Sensori

A self-supervised foundation model that learns general-purpose representations of human health and disease directly from 24 hours of raw tri-axial wrist movement.

It was developed and evaluated across four population-based cohorts from the United Kingdom (UK Biobank, English Longitudinal Study of Ageing), China (China Kadoorie Biobank) and the United States (National Health and Nutrition Examination Survey), comprising 122,640 participants contributing 683,617 person-days of free-living recordings. Results are reported in the paper and on the project website.

• Code and tutorials: https://github.com/OxWearables/Sensori
• Project website: https://oxwearables.github.io/Sensori
• Licence: Sensori is available under the Academic Use Licence, which permits internal academic, non-commercial research subject to its conditions. For commercial use, contact Oxford University Innovation as described in the licence.

Usage

git clone https://github.com/OxWearables/Sensori.git
cd Sensori

conda create -n sensori python=3.13 pip
conda activate sensori
pip install -e .

python scripts/get_npy.py --file /path/to/participant_001.cwa --output /path/to/processed
python -m sensori.inference --data-path /path/to/processed

get_npy.py turns a raw recording into the (2880, 300, 3) float32 arrays the model expects (10 Hz, units of g, one complete calendar day each); sensori.inference downloads these weights on first use and writes one embedding per valid day. Runnable notebooks for preprocessing, activity recognition and health probing are in tutorials/.

Citation

A preprint is forthcoming; citation details will be added on release.