KEEP: Integrating Medical Ontologies with Clinical Data for Robust Code Embeddings.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Effective representation of structured medical codes is crucial for machine learning in healthcare.
• Current methods present a trade-off between knowledge graph-based (formal relationships) and data-driven (empirical patterns) approaches.
• Existing methods often fail to integrate both formal knowledge and real-world patterns effectively.

Purpose of the Study:

• To present KEEP (Knowledge-preserving and Empirically refined Embedding Process), an efficient framework to bridge the gap in medical code representation.
• To combine knowledge graph embeddings with adaptive learning from clinical data.
• To enable a unified representation supporting multiple downstream applications and model architectures.

Main Methods:

• KEEP generates initial embeddings from medical knowledge graphs.
• It employs regularized training on patient electronic health records (EHR) to adaptively integrate empirical patterns.
• The process preserves ontological relationships while learning from real-world clinical data, producing final embeddings without task-specific training.

Main Results:

• KEEP outperforms traditional and Language Model-based approaches in capturing semantic relationships of medical codes.
• The framework demonstrates superior performance in predicting clinical outcomes using structured EHR data from UK Biobank and MIMIC-IV.
• KEEP exhibits minimal computational requirements, making it suitable for resource-constrained environments.

Conclusions:

• KEEP offers an efficient and effective solution for representing structured medical codes in machine learning.
• The framework successfully integrates formal medical knowledge with empirical patterns from clinical data.
• KEEP provides a versatile and computationally efficient tool for advancing machine learning applications in healthcare.