An Informatics Approach to Characterizing Rarely Documented Clinical Information in Electronic Health Records: Spiritual Care as an Exemplar.

Abstract

Electronic health records (EHRs) contain valuable patient information, yet certain aspects of care remain infrequently documented and difficult to extract. Identifying these rarely documented elements requires advanced informatics approaches to uncover clinical documentation patterns that would otherwise remain inaccessible for research and quality improvement.This study developed and validated an informatics approach using natural language processing (NLP) to detect and characterize rarely documented elements in EHRs, using spiritual care documentation as an exemplar case.Using EHR data from a Midwestern US hospital (2010-2023), we fine-tuned Spiritual-BERT, an NLP model based on Bio-Clinical-BERT. The model was trained on 80% of a manually annotated, gold-standard corpus of EHR notes, and its performance was validated using the remaining 20% of the corpus, alongside 150 synthetic notes generated by GPT-4 and curated by clinical experts. We applied Spiritual-BERT to identify spiritual care documentation and analyzed patterns across diverse patient populations, provider roles, and clinical services.Spiritual-BERT demonstrated high accuracy in capturing spiritual care documentation (F1-scores: 0.938 internal validation, 0.832 external validation). Analysis of nearly 3.6 million EHR notes from 14,729 older adults revealed that 2% of clinical notes contained spiritual care references, while 73% of patients had spiritual care documented in at least one note. Significant variations were observed across provider types: chaplains documented spiritual care in 99.4% of their notes, compared to 1.7% for nurses and 1.2% for physicians. Documentation patterns also varied based on ethnicity, language, and medical diagnosis.This study demonstrates how advanced NLP techniques can effectively identify and characterize rarely documented elements in EHRs that would be challenging to detect through traditional methods. This approach revealed distinct documentation patterns across provider types, clinical settings, and patient characteristics, with promise for analyzing other under-documented clinical information.