GPT-PPG
PPGDOI
Key facts
Summary
Decoder-only GPT pretrained on >200M 30-s single-channel PPG segments from UCSF ICU bedside monitors (private). Sizes 19M / 85M / 345M / 1B. Predicts a per-sample logit-Laplace distribution for the next 1-s patch, so one model embeds, completes and denoises PPG.
Datasets used
Stanford wrist PPG AF (DeepBeat)
PPG
WESAD
PPGECGIMUTempEDA
PPG-DaLiA
PPGECGIMUTempResp
IEEE Signal Processing Cup 2015 (TROIKA)
PPGIMUECG
BIDMC PPG and Respiration (Pimentel 2016 / PhysioNet)
PPGECGRespSpO₂
PulseDB v2.0
PPGECGABP
UCSF ICU bedside-monitor waveforms (private; GPT-PPG / SiamQuality / SQUWA pretraining)
PPGECGResp
Source paper
GPT-PPG: a GPT-based foundation model for photoplethysmography signals
Related Research
- Stanford wrist PPG AF (DeepBeat)
Torres-Soto & Ashley 2020, Synapse syn21985690
- WESAD
Wearable Stress and Affect Detection
- PPG-DaLiA
PPG dataset for motion compensation / HR
- Generalist vs Specialist Time Series Foundation Models: Investigating Potential Emergent Behaviors in Assessing Human Health Using PPG Signals
Saurabh Kataria, Yi Wu, Zhaoliang Chen, Hyunjung Gloria Kwak, Yuhao Xu, Lovely Yeswanth Panchumarthi
- Aether
compute-optimal physiological waveform FM family (20M–2.1B)
- AnyPPG
Foundation model