MIND-LLM

MIND-LLM

 

Full Title

Optimizing De-Identification and Data Utility in Mental Health Progress Notes Using Large Language Models

 

Background

Unstructured clinical text, such as mental health progress notes, is central to patient care but remains difficult to access, share, and study due to privacy, accessibility, and interoperability barriers. While large language models have shown strong performance in automatically de-identifying protected health information, most existing approaches simply remove sensitive data rather than replacing it, which can strip away important contextual and temporal information. 

This project proposes an LLM-based multi-agent system that both de-identifies and intelligently replaces PHI with plausible substitutes to preserve clinical utility while maximizing privacy. This work leverages synthetic health data generation, an emerging approach for overcoming data scarcity and privacy limitations in healthcare machine learning, to support model development and evaluation in a privacy-preserving way.

 

Aim

This seed project aims to develop and evaluate a privacy-preserving approach for de-identification of mental health notes that maintains clinical utility. To achieve this goal, this study first generated synthetic clinical notes that capture mental health symptoms and diagnoses using the SOAP (Subjective, Objective, Assessment, Plan) format and developed a LLM-powered, multi-agent de-identification system to remove identifiable health information, and evaluate the utility, fidelity, and privacy of de-identified progress notes.
 

Publications

  • Wang L, Juneja S, Naar S, Lee M, Sha K, Shah K, Smith A, Green S, Feng Y. LOGAR: A Local–Online Generate–Audit–Refine Framework for Mental Health De-identification. the 11th IEEE/ACM conference on Connected Health: Applications, Systems and Engineering Technologies; c2026.
  • Wang L, Juneja S, Naar S, Lee M, Sha K, Shah K, Green S, Feng Y. Generating High-Fidelity Synthetic Mental Health Records with Expert-in-the-Loop Verification. the 11th IEEE/ACM conference on Connected Health: Applications, Systems and Engineering Technologies; c2026

 

Conference Activities

Coming soon!

 

Funding Source

Florida State University