Learning Low-Dimensional Representations of Medical Concepts

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Low-dimensional representations (embeddings) are valuable for various medical informatics tasks.
• Existing methods often rely solely on text data, limiting their applicability.

Purpose of the Study:

• To develop a method for learning medical concept embeddings from claims data and clinical narratives.
• To enable privacy-preserving embedding generation.
• To establish a framework for evaluating these embeddings.

Main Methods:

• Utilizing neural language modeling techniques adapted for healthcare data.
• Learning embeddings from claims data (e.g., ICD9 codes, medications, procedures, lab tests).
• Implementing a privacy-preserving approach using co-occurrence counts from clinical narratives.

Main Results:

• Successfully generated low-dimensional embeddings for diverse medical concepts.
• Demonstrated the utility of embeddings for tasks like cohort selection and patient summarization.
• Validated a privacy-preserving method for embedding generation.

Conclusions:

• Learned medical concept embeddings from claims data and clinical narratives offer significant potential for medical informatics.
• The proposed methods facilitate privacy-preserving learning and provide a framework for quantitative analysis.
• These embeddings are expected to improve efficiency and effectiveness in healthcare data analysis.