Ontology-Driven Depression Detection on Social Media via Large Language Model Reasoning

Abstract

Early detection of depression is a critical challenge for global mental health, as traditional clinical assessments are often delayed by social stigma and limited medical resources. While social media offers a unique window into real-world behavior, existing automated methods often lack clinical interpretability and symbolic reasoning. This paper proposes a novel framework that integrates a formal depression ontology with Large Language Models (LLMs) to enhance screening accuracy and transparency. By transforming ontological hierarchies and deterministic inference rules into a natural language style representation, we enable LLMs to perform step-by-step symbolic reasoning beyond simple sentiment analysis. Experimental results on a dataset of 200 Twitter users demonstrate that our ontology-driven framework significantly outperforms the baseline, achieving an accuracy of 88.0% and a precision of 91.5%. The model successfully reduces false positives by verifying temporal persistence and providing a traceable reasoning chain from digital footprints to clinical symptoms. This approach offers a promising path toward explainable AI-assisted mental health monitoring.