Analiza zmožnosti velikih jezikovnih modelov za sklepanje o duševnem zdravju iz senzorskih podatkov mobilnih naprav

Depression is a widespread mental health disorder, yet current screening and diagnostic practices rely heavily on self-reporting and clinical assessments, which are often subjective and resource-intensive. Passive mobile sensing offers unobtrusive and continuous monitoring of behavioural patterns associated with mental health, but existing approaches heavily utilise traditional machine learning approaches. At the same time, large language models~(LLMs) are increasingly used across various domains. However, their application for mental health inference from structured numerical data remains unexplored. This thesis investigates the capabilities of LLMs for depression inference from mobile sensor data. We conduct a comprehensive analysis across two datasets and two task formulations: 14-day history depression prediction and prediction of changes in depression levels. We also evaluate two small LLMs, GPT-4o mini and GPT-4.1 mini, to enable privacy-preserving on-device inference. Moreover, we compare different prompting strategies to established baselines. With peak average F1-scores of 45.6% and 24.8% for the first and second tasks, respectively, our findings show that off-the-shelf small LLMs are not able to reliably detect depression. Nevertheless, this work represents the first systematic evaluation of LLMs for mobile sensing–based depression inference, providing valuable insights into their current limitations and outlining pathways toward future privacy-preserving, on-device applications.