Wavelet-Based Masked Multiscale Reconstruction for PPG Foundation Models

Megha Thukral, Cyrus Tanade, Simon A. Lee, Juhyeon Lee, Sharanya Arcot Desai

Keywords: Wearables, Foundation Models, Masked Reconstruction, Discrete Wavelet Decomposition

TL;DR: Masked multiscale reconstruction uses multiresolution wavelet decomposition to pretrain a large-scale ~50K hours wavelet-based PPG foundation model, learning rich time–frequency features and transferring with superior performance across health tasks.

Abstract:

We introduce Masked Multiscale Reconstruction (MMR), a self-supervised pretraining framework for photoplethysmography (PPG) signals that leverages the discrete wavelet transform.
MMR is pretrained on $\sim$18M unlabeled 10-second PPG segments collected from over $\sim$41K smartwatch users largely in naturalistic field settings.
The pretraining task is defined to randomly mask out subsets of wavelet coefficients derived from multi-resolution decomposition of raw PPG signals and train the encoder to reconstruct them.
This enables the model to capture patterns across scales from fine-grained waveform morphology to long-term temporal dynamics crucial for diverse downstream tasks.
On 10 of 13 health-related tasks, MMR trained on large-scale wearable PPG data outperforms or matches state-of-the-art open-source PPG foundation models and other self-supervised baselines.
An ablation study of wavelet design further underscores the value of wavelet-based representations, paving the way toward robust and generalizable PPG foundation models.

Submission Number: 52