Reproduce “WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Peng Cao, Zhijian Yang, Tennison Liu, Jonathan Wang, Jiang Wu, Magdalena Proszewska, Arvind Pillai, Mingwu Gao and 12 more
WavesFM, a hierarchical foundation model developed by Google Research, learns unified subject-level representations from multi-day, high-resolution wearable sensor waveforms by simultaneously capturing micro-scale morphological signatures and macro-scale longitudinal dynamics. It demonstrated superior performance across 58 health-related tasks and robustness to data missingness, improving upon existing methods.
Key insights
- Decomposing the learning process into separate morphological and temporal stages effectively addresses both computational complexity and multi-scale physiological dynamics.
- Jointly modeling micro-scale waveform morphology and macro-scale temporal rhythms through a hierarchical approach leads to more accurate and comprehensive health phenotyping.
- Aggressive multi-scale masking combined with a dual-branch decoder for masked embedding prediction prevents interpolation and identity shortcuts, forcing the model to learn robust long-horizon dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.