A Foundation Model for Wearable Movement Data in Mental Health Research
Researchers have developed a new foundation model called PAT (Pretrained Actigraphy Transformer) specifically for analyzing wearable movement data in mental health research. This open-source model uses self-supervised learning on actigraphy sequences to predict psychiatric outcomes, outperforming traditional time-series models. PAT demonstrated significant improvements in predicting benzodiazepine use, depression, and sleep abnormalities, while also offering interpretable attention maps to highlight key activity periods. AI
IMPACT Enables more accurate and interpretable analysis of wearable sensor data for mental health research.