Using language models to identify relevant new information in inpatient clinical notes

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Health Information Management

Background:

• Redundant information in electronic health records (EHR) is prevalent.
• This redundancy can hinder clinician efficiency and patient care.
• Automated methods are needed to distinguish new information from repeated data.

Purpose of the Study:

• To investigate language models for identifying new information in inpatient clinical notes.
• To evaluate the performance of these models against expert-defined standards.
• To quantify the extent of redundant information in clinical notes.

Main Methods:

• Utilized language models to process inpatient clinical notes.
• Developed and evaluated methods for identifying novel versus redundant information.
• Compared model performance using precision, recall, and F1-measure metrics.

Main Results:

• The optimal language model achieved a precision of 0.743, recall of 0.832, and F1-measure of 0.784.
• Inpatient and outpatient progress notes showed similar redundancy rates (approx. 76.6%).
• Advanced practice providers' notes exhibited higher redundancy than physicians' notes.

Conclusions:

• Language models show promise for identifying new information in EHR notes.
• Automated redundancy detection can support clinical information synthesis.
• Future work will incorporate semantic analysis and visualization for enhanced information retrieval.