A new foundation model called SOTER aims to treat wearable health data as the messy, multi-signal problem it actually is.
The paper, 'SOTER: A Generative Time-Series Foundation Model for Wearable Human Physiological Signals' (arXiv:2609.16804), describes a model built specifically for signals like heart rate and motion from wearable devices. It combines three parts: a backbone that tracks relationships between signal channels, a routing layer that sends data to specialized experts based on frequency band, and a decoder that can predict or fill in values at any timestamp, including irregular ones. The authors pre-trained it on 226 billion data points from five public physiological datasets, then tested the same unmodified model on six wearable benchmarks covering forecasting, classification, and missing-data imputation. It posted the best forecasting accuracy on four of six datasets, the best classification scores, and the lowest imputation error on all six, even holding up when the researchers added artificial noise to the input.
That durability matters because wearable data is rarely clean: a smartwatch drops samples, sensors run on different clocks, and heart rate and motion signals arrive at different rates. Most general-purpose time-series models assume tidier inputs than that, which is why domain-specific systems like SOTER keep beating them on real device data rather than curated benchmarks.
Worth noting: these are the authors' own benchmark numbers, not an independent replication, and 'generative foundation model' is doing some of its own marketing work here too.