Decoding substance use disorder severity from clinical notes using a large language model.

Area of Science:

• Computational linguistics
• Medical informatics
• Psychiatry

Background:

• Substance Use Disorder (SUD) significantly impacts health and society, necessitating accurate identification and treatment.
• Current diagnostic coding systems (e.g., ICD-10) lack the granularity needed for comprehensive SUD assessment.
• Clinical notes contain vital, granular SUD information but are challenging for traditional Natural Language Processing (NLP) methods to parse.

Purpose of the Study:

• To investigate the efficacy of Large Language Models (LLMs) in extracting SUD severity information from unstructured clinical notes.
• To develop and evaluate an LLM-based workflow for enhancing SUD diagnosis granularity.
• To improve risk assessment and treatment planning for individuals with SUD.

Main Methods:

• A novel workflow utilizing zero-shot learning with carefully crafted prompts for LLMs was developed.
• The open-source LLM, Flan-T5, was employed for experimentation.
• Performance was evaluated against a rule-based approach, focusing on 11 categories of SUD diagnoses.

Main Results:

• LLMs demonstrated superior recall in extracting SUD severity information compared to the rule-based approach.
• The proposed LLM workflow effectively parsed diverse clinical language, overcoming limitations of traditional NLP.
• Successful extraction of severity information was achieved across 11 distinct SUD diagnostic categories.

Conclusions:

• LLMs show significant promise for accurately extracting granular SUD severity data from clinical text.
• This approach can enhance the limitations of existing diagnostic coding systems.
• The findings support the use of LLMs for improved clinical decision-making in SUD patient care.