A Foundation Model for Wearable Movement Data in Mental Health Research.

Area of Science:

• Digital health
• Machine learning for healthcare
• Wearable sensor data analysis

Background:

• Smartwatch data offers insights into temporal behavioral trends valuable for mental health research.
• Foundation models for analyzing health wearable data are underdeveloped compared to other health data types.
• Actigraphy (physical activity intensity measurement) sequences capture detailed movement patterns.

Purpose of the Study:

• To develop and evaluate the Pretrained Actigraphy Transformer (PAT), an open-source foundation model for wearable movement time series.
• To enable week-long temporal modeling and psychiatric outcome evaluation using wearable data.
• To demonstrate reproducibility on public datasets.

Main Methods:

• Designed transformers with patch embeddings and utilized self-supervised masked autoencoder pretraining.
• Pretrained PAT on minute-level, week-long actigraphy sequences.
• Trained and evaluated PAT on data from 21,538 U.S. participants in the National Health and Nutrition Examination Survey (NHANES).

Main Results:

• PAT consistently outperformed non-foundation model baselines in mental health prediction tasks (e.g., benzodiazepine/SSRI use, depression, sleep abnormalities).
• PAT showed significant improvements over LSTM, 1-D CNN, and ConvLSTM models in predicting benzodiazepine usage.
• Interpretable attention maps from PAT highlight key activity periods for clinical predictions, enhancing model transparency.

Conclusions:

• PAT is an adaptable, scalable, and easily deployable foundation model for advancing clinical insights from wearable sensor data.
• The model offers potential for improved mental health research and clinical applications.
• PAT demonstrates the value of foundation models in analyzing complex time-series data from wearables.