Building a wearable health model from scratch for each clinical endpoint costs millions in labeled data and months of engineering time per outcome.
Thirty-five health outcomes used to require thirty-five separate models
Most wearable health research requires expensive retrospective labeling for every new clinical outcome, and that bottleneck makes scaling past a handful of endpoints nearly impossible. The field has been stuck rebuilding the foundation every single time.
One encoder now handles cardiovascular, metabolic, sleep, and mental health signals simultaneously
Google Research Introduces SensorFM ingests 34 one-minute aggregate features from five sensors, including PPG, accelerometer, EDA, skin temperature, and altimeter, across a 24-hour context window, then outputs a pretrained representation researchers can fine-tune for downstream clinical tasks. The backbone is a ViT-1D encoder trained with a masked-autoencoder objective. Four model variants scale from 138K to 110M parameters, each paired with a proportional slice of the training corpus.
Clinical AI researchers are the first to feel this shift
- Digital health researchers who need pretrained wearable representations without building their own data pipelines from scratch
- Clinical ML engineers evaluating models against IRB-approved prospective studies rather than synthetic benchmarks
- Health AI product teams who need a single encoder that covers cardiovascular, metabolic, sleep, and mental health endpoints in one pass
The pretraining corpus spans over 5 million consented participants across 100-plus countries, sampled from 20-plus Fitbit and Pixel Watch models between September 2024 and September 2025. That geographic and device breadth is what separates this from every prior wearable model, and it sets a bar that academic labs cannot easily replicate.
A trillion minutes of sensor data just became a public research asset
- Fine-tune the base encoder on your own labeled cohort for specific clinical outcomes
- Benchmark against 35 evaluated endpoints across cardiovascular, metabolic, sleep, and mental health tasks
- Select model size from XXS to Base depending on compute budget and downstream task complexity
- Apply zero-shot or few-shot transfer to new health prediction tasks without relabeling from scratch
Pricing not listed — check our directory.
SensorFM was evaluated on 13,985 subjects across three separate IRB-approved prospective studies, which is a meaningful methodological distinction, but the model is not a clinical device and downstream deployment would still require regulatory review in any medical context.
Apple has its own wearable sensor research program and has published work on health prediction from Watch data, though nothing at this pretraining scale has been publicly documented. If Google releases SensorFM weights broadly, it changes who can afford to do serious wearable health AI research overnight.
The cost of entry into wearable health AI just dropped for everyone outside Big Tech
The research implications here extend well beyond Google’s own product roadmap, and any team currently training health models on small labeled datasets should be watching this closely. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.