Multimodal Self-Supervised Learning for Wearable Sleep Staging Using Photoplethysmography and Accelerometer Signals
Juhyeon Lee, Simon A. Lee, Cyrus Tanade, Viswam Nathan, Megha Thukral, Hao Zhou, Keum San Chun, Sharanya Arcot Desai (Samsung Research America)
Summary
Joint SSL pretraining of PPG and accelerometer encoders on 23,000 hours of unlabeled wearable data, then a causal Mamba sequence model for staging so it could run in real time. On DREAMT (dreamt), 4-class staging improves over fully supervised baselines: AUROC 0.59 → 0.72, Cohen's κ 0.06 → 0.30. The pretraining corpus source isn't named in the abstract (authors are Samsung; likely Galaxy Watch data, unconfirmed).
Key facts
- Form Factor
- wrist
- License
- Not stated
Reported benchmarks
Sleep staging
DREAMT · AUROC ↑
DREAMT 4-class staging; SSL PPG+ACC pretrain (23,000 h unlabeled) + causal Mamba head; vs supervised 0.59; abstract-only, split unknown
Sleep staging
DREAMT · Cohen's κ ↑
DREAMT 4-class; vs supervised 0.06; abstract-only
Updated: Oct 11, 2026
Related Research
- Sleep staging leaderboard
Benchmark results for sleep staging
- LIMA-Mamba: A lightweight and distribution-robust Mamba network for PPG-based biometrics
Research paper
- 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 Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models
Audits frozen representations of four encoders on 37,784 MC-MED encounters via concept decodability,