Pulse-PPG

Pulse-PPG is an open-source foundation model for photoplethysmography (PPG), the optical signal captured by the green-light sensors in consumer smartwatches and fitness bands. PPG underpins continuous estimates of heart rate, blood pressure, stress, and sleep, but raw field-collected PPG is notoriously noisy — corrupted by motion artifacts, skin tone, sensor fit, and ambient light. Most prior PPG foundation models were pretrained on clean, clinically collected datasets, which limits how well they transfer to the messy data that real wearables produce.

Pulse-PPG takes the opposite approach: it is, to the authors' knowledge, the first PPG foundation model trained directly on raw, uncurated signals from a free-living field study rather than a controlled clinical setting. It was developed by a team led by Mithun Saha and Santosh Kumar at the University of Memphis together with Maxwell A. Xu and James M. Rehg at the University of Illinois Urbana-Champaign, first released as a preprint in February 2025 and published at UbiComp / ACM IMWUT 2025.

The central finding is that pretraining on field data yields markedly better cross-dataset generalization than pretraining on clinical data — a field-trained model handles both lab and field downstream tasks well, whereas clinically-trained baselines degrade sharply when moved out of the lab. By releasing the code and pretrained weights, the authors aim to give the mobile health community a robust, reusable starting point.

#Key Features

• Field-trained pretraining: Learns representations from raw PPG collected over a 100-day field study of 120 participants wearing everyday devices, capturing real-world variability that clinical corpora miss.
• Self-supervised, label-free: Trained with the RelCon relative-contrastive objective using a learned motif-distance (MotifDist) measure, so no annotated labels are needed during pretraining.
• Variable-length inputs: A temporal pooling step collapses the time axis, letting the same model accept inputs of different durations despite being pretrained on 4-minute windows.
• Strong cross-dataset transfer: Outperforms clinically-pretrained PPG foundation models across lab and field benchmarks, demonstrating better out-of-distribution robustness.
• Fully open: MIT-licensed code, downloadable pretrained weights on Zenodo, and a Colab demo for quick evaluation.

#Technical Details

Pulse-PPG uses a convolutional (ResNet-style) backbone with temporal pooling, configured through a modular class-based config system. Pretraining follows the RelCon framework, a self-supervised relative-contrastive approach in which a learnable motif-distance function (MotifDist) defines which signal segments are treated as similar or dissimilar, allowing the model to learn from raw waveforms without labels. The pretraining corpus is raw PPG gathered during a 100-day free-living study of 120 participants, processed in 4-minute windows. The released model is evaluated on four downstream datasets — PPG-BP (blood pressure), PPG-DaLiA (activity/heart rate), SDB (sleep-disordered breathing), and WESAD (stress and affect) — spanning both controlled-lab and field conditions, and is benchmarked against state-of-the-art PPG foundation models trained on clinical data, which it surpasses on overall generalization.

#Applications

Pulse-PPG is aimed at researchers and developers building digital-health applications on consumer wearables. Because it produces general-purpose PPG embeddings, it can be fine-tuned or linearly probed for tasks such as blood pressure estimation, heart-rate and activity inference, stress and affect detection, and sleep-related monitoring — particularly in deployments where data is collected in everyday, free-living conditions rather than the clinic. Its open weights and Colab demo lower the barrier for mobile-health labs that lack the large field studies needed to pretrain such a model from scratch.

#Impact

Pulse-PPG addresses a persistent gap between PPG models that perform well on clean clinical recordings and the noisy reality of consumer devices. By showing that field-trained pretraining generalizes better than clinical pretraining across both settings, it offers practical evidence that data realism matters as much as data cleanliness for wearable foundation models, and provides one of the first openly released, field-trained PPG backbones for the community to build on. As a relatively new release, its long-term adoption and performance on additional sensors and populations remain to be established.