LENS: LLM-Powered Mental Health Narrative Generation from Sensor Data

This paper introduces LENS, a novel framework that leverages LLMs to generate clinically relevant narratives from multimodal sensor data for mental health assessment. The scarcity of paired sensor-text data and the inability of LLMs to directly process time-series data are key challenges addressed. The creation of a large-scale dataset and the development of a patch-level encoder for time-series integration are significant contributions. The paper's focus on clinical relevance and the positive feedback from mental health professionals highlight the practical impact of the research.

• •LENS framework bridges the gap between multimodal sensor data and LLMs for mental health assessment.
• •Addresses the challenge of scarce sensor-text datasets by creating a large-scale dataset from EMA responses.
• •Employs a patch-level encoder to integrate time-series sensor data directly into LLMs.
• •Demonstrates superior performance compared to baselines and receives positive feedback from mental health professionals.