Interpretable depression assessment using a large language model

Area of Science:

• Computational linguistics
• Clinical psychology
• Artificial intelligence

Background:

• Automated depression detection from conversational text is gaining traction.
• Existing methods using large language models (LLMs) lack interpretability, hindering clinical use.
• Interpretability is crucial for understanding model predictions and ensuring clinical validity.

Purpose of the Study:

• To develop a novel, interpretable framework for automatic depression assessment using LLMs.
• To extract clinically relevant, interpretable factors from conversational text indicative of depression.
• To predict depression severity scores using these interpretable factors and linear regression.

Main Methods:

• Utilized LLM prompting to identify and extract interpretable depression-related features from clinical interview transcripts.
• Employed linear regression to predict depression severity scores based on extracted features.
• Evaluated the framework on the DAIC-WOZ benchmark dataset and an independent E-DAIC dataset.

Main Results:

• The proposed method achieved state-of-the-art performance in predicting depression severity (PHQ-8 scores).
• Achieved a mean absolute error (MAE) of 2.91 on the DAIC-WOZ test set.
• Demonstrated generalization capabilities with an MAE of 2.86 on the independent E-DAIC test set.

Conclusions:

• Interpretable LLM-based approaches can effectively identify key behavioral and linguistic features of depression.
• The developed framework shows significant promise for enhancing the clinical utility of automated depression assessment tools.
• This approach bridges the gap between LLM capabilities and the need for transparency in clinical decision-making.