ARCH: Large-scale Knowledge Graph via Aggregated Narrative Codified Health Records Analysis

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Electronic health record (EHR) systems contain vast amounts of structured and unstructured clinical data.
• Challenges exist in representing, extracting, and quantifying information from complex, large-scale EHR data.
• Developing robust methods for EHR data analysis is crucial for advancing clinical research and care.

Approach:

• Proposed the Aggregated Narrative Codified Health (ARCH) records analysis to generate a large-scale knowledge graph (KG).
• ARCH derives embeddings from EHR concept co-occurrence, quantifies feature relatedness using statistical measures, and refines linkages via sparse embedding regression.
• Validated the ARCH KG using 12.5 million patient records from the Veterans Affairs (VA) healthcare system.

Key Points:

• ARCH generates high-quality clinical embeddings and KG for over 60,000 EHR concepts, achieving high accuracy in detecting similar and related concepts.
• The approach significantly improves the detection of drug side effects by integrating both codified and NLP data, outperforming existing methods like PubmedBERT, BioBERT, and SAPBERT.
• Incorporating ARCH features enhances the robustness of phenotyping algorithms and aids in sub-typing diseases like Alzheimer's, revealing distinct patient subgroups with varying mortality rates.

Conclusions:

• The ARCH algorithm effectively generates large-scale, high-quality semantic representations and knowledge graphs from both codified and NLP EHR features.
• These representations are highly valuable for a wide range of predictive modeling tasks in clinical research and practice.
• ARCH offers a powerful tool for unlocking the full potential of EHR data for improved healthcare outcomes.