Foundation model

PPGPT

PPG foundation model with next-feature token prediction (AAAI 2026)

From “PPGPT: Transferring Next-Token Modeling from Language to PPG Signals”
Venue listedCatalog paper date: Mar 14, 2026Added to atlas: Nov 16, 2026Source

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

PPGPT is a photoplethysmography foundation model described in AAAI 2026. The authors reformulate next-token prediction as next-feature token prediction (NFTP) and use a dual-stream encoder that jointly models temporal dynamics and local-global morphological patterns. The official repository names that encoder Hierarchical Spatio-Temporal Fusion (HSTF) and pairs it with a DeepSeekMoE decoder. The paper says the model was pre-trained on a mixed dataset of 1.6 billion data points and evaluated on the authors' BioMTL benchmark: 172 subjects over 285 days, seven tasks. The official README specifies 1,808 four-second fingertip-PPG windows at 1000 Hz, with regression labels for SBP, DBP, blood glucose, heart rate and BMI, and classification labels for mental stress, hypertension and type-2 diabetes. The authors report a 16.5% improvement in F1-score and a 25.9% reduction in MAE versus competing methods, plus few-shot results. Code is on GitHub under MIT; the repository has no trained weights, and BioMTL is released under a data-use agreement.

Signals
PPG
Year
2026
Published

AAAI 2026 (Proceedings of the AAAI Conference on Artificial Intelligence, 40(34), 28573–28581, doi:10.1609/aaai.v40i34.40088), 14 March 2026

Authors

Zexing Zhang, Huimin Lu, Qingxin Zhao (Changchun University of Technology)

Architecture

dual-stream Hierarchical Spatio-Temporal Fusion (HSTF) encoder with a DeepSeekMoE decoder; pre-training objective is next-feature token prediction (NFTP). Official code at https://github.com/PPGPT/AAAI2026 (MIT).

Pretraining data

a mixed dataset of 1.6 billion data points (composition not itemized on the publisher abstract or the official repository README).

Evaluation

authors’ BioMTL benchmark, 1,808 four-second fingertip-PPG windows (1000 Hz) from 172 subjects over 285 days, seven tasks (SBP, DBP, blood glucose, heart rate, BMI; mental stress, hypertension, type-2 diabetes). Dataset access is by data-use agreement. The repository does not include trained weights.

Result

the authors report a 16.5% improvement in F1-score and a 25.9% reduction in MAE versus competing methods on BioMTL, plus few-shot results. Baselines and per-task scores are in the paper.

At a glance
Type
Foundation model
Modalities
PPG
Form factor
Not reported
Parameters
Not reported
Architecture
Dual-stream Hierarchical Spatio-Temporal Fusion (HSTF) encoder with a DeepSeekMoE decoder; next-feature token prediction (NFTP)
Venue
AAAI 2026 (Proceedings of the AAAI Conference on Artificial Intelligence, 40(34), 28573–28581)
Year
2026
arXiv
Not reported
Weights
Not released
Caveats
Not reported
Datasets
Not reported
Added
Nov 16, 2026
Updated
Nov 16, 2026
PPGPT: ChatPPG research summary card

Key results

25.9%
F1-score
The authors report a 16.5% improvement in F1-score and a 25.9% reduction in MAE versus competing methods, plus few-shot results

How it works

How PPGPT works: Input signals → Foundation model → Downstream tasksInput signals: PPG (Photoplethysmogram). Pretraining data: Mixed dataset of 1.6 billion data points (composition not itemized on the publisher abstract or the official repository README). Foundation model: PPGPT (Masked reconstruction).Input signalsPPG: PhotoplethysmogramPPGPhotoplethysmogramPretraining data: Mixed dataset of 1.6 billion data points (composition not itemized on the publisher abstract or the official repository README)Pretraining dataMixed dataset of 1.6 billion data points…Foundation modelMASK → RECONSTRUCTPPGPTMasked reconstruction
Pretraining data
Mixed dataset of 1.6 billion data points (composition not itemized on the publisher abstract or the official repository README)
Architecture
Dual-stream Hierarchical Spatio-Temporal Fusion (HSTF) encoder with a DeepSeekMoE decoder; next-feature token prediction (NFTP)
Inputs
PPG
Approach
PPG · foundation model · AAAI

Datasets used

No public datasets linked yet.

Papers· 1

Authors· 3

Cite this model

Cite the publication recorded for this model. Check it against the publisher before submitting.

Zhang, Z., Lu, H., & Zhao, Q. (2026). PPGPT: Transferring Next-Token Modeling from Language to PPG Signals. AAAI 2026. https://doi.org/10.1609/aaai.v40i34.40088
Zhang, Zexing, Huimin Lu, and Qingxin Zhao. 2026. "PPGPT: Transferring Next-Token Modeling from Language to PPG Signals." AAAI 2026. https://doi.org/10.1609/aaai.v40i34.40088.
Zhang, Z., Lu, H. and Zhao, Q. (2026) 'PPGPT: Transferring Next-Token Modeling from Language to PPG Signals', AAAI 2026. doi: 10.1609/aaai.v40i34.40088.
Zhang, Zexing, et al. "PPGPT: Transferring Next-Token Modeling from Language to PPG Signals." AAAI 2026, 2026. https://doi.org/10.1609/aaai.v40i34.40088.
Z. Zhang, H. Lu, and Q. Zhao, "PPGPT: Transferring Next-Token Modeling from Language to PPG Signals," AAAI 2026, 2026, doi: 10.1609/aaai.v40i34.40088.
@inproceedings{zhang2026ppgpt,
  title         = {{PPGPT: Transferring Next-Token Modeling from Language to PPG Signals}},
  author        = {Zhang, Zexing and Lu, Huimin and Zhao, Qingxin},
  year          = {2026},
  booktitle     = {AAAI 2026},
  doi           = {10.1609/aaai.v40i34.40088}
}

History

  1. Added to the ChatPPG research hub