An Informatics Approach to Characterizing Rarely Documented Clinical Information in Electronic Health Records:...

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Health Services Research

Background:

• Electronic health records (EHRs) contain valuable patient data, but some clinical aspects are under-documented and hard to extract.
• Advanced informatics methods are needed to identify these rare documentation patterns for research and quality improvement.

Purpose of the Study:

• To develop and validate a natural language processing (NLP) informatics approach for detecting and characterizing rarely documented elements in EHRs.
• To use spiritual care documentation as an exemplar case to demonstrate the approach's efficacy.

Main Methods:

• Fine-tuned Spiritual-BERT, an NLP model based on Bio-Clinical-BERT, using EHR data from a Midwestern US hospital (2010-2023).
• Trained the model on a manually annotated corpus, validating it with a separate corpus and GPT-4 generated synthetic notes.
• Applied the model to identify spiritual care documentation and analyze patterns across patient populations, provider roles, and clinical services.

Main Results:

• Spiritual-BERT achieved high accuracy in identifying spiritual care documentation (F1-scores: 0.938 internal, 0.832 external validation).
• Analysis of nearly 3.6 million EHR notes showed 2% of notes contained spiritual care references, with 73% of patients having it documented at least once.
• Significant documentation variations were found across provider types (chaplains: 99.4%, nurses: 1.7%, physicians: 1.2%), ethnicity, language, and diagnosis.

Conclusions:

• Advanced NLP techniques can effectively identify and characterize under-documented elements in EHRs, overcoming limitations of traditional methods.
• The approach revealed distinct documentation patterns related to provider types, clinical settings, and patient characteristics.
• This method shows promise for analyzing other types of under-documented clinical information within EHRs.