Local and Online Large Language Models for Mental Health Summaries
Dahana Moz Ruiz
Co-Presenters: Individual Presentation
College: The Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Computer Science
Faculty Research Mentor: Daehan Kwak
Abstract:
Electronic mental health records (EMHRs) often contain extensive, text-heavy reports, making it challenging for medical professionals to review them efficiently. Traditional methods of analyzing these records are time-consuming and may hinder timely interventions. This study explores the use of local and online large language models (LLMs) to address this issue by summarizing EMHRs into actionable insights.Local LLMs, such as Mistral and LLaMA, offer advantages in privacy and security, which are crucial for handling sensitive mental health data. However, local models often struggle to produce accurate results for longer records due to limitations in computational resources, model size, and insufficient fine-tuning. In contrast, online models like BART and T5 demonstrate greater summarization accuracy and scalability but raise concerns about data security and compliance with privacy regulations.This research focuses on developing a framework for EMHR summarization tailored to specific contexts, such as summarizing by doctor, event, timeline, or mental health issue. The study evaluates local and online LLMs across dimensions like accuracy, efficiency, security, and applicability. Diverse datasets of mental health records will be used to test these models, ensuring robustness and relevance.Preliminary findings highlight the trade-offs between the privacy benefits of local models and the broader capabilities of online models. Future directions include refining local LLMs to overcome token limitations and exploring hybrid approaches that leverage the strengths of both local and online models. By enabling efficient and secure summarization of EMHRs, this research aims to improve the workflow of mental health professionals and enhance patient care outcomes.By providing a structured comparison of local and online LLMs, this study aims to enhance EMHR utility in mental health care. The findings provide insights for creating secure, effective tools to improve workflows and outcomes, highlighting the potential of artificial intelligence in mental health care.