Harmonized representation learning on dynamic EHR graphs.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Deep learning methods for electronic health records (EHR) show promise but are limited by the inherent inconsistencies and lack of detail in structured diagnosis codes.
• Existing EHR representation learning often overlooks the rich information present in unstructured clinical notes, hindering accurate patient condition representation.

Purpose of the Study:

• To develop a unified graph representation learning framework, HORDE, that integrates heterogeneous medical entities from both structured and unstructured EHR data.
• To improve the robustness and clinical value of EHR representations by incorporating natural language from clinical notes.
• To enhance downstream clinical tasks and enable novel analyses of diagnosis code consistency.

Main Methods:

• Proposed HORDE, a unified graph representation learning framework for embedding heterogeneous medical entities.
• Fused information from structured diagnosis codes and unstructured clinical notes (natural language).
• Evaluated HORDE on clinical tasks including subsequent code prediction and patient severity classification.

Main Results:

• HORDE significantly improved performance on conventional clinical tasks compared to existing methods.
• Demonstrated robustness to inconsistencies in structured EHR codes.
• Showcased promising results in a novel analysis of diagnosis code assignment consistency.

Conclusions:

• Integrating unstructured clinical notes with structured data via HORDE enhances EHR representation learning.
• The HORDE framework offers improved accuracy and robustness for clinical tasks and provides new insights into EHR data quality.