AI Foundation Models for Wearable Movement Data in Mental Health Research
Author Information
Author(s): Ruan Franklin Y., Zhang Aiwei, Oh Jenny Y., Jin SouYoung, Jacobson Nicholas C.
Primary Institution: Dartmouth College
Hypothesis
We hypothesize that PAT, our explainable foundation model, can deliver state-of-the-art results when fine-tuned for various actigraphy studies, even in datasets containing small amounts of labeled actigraphy data.
Conclusion
The Pretrained Actigraphy Transformer (PAT) outperforms traditional models in predicting mental health outcomes using wearable movement data.
Supporting Evidence
- PAT consistently outperformed baseline models in predicting benzodiazepine usage, SSRIs, sleep disorders, sleep abnormalities, and depression.
- Models were pretrained on a large dataset of actigraphy data, allowing for effective fine-tuning on smaller datasets.
- The use of attention mechanisms in PAT allows for better understanding of which movement patterns are most relevant to health outcomes.
Takeaway
This study created a smart computer program that can understand movement data from wearables to help predict mental health issues, even with small amounts of information.
Methodology
The study used a transformer architecture to analyze actigraphy data from over 29,000 participants, employing a masked autoencoder for pretraining and fine-tuning on specific mental health tasks.
Potential Biases
Potential biases may arise from the reliance on specific datasets and the representativeness of the participant pool.
Limitations
The study acknowledges challenges in adapting traditional models to the unique structure of actigraphy data and the limited number of participants in some studies.
Participant Demographics
Data was collected from a nationally representative sample of U.S. participants across multiple years.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website