Teaching PPG How not Who: Fixed-Effects Distillation from ECG

Q: What is the clinical evidence from this study?

Population: Continuous cardiovascular monitoring (n=643). Intervention: Fixed-effects distillation (FE-COS) vs. Pooled distillation / standard contrastive learning. Primary outcome: Within-recording state agreement on held-out recordings (PaPaGei-S backbone).

Q: What does this research mean for the field?

Standard ECG-to-PPG knowledge distillation predominantly memorizes static subject identity rather than physiological state, but fixed-effects distillation isolates dynamic state changes and more than doubles state agreement. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Zhongli Wu; Zhuangzhi Gao; Yuankai Wang; Gregory Y. H. Lip; Bilal H. Kirmani; Yalin Zheng

doi:10.48550/arxiv.2610.10662

October 7, 2026Open Access

Teaching PPG How not Who: Fixed-Effects Distillation from ECG

View Full Paper

Ask AI

Bookmark

Share

Key result

Fixed-effects distillation achieves ~2.3-fold higher within-recording state agreement versus pooled distillation.

n=643

Why the study?

ECG-to-PPG distillation models primarily learn individual identity traits rather than cardiovascular state changes, limiting wearable monitoring utility.

Key Points

To determine why ECG-to-PPG distillation transfers static individual identity rather than dynamic physiological state, and to develop an objective that prioritizes within-person physiological tracking.
Decomposed frozen ECG teacher targets into per-recording static means (traits) and within-recording variations (states) across 282 training recordings.
Designed fixed-effects distillation by subtracting recording-level means from predictions and targets to eliminate identity traits while retaining a pooled anchor.
Evaluated representations across 34 training runs, two distinct PPG backbones (PaPaGei-S and AnyPPG), and two additional validation databases.
The static trait held 40% to 59% of the target energy, allowing a constant baseline predictor to achieve an uninformative raw cosine score of 0.793 compared to 0.823 for the best linear model.
Standard cross-recording objectives memorized identity, achieving a trait agreement of 0.90 on training windows that degraded to 0.33 on held-out recordings.
Fixed-effects distillation more than doubled within-recording state agreement with the teacher across both backbones and improved dynamic within-person labels while leaving static age and sex predictions unchanged.

Structured PICO

P

Population

A machine learning methodology study utilizing over 600 synchronous ECG and PPG recordings from multiple open-access databases to improve PPG-only foundation models for continuous monitoring.

I

Intervention

Fixed-effects distillation (FE-COS) from an ECG teacher to a PPG-only model, which subtracts each recording's mean from the prediction and target to cancel the trait (identity) component, combined with a pooled anchor.

C

Comparator

Pooled distillation objectives (pooled least-squares fit, cross-recording InfoNCE, mix-NCE) that do not explicitly separate the per-recording mean from within-recording deviations.

O

Outcome

State agreement (within-recording agreement between the predicted and true target after removing the per-recording mean).

Main Result

Absolute Event Rate: 0.307% vs 0.136%

Conditioning cross-modal ECG-to-PPG distillation on the recording mean prevents the model from memorizing patient identity, significantly improving its ability to track within-person cardiovascular changes relevant for wearable monitoring.

Limitations

White noise on the PPG lowers the performance of every objective
The gain lies beyond heart rhythm and is restricted to what PPG can observe, as features like the T/R amplitude ratio remain unreadable by PPG
White noise on the PPG lowers every objective
The T/R amplitude ratio, to which PPG has no path, is not read well by any objective
On MIMIC-III, every blood-pressure read-out stayed above the zero-prediction error

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

May improve PPG model generalizability for dynamic states across patients; extends cross-modal distillation methods but remains hypothesis-generating.

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.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Wu et al. (2026) studied Continuous cardiovascular monitoring (n=643). Fixed-effects distillation (FE-COS) vs. Pooled distillation / standard contrastive learning was evaluated on Within-recording state agreement on held-out recordings (PaPaGei-S backbone). Fixed-effects distillation conditioned on the recording removed the identity trait and more than doubled within-recording state agreement on held-out data from 0.136 to 0.307 compared to pooled distillation.

synapsesocial.com/papers/6acf0f522fe143470d0586ae https://doi.org/10.48550/arxiv.2610.10662

View Full Paper

Ask AI

Bookmark

Share

Teaching PPG How not Who: Fixed-Effects Distillation from ECG

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Teaching PPG How not Who: Fixed-Effects Distillation from ECG

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading