Interpretable depression assessment using a large language model

Detecting depression from conversational text using large language models (LLMs) has garnered significant interest. However, the limited interpretability of existing methods presents a major challenge for clinical application. To address this, we propose a novel framework for automatic depression assessment, which employs LLM prompting to extract interpretable factors linked to depression from text and uses linear regression to predict severity scores. We evaluated our approach using a benchmark dataset (DAIC-WOZ; n = 186), predicting Patient Health Questionnaire (PHQ)-8 scores from clinical interview transcripts. Our method identifies key behavioral and linguistic features indicative of depression while also achieving state-of-the-art performance with a mean absolute error (MAE) of 2.91 on the test set. The resulting model further generalizes to an independent test dataset (E-DAIC; n = 86) with an MAE of 2.86. These findings suggest that interpretable LLM-based approaches hold significant promise for enhancing the clinical utility of automated depression assessment.Author summary: Depression is a common and serious mental health concern, and there is a growing need to develop fast and accessible screening tools. Recently, detecting depression from conversational texts using large language models (LLMs) has emerged as a promising solution. However, most LLM-based methods operate as “black-box” models that provide little insight into how decisions are made, limiting their use in clinical settings. In this study, we propose a novel framework to enhance the interpretability of LLM-based depression assessment. Rather than asking an LLM to provide a single overall assessment, we prompt it to evaluate a set of specific depression-related factors in the text, spanning clinical symptoms, linguistic patterns, and cognitive distortions. These factor scores are then used in a linear regression model to predict depression severity, enabling a transparent understanding of which features contribute to the prediction. When evaluated on a benchmark clinical interview dataset, our method achieves state-of-the-art performance while also identifying key behavioral and linguistic markers of depression. Moreover, the resulting model further generalizes to an independent test dataset. These findings suggest that interpretable LLM-based approaches hold significant promise for enhancing the clinical utility of automated depression assessment.