A Foundation Model for Wearable Movement Data in Mental Health Research

Researchers developed PAT (Pretrained Actigraphy Transformer), an open-source foundation model that analyzes wearable movement data to predict mental health outcomes including depression, sleep disorders, and medication use. Trained on data from over 21,000 U.S. participants, PAT significantly outperforms traditional deep learning models while providing interpretable insights into behavioral patterns relevant to clinical decision-making.

The development of PAT addresses a significant gap in health AI research. While foundation models have revolutionized clinical imaging and natural language processing, wearable sensor data—collected by billions of smartwatch users globally—has received comparatively limited attention from the machine learning community. This foundation model leverages transformer architectures with masked autoencoder pretraining on minute-level actigraphy sequences, enabling the model to capture week-long behavioral patterns that correlate with psychiatric outcomes.

The research builds on growing recognition that passive sensor data from consumer devices contains clinically actionable information. Smartwatches and fitness trackers generate continuous, objective measurements of physical activity that traditional clinical assessments cannot capture. By training on NHANES data, PAT benefits from a nationally representative cohort, enhancing generalizability across diverse populations—a critical consideration for health AI systems.

The performance improvements are substantial: PAT achieved 55.6% improvement over LSTMs in predicting benzodiazepine use, suggesting the transformer architecture captures temporal dependencies in actigraphy that recurrent models miss. Beyond raw accuracy, the model's attention mechanisms provide interpretability, identifying specific daily activity periods most predictive of mental health conditions. This transparency is essential for clinical adoption, as physicians require actionable reasoning rather than black-box predictions.

The availability of open-source code democratizes access to foundation model technology for researchers without massive computational resources. As wearable adoption continues accelerating, PAT establishes a reusable framework that could integrate into clinical workflows, enabling personalized mental health monitoring and early intervention. Future development likely involves multi-modal integration with other sensor types and expanded outcome prediction tasks.

• →PAT is the first open-source foundation model designed specifically for wearable actigraphy time-series analysis in mental health prediction.
• →The model achieved 55.6% improvement over LSTMs in benzodiazepine use prediction, demonstrating transformer superiority for temporal health data.
• →Training on 21,538 NHANES participants provides strong generalization potential across diverse demographic groups.
• →Attention mechanisms offer clinical interpretability by highlighting specific daily activity periods most predictive of mental health outcomes.
• →Open-source release enables wider adoption by researchers and clinicians lacking proprietary foundation model access.