Paper
JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare
NeurIPS 2025 Workshop TS4H (Learning from Time Series for Health), poster; OpenReview QqQDjLgHab, published 2025-09-23
TL;DR
A masked joint-embedding (JEPA-style) pretraining framework for *high-level behavioural* wearable time series (daily/sporadic metrics such as resting HR, SpO2, sleep stages, activity) with heavy noise, missing values and irregular sampling. Per the blog: 63 metric channels. Each observation is a (time, value, metric-type) triplet token (d=64). The context encoder sees a random ~30% subset, and a predictor regresses an EMA target encoder's embeddings. Mamba and Transformer variants. Pretrained on ~3M person-days.
- Form factor: wrist
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Signals
Keywords
Models
How it works
Cite this paper
APA
Xie, E., Chang, W., Martinez, R. R., & Ballinger, B. (2025). JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare. NeurIPS 2025 Workshop TS4H , poster. https://openreview.net/forum?id=QqQDjLgHab
BibTeX
@inproceedings{xie2025jets,
title = {{JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare}},
author = {Xie, Erik and Chang, Wyatt and Martinez, Raquel Rodriguez and Ballinger, Brandon},
year = {2025},
booktitle = {NeurIPS 2025 Workshop TS4H , poster},
url = {https://openreview.net/forum?id=QqQDjLgHab}
}Built from the hub record. Check it against the publisher before submitting.
Added to the ChatPPG research hub
