ARCH: Large-scale knowledge graph via aggregated narrative codified health records analysis

Area of Science:

• Biomedical Informatics
• Computational Health

Background:

• Electronic health record (EHR) systems contain complex, heterogeneous data including codified entries and free-text clinical notes.
• Extracting meaningful insights from EHR data is challenging due to feature representation and information extraction complexities.

Purpose of the Study:

• To develop an efficient Aggregated naRrative Codified Health (ARCH) records analysis method.
• To generate a large-scale knowledge graph (KG) integrating both codified and narrative EHR features.

Main Methods:

• ARCH analyzes EHR data from 12.5 million patients, deriving embedding vectors and quantifying feature relatedness with p-values.
• A sparse embedding regression is employed to build a KG by removing indirect feature linkages.
• The method was validated on tasks including relationship detection, drug side effect prediction, and disease phenotyping.

Main Results:

• ARCH generated high-quality clinical embeddings and a KG for over 60,000 EHR concepts.
• The system achieved high accuracy in detecting EHR concept relationships (AUCs up to 0.926) and predicting drug side effects (AUC 0.826).
• ARCH successfully sub-typed Alzheimer's disease patients into groups with different mortality rates.

Conclusions:

• The ARCH algorithm effectively generates large-scale, high-quality semantic representations and a KG from EHR data.
• These resources are valuable for diverse predictive modeling tasks in healthcare.
• ARCH enhances the performance of weakly supervised phenotyping algorithms.