WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms
Researchers have introduced WavesFM, a novel foundation model designed to analyze longitudinal data from wearable sensors. Addressing the challenges of high sampling frequencies, extreme sequence lengths, and scarce labeled data in physiological monitoring, WavesFM employs a two-stage self-supervised learning framework. The first stage utilizes a segment-level encoder to extract local embeddings from short waveform segments, while the second stage uses a temporal encoder to model these embeddings over multi-day periods. This hierarchical approach effectively captures both local signal semantics and complex circadian variations without the computational burden of processing raw high-resolution data directly. Pretrained on massive datasets comprising over 6.8 million hours of recordings from 324,000 individuals for the first stage and 5.3 million hours from 10,000 individuals for the second, the model demonstrates superior performance. It was evaluated across 58 diverse tasks related to demographics, lifestyle, health conditions, and medications, showcasing its potential to significantly advance the inference of health-related phenotypes from continuous wearable sensor data like photoplethysmography and accelerometry.
Wire timeline
WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms
Researchers have introduced WavesFM, a novel foundation model designed to analyze longitudinal data from wearable sensors. Addressing the challenges of high sampling frequencies, extreme sequence lengths, and scarce labeled data in physiological monitoring, WavesFM employs a two-stage self-supervised learning framework. The first stage utilizes a segment-level encoder to extract local embeddings from short waveform segments, while the second stage uses a temporal encoder to model these embeddings over multi-day periods. This hierarchical approach effectively captures both local signal semantics and complex circadian variations without the computational burden of processing raw high-resolution data directly. Pretrained on massive datasets comprising over 6.8 million hours of recordings from 324,000 individuals for the first stage and 5.3 million hours from 10,000 individuals for the second, the model demonstrates superior performance. It was evaluated across 58 diverse tasks related to demographics, lifestyle, health conditions, and medications, showcasing its potential to significantly advance the inference of health-related phenotypes from continuous wearable sensor data like photoplethysmography and accelerometry.
cs.AI updates on arXiv.org