The reports of my capabilities are greatly exaggerated – small LLMs for depression inference from mobile sensing data

Modern large language models (LLMs) trained on huge amounts of textual data excel in a number of tasks related to generation and analysis of textual and even multimodal input. Automated infer- ences related to mental health status are crucial as humanity faces an epidemic of mental illness. The potential of LLMs to tackle this problem given texts written by target participants has already been documented. However, the ability of LLMs to infer a user’s health status from mobile sensing data alone has received only limited attention, while the practical feasibility of such LLMs running directly on end-user devices has not been addressed. In this paper, we conduct a preliminary analysis of the potential of a state-of-the-art mobile-ready LLM to infer a user’s depression level from mobile sensing traces. Our investigation reveals that expectations based on the success of LLMs in other tasks are not justified in the case of inferring mental health status from sensor data. We discuss augmentations that could improve LLM-based inference in the future.