ICD2Vec: Mathematical representation of diseases

Area of Science:

• Medical Informatics
• Computational Biology
• Health Data Science

Background:

• International Classification of Diseases (ICD) codes are the global standard for disease reporting.
• Current ICD codes use a hierarchical structure with human-defined relationships.
• Vector representations can capture complex, nonlinear relationships within medical ontologies.

Purpose of the Study:

• To develop a universal framework, ICD2Vec, for creating mathematical representations of diseases.
• To validate ICD2Vec by assessing semantic and biological relationships between vectorized ICD codes.
• To introduce IRIS, a novel risk score derived from ICD2Vec, and evaluate its clinical utility.

Main Methods:

• Developed ICD2Vec to encode disease information into mathematical vectors.
• Mapped symptom and disease vectors to similar ICD codes to establish arithmetical and semantic relationships.
• Validated ICD2Vec by comparing cosine similarities of vectorized codes with known biological relationships.
• Derived and tested the IRIS risk score using large UK and South Korean healthcare datasets.

Main Results:

• ICD2Vec demonstrated semantic compositionality, linking symptoms to relevant diseases (e.g., COVID-19 similarities).
• Cosine similarities from ICD2Vec significantly correlated with biological disease-to-disease relationships.
• The IRIS score showed significant predictive power for eight major diseases, including coronary artery disease (CAD).
• IRIS improved risk prediction for CAD when combined with established risk factors.

Conclusions:

• ICD2Vec provides quantitative, semantically rich vector representations of ICD codes, correlating with biological significance.
• The IRIS score, derived from ICD2Vec, is a significant predictor of major diseases.
• ICD2Vec offers a valuable tool for diverse research and clinical applications, with significant clinical implications.