Teaching PPG How not Who: Fixed-Effects Distillation from ECG
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.
ECG is widely used as a privileged teacher for PPG-only models (AnyPPG-style alignment, ECG reconstruction). The authors show much of what is transferred is identity (trait), not state: a per-recording mean holds 40–59% of a frozen ECG teacher's target, pooled students memorise it, and the raw alignment cosine is misleading (a constant predictor scores 0.793). Across 34 runs, more identity memorised → less state learned.
- Venue: arXiv 2610.10662 (v1 2026-10-07, cs.LG)
- Signals: PPG, ECG
- Form factor: clinical
- Venue
- arXiv 2610.10662 (v1 2026-10-07, cs.LG)
- Published
- 2026
- DOI
- Not reported
- PMID
- Not reported
- Licence
- Not reported
- Form factor
- Clinical
- Devices
- Datasets
- Subjects
- Not reported
- Code
- Not reported
- Added
- Oct 14, 2026
Abstract
ECG is widely used to teach PPG-only models, yet what it teaches is unexamined. Wearables are valued for tracking how a person's cardiovascular state changes, but ECG-to-PPG distillation mostly learns who the person is. A per-recording mean, the trait, holds 40-59% of a frozen ECG teacher's target, and pooled students memorise it without carrying it to new recordings. The raw alignment cosine misses this, since a constant predictor scores 0.793. Across 34 runs, the more identity a student memorises, the less state it learns. Fixed-effects distillation subtracts each recording's mean from prediction and target, so the trait cancels exactly, while a pooled anchor keeps it. State agreement more than doubles, within-person labels improve while age and sex do not, and the gain holds on two backbones and two further databases. Conditioning on the recording turns distillation toward the within-person changes that wearables monitor.
Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)
How it works
Devices· 1
Illustrations by form factor; product photos are added as licences clear.
Datasets used· 3
Cite
@article{teachingppghownotwhofixe2026,
title = {{Teaching PPG How not Who: Fixed-Effects Distillation from ECG}},
author = {Zhongli Wu and Zhuangzhi Gao and Yuankai Wang and Gregory Y. H. Lip and Bilal H. Kirmani and Yalin Zheng},
year = {2026},
journal = {arXiv 2610.10662 (v1 2026-10-07, cs.LG)},
eprint = {2610.10662},
url = {https://arxiv.org/abs/2610.10662}
}History
Added to the ChatPPG research hub
More on ChatPPG
- LIMA-Mamba: A lightweight and distribution-robust Mamba network for PPG-based biometrics
From the title only, it is a Mamba (selective state-space) network for PPG-based biometric recogniti
- A Comparative Study of Signal Representations Methods and Deep Learning Architectures for PPG-Based Cuffless Blood Pressure Estimation
Seven 1D→2D transforms × several CNN variants on the UCI BP database. Best: CWT, errors reported as
- A meta-learning method with reduced data requirements for training and updating deep learning models on wearable devices
This paper proposes a meta-learning method that cuts how much labelled data a new wearable needs bef