Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PPG-Distill: Efficient Photoplethysmography Signals Analysis via Foundation Model Distillation

Juntong Ni, Saurabh Kataria, Shengpu Tang, Carl Yang, Xiao Hu, Wei Jin

TL;DR

The paper addresses the challenge of deploying large PPG foundation models on resource-constrained wearables. It introduces PPG-Distill, a knowledge distillation framework that combines global prediction/feature matching with patch-level morphology and rhythm distillation to preserve both local waveform details and cross-patch temporal structure. Empirical results on DaLiA and StanfordAF show substantial performance gains for lightweight students while achieving up to 7× faster inference and up to 19× memory savings, enabling practical on-device deployment. This approach opens avenues for robust, efficient PPG analysis across diverse tasks in wearable health monitoring.

Abstract

Photoplethysmography (PPG) is widely used in wearable health monitoring, yet large PPG foundation models remain difficult to deploy on resource-limited devices. We present PPG-Distill, a knowledge distillation framework that transfers both global and local knowledge through prediction-, feature-, and patch-level distillation. PPG-Distill incorporates morphology distillation to preserve local waveform patterns and rhythm distillation to capture inter-patch temporal structures. On heart rate estimation and atrial fibrillation detection, PPG-Distill improves student performance by up to 21.8% while achieving 7X faster inference and reducing memory usage by 19X, enabling efficient PPG analysis on wearables.

PPG-Distill: Efficient Photoplethysmography Signals Analysis via Foundation Model Distillation

TL;DR

The paper addresses the challenge of deploying large PPG foundation models on resource-constrained wearables. It introduces PPG-Distill, a knowledge distillation framework that combines global prediction/feature matching with patch-level morphology and rhythm distillation to preserve both local waveform details and cross-patch temporal structure. Empirical results on DaLiA and StanfordAF show substantial performance gains for lightweight students while achieving up to 7× faster inference and up to 19× memory savings, enabling practical on-device deployment. This approach opens avenues for robust, efficient PPG analysis across diverse tasks in wearable health monitoring.

Abstract

Photoplethysmography (PPG) is widely used in wearable health monitoring, yet large PPG foundation models remain difficult to deploy on resource-limited devices. We present PPG-Distill, a knowledge distillation framework that transfers both global and local knowledge through prediction-, feature-, and patch-level distillation. PPG-Distill incorporates morphology distillation to preserve local waveform patterns and rhythm distillation to capture inter-patch temporal structures. On heart rate estimation and atrial fibrillation detection, PPG-Distill improves student performance by up to 21.8% while achieving 7X faster inference and reducing memory usage by 19X, enabling efficient PPG analysis on wearables.

Paper Structure

This paper contains 19 sections, 3 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Global KD
  4. PPG-Distill
  5. PPG Morphology Distillation
  6. PPG Rhythm Distillation
  7. Joint Optimization
  8. Experiment
  9. Experimental Setting
  10. Results
  11. Efficiency Analysis
  12. Conclusion and Future Work
  13. Related Work
  14. PPG Signal Analysis
  15. Foundation Model for PPG Signal
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Illustration of our motivation. PPG foundation models are pretrained and finetuned for downstream tasks, but direct deployment on wearables is costly. KD produces efficient student models suitable for wearable deployment.
  • Figure 2: Real PPG signals from the StanfordAF dataset, segmented into patches by red lines (patch size = 40).
  • Figure 3: Overall framework of PPG-Distill.
  • Figure 4: Inference throughput (Batch/s) and parameter size comparison across GPT-PPG-19m, PaPaGei, and GPT-PPG-1m.
  • Figure 5: Effect of hyperparameters ($\alpha$, $\beta$, $\gamma$) on MAE for the DaLia dataset (Teacher: GPT-PPG-19m, Student: GPT-PPG-1m).