LLM Assistance for Pediatric Depression
Project Overview
The document explores the innovative use of Large Language Models (LLMs) in education, particularly in enhancing mental health screening among pediatric patients through the analysis of electronic health records (EHRs). It addresses the limitations of traditional depression screening methods for young individuals and introduces a zero-shot approach with LLMs to improve the extraction of depressive symptoms from free-text clinical notes. The study showcases the effectiveness of models like FLAN-T5, which can achieve high precision in identifying depressive symptoms, ultimately contributing to more consistent and efficient mental health assessments in educational settings. While acknowledging certain limitations, the findings highlight the promising role of LLMs as supportive tools in mental health diagnostics, suggesting that their integration could significantly benefit educational institutions in addressing students' mental health needs. This underscores the broader potential of generative AI to enhance educational outcomes by providing timely and accurate insights into student welfare.
Key Applications
LLM Assistance for Pediatric Depression
Context: Pediatric primary care settings focusing on young individuals aged 6-24 years experiencing depressive symptoms.
Implementation: LLMs were applied to analyze free text from electronic health records (EHRs) to extract symptoms related to depression, building on traditional screening tools like PHQ-9.
Outcomes: The study found that LLMs, particularly FLAN-T5, achieved high precision (up to 80%) in identifying depressive symptoms, improving the reliability of mental health screening.
Challenges: Challenges included the complexity of clinical notes, the need for human oversight, and the risk of overgeneralization in symptom classification.
Implementation Barriers
Technical / Ethical Barrier
Complexity of clinical notes, potential misinterpretation of PHQ-9 scores leading to diagnostic errors, and concerns about hallucinated content and biased outputs from LLMs affecting reliability.
Proposed Solutions: Employing LLMs for evidence extraction rather than direct classification to ensure human oversight and increase interpretability, while also ensuring human oversight in the diagnostic process to support clinical decision-making.
Data Barrier
Scarcity of annotated datasets for training models and the variability in symptom expression among pediatric patients.
Proposed Solutions: Using zero-shot approaches with LLMs to leverage unannotated data for symptom extraction.
Project Team
Mariia Ignashina
Researcher
Paulina Bondaronek
Researcher
Dan Santel
Researcher
John Pestian
Researcher
Julia Ive
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Mariia Ignashina, Paulina Bondaronek, Dan Santel, John Pestian, Julia Ive
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai