Automated Insomnia Phenotyping from Electronic Health Records: Leveraging Large Language Models to Decode Clinical...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Research

Background:

• Insomnia is common but frequently underdiagnosed.
• Inconsistent electronic health record (EHR) documentation hinders insomnia research.
• Accurate phenotyping is crucial for understanding treatment patterns and outcomes.

Purpose of the Study:

• To develop and evaluate an automated system for insomnia phenotyping using generative large language models (LLMs).
• To improve the identification and documentation of insomnia within unstructured clinical notes.
• To establish a scalable framework for clinical phenotyping of underdiagnosed conditions.

Main Methods:

• Utilized generative large language models (LLMs), specifically Llama 70B and Llama 405B.
• Employed prompt engineering with few-shot learning and chain-of-thought reasoning.
• Evaluated the system on two diverse corpora: MIMIC-III (inpatient) and University of Kansas Health System (KUMC) (outpatient) clinical notes.

Main Results:

• Achieved high F1 scores: 93.0 on the MIMIC corpus and 85.7 on the KUMC corpus.
• Demonstrated superior performance compared to domain-adapted BERT-based classifiers.
• Successfully phenotyped insomnia directly from unstructured clinical text.

Conclusions:

• The developed LLM-based framework provides a scalable and interpretable solution for clinical insomnia phenotyping.
• This approach can enhance the analysis of insomnia prevalence, treatment, and outcomes.
• The methodology serves as a model for phenotyping other underdiagnosed conditions in EHRs.