Subcategorizing EHR diagnosis codes to improve clinical application of machine learning models

Area of Science:

• Health Informatics
• Clinical Data Analysis
• Predictive Modeling

Background:

• Electronic health record (EHR) data is vital for research and clinical decision support.
• Identifying all diagnoses within a patient encounter from EHR data is a significant challenge.
• Current methods often fail to capture the full diagnostic picture during a clinical encounter.

Purpose of the Study:

• To assess the feasibility of a novel schema for identifying and subcategorizing all structured diagnosis codes within a patient encounter.
• To improve the granularity and utility of diagnosis data extracted from EHRs.
• To enhance secondary use of EHR data for research and clinical applications.

Main Methods:

• Utilized EHR data from an interhospital transport repository with complete encounter-level data.
• Normalized diagnosis codes using the Unified Medical Language System (UMLS) and integrated additional EHR data.
• Developed and validated six diagnosis subcategories (e.g., primary, history, comorbidity, discharge).
• Employed random forest models to compare predictive performance of standard vs. subcategorized diagnoses for post-transfer mortality.

Main Results:

• Successfully identified and validated six distinct diagnosis subcategories.
• The subcategorized diagnosis model significantly outperformed the standard model in predicting mortality (testing AUROC 0.81 vs. 0.46).
• Subcategories represented varying proportions of diagnoses, including primary (10%), history (9%), problem list (20%), comorbidity (24%), discharge (6%), and unmapped (31%).

Conclusions:

• Merging structured diagnosis codes with additional EHR and secondary data sources enhances understanding of diagnostic roles.
• Subcategorization of diagnoses improves predictive model performance, demonstrating the value of granular diagnostic information.
• Further research is needed to explore benefits in prognostic model interpretation and clinical decision support.