Wearable Prompt: In-Context Learning for Depression and Anxiety Prediction from Consumer Smart Ring Metrics

The study investigated the use of lightweight language models (LLMs) to predict depression and anxiety symptoms from Oura Ring smart ring data. The research involved 1,285 participants from a Finnish birth cohort born in 1986, who provided data from 4 to 8 days. Data on activity, sleep, heart rate, heart rate variability, and demographic characteristics were converted into text prompts for the Llama 3.1, BioMistral, and Qwen 2.5 models. Results showed that prompt design is critical for prediction accuracy. The Llama 3.1 model with four in-context examples achieved the best results with 0.92 accuracy, 0.82 macro-F1 score, and 0.69 F1 score for the positive class. The findings suggest that lightweight language models can be a promising tool for predicting mental health from consumer wearable device data.