a population aged 18-92yr), which illustrates the domain shift problem. And fusion with demographics pushes the best result to 8.22 years.

The age-adjusted PPG age gap correlates with diastolic blood pressure (r=-0.188, p=1.2e-8), consistent with Apple's finding that PPG captures vascular aging.

Project structure

PPGAge/
├── src/
│   ├── run_pulsedb_age.py     # Main benchmark pipeline (5-fold CV)
│   ├── analyze_age_gap.py     # PPG age gap vs cardiovascular markers
│   ├── run_ppgbp_age.py       # PPG-BP dataset analysis
│   └── analyze_pulsedb_demographics.py
├── data/
│   ├── vitaldb_fileids.json   # All 2938 VitalDB Google Drive file IDs
│   └── download_list.json     # Selected subjects to download
├── results/
│   └── pulsedb_vitaldb/       # Plots and full output log
├── pulseppg/                  # Pulse-PPG (clone separately, MIT license)
├── papagei-foundation-model/  # PaPaGei (clone separately)
├── ppg-vascularage/           # AI-PPG Age weights + patched net1d.py
└── download_vitaldb.py        # Script to download VitalDB segments

Resources

Resource	Link
PpgAge paper (Apple, Nature Comms 2025)	doi
AI-PPG Age paper (Comms Medicine 2025)	doi
Pulse-PPG (MIT)	github.com/maxxu05/pulseppg
PaPaGei	github.com/Nokia-Bell-Labs/papagei-foundation-model
AI-PPG Age weights	HuggingFace Ngks03/PPG-VascularAge
PulseDB dataset	github.com/pulselabteam/PulseDB

Setup

git clone https://github.com/Misterbra/ppg-age-benchmark
cd ppg-age-benchmark
pip install -r requirements.txt

# Clone the model repos
git clone https://github.com/maxxu05/pulseppg
git clone https://github.com/Nokia-Bell-Labs/papagei-foundation-model

# Download weights
cd pulseppg && bash download_model.sh && cd ..
# For PaPaGei: see papagei-foundation-model/README.md
# For AI-PPG Age: huggingface-cli download Ngks03/PPG-VascularAge --local-dir ppg-vascularage/

# Download VitalDB segments (needs Google Drive access, see script)
python download_vitaldb.py

Run

python src/run_pulsedb_age.py
# Results saved to results/pulsedb_vitaldb/

Citation

@article{brag2026ppgage,
  title={Benchmarking Open-Source PPG Foundation Models for Biological Age Prediction},
  author={Brag, N.},
  year={2026},
  note={Preprint}
}