Patient Representation From Structured Electronic Medical Records Based on Embedding Technique: Development and...

Area of Science:

• Computational Medicine
• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Secondary use of structured electronic medical record (sEMR) data faces challenges due to data diversity, sparsity, and high dimensionality.
• Effective data representation is critical for advanced sEMR data applications.

Purpose of the Study:

• To adapt natural language processing embedding techniques for sEMR data representation.
• To evaluate the feasibility and superiority of embedding-based features and patient representations in clinical settings.

Main Methods:

• Utilized the Skip-gram algorithm to embed 13,757 medical concepts from 104,752 patient records into 200-dimensional vectors.
• Patient representation was achieved by averaging the vectors of all medical concepts within a record.
• Evaluated concept associations using cosine similarity and patient clustering using Hopkins statistic, Silhouette index, and Davies-Bouldin index.

Main Results:

• Reduced patient representation dimensionality from 13,757 to 200.
• Achieved high cosine similarity (0.973) for clinically relevant concepts.
• Embedding-based patient clustering demonstrated superior performance (SI 0.862, F1 scores up to 0.944) compared to reference methods.

Conclusions:

• Embedding-based representations effectively capture clinical associations among medical concepts.
• Patient-level embeddings offer a user-friendly input for machine learning, enhancing performance.
• This approach is expected to benefit various secondary uses of sEMR data.