Generating Medically-Informed Explanations for Depression Detection using LLMs

TL;DR

Addressing the need for timely, interpretable depression detection from social media, the paper presents LLM-MTD, a multi-task framework that uses a pre-trained large language model to simultaneously classify posts for depression and generate medically-grounded explanations. The approach optimizes a combined loss $\mathcal{L}_{total} = \lambda \mathcal{L}_{cls} + (1-\lambda)\mathcal{L}_{gen}$, balancing accuracy with explanation quality. On the Reddit Self-Reported Depression Dataset (RSDD), LLM-MTD achieves state-of-the-art AUPRC and overall improvements over baselines, with human evaluators rating explanations as highly relevant, complete, and medically accurate. The work demonstrates that integrating explainability into depression-detection pipelines enhances interpretability and clinical utility, and points to future work in temporal and multimodal data integration and broader generalization.

Abstract

Early detection of depression from social media data offers a valuable opportunity for timely intervention. However, this task poses significant challenges, requiring both professional medical knowledge and the development of accurate and explainable models. In this paper, we propose LLM-MTD (Large Language Model for Multi-Task Depression Detection), a novel approach that leverages a pre-trained large language model to simultaneously classify social media posts for depression and generate textual explanations grounded in medical diagnostic criteria. We train our model using a multi-task learning framework with a combined loss function that optimizes both classification accuracy and explanation quality. We evaluate LLM-MTD on the benchmark Reddit Self-Reported Depression Dataset (RSDD) and compare its performance against several competitive baseline methods, including traditional machine learning and fine-tuned BERT. Our experimental results demonstrate that LLM-MTD achieves state-of-the-art performance in depression detection, showing significant improvements in AUPRC and other key metrics. Furthermore, human evaluation of the generated explanations reveals their relevance, completeness, and medical accuracy, highlighting the enhanced interpretability of our approach. This work contributes a novel methodology for depression detection that combines the power of large language models with the crucial aspect of explainability.