DS Ph.D. Qualifier Presentation | Avantika Shrestha | Tuesday, May 6, Gordon Library 303 Conference Room | Language Model-Based Depression Screening Using Clinical Interview Transcripts

DATA SCIENCE

Ph.D. Qualifier Presentation

Avantika Shrestha

Tuesday, May 6, 2025

1:00PM - 2:00PM

Gordon Library 303 Conference Room
 

Committee: 

• Elke Rundensteiner,  PhD Advisor, CS/DS

• Raha Moraffah,  Co-Advisor, CS
• Daniel Treku,  Co-Advisor, Business School

Title: Language Model-Based Depression Screening Using Clinical Interview Transcripts

Abstract: 

Depression is a disorder that affects millions globally, impacting the well-being of individuals. However, the current healthcare systems are overburdened due to low retention rates and increasing demand. There has been active research on employing the capabilities of AI models to combat this issue via automated screening using state-of-the-art language models. As such, clinical interview transcripts are useful for depression screening as they provide rich information regarding patient mental health.  Classical models like Bidirectional Encoder Representations from Transformers (BERT) are well-established for depression screening but are often used for shorter text data due to their limited token length allowance. In contrast, Large Language Models (LLMs) can utilize their large context window for large bodies of text, allowing more information to be leveraged. However, BERT models are better suited for text classification due to specialized token embeddings and LLMs perform better as text generative models. Thus, deploying the LLMs for summarization of the long transcripts and the BERT models to train for depression screening for said summarizations can be more beneficial. In this research, we compare the capabilities of LLMs, BERT variants, and our proposed two-step LLM and BERT pipeline on DAIC-WOZ benchmark dataset with 15 model pipelines. Our findings suggest that while LLMs perform better than the BERT variants on their own, leveraging the LLMs as a summarization tool for the transcripts and performing classification with BERT variants on the summaries provides better performance on average with a balanced accuracy score of 0.94.