Refining electronic medical records representation in manifold subspace

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic medical records (EMR) contain rich patient health data crucial for downstream applications.
• EMR data presents challenges like incompleteness, unstructured formats, and redundancy, necessitating effective preprocessing.
• Traditional word embedding models overlook geometric features, leading to inaccurate word similarities and loss of medical information.

Purpose of the Study:

• To develop an improved biomedical word embedding framework for processing electronic medical record data.
• To address the limitations of classic distributed word representations in capturing nuanced word similarities.

Main Methods:

• Propose a biomedical word embedding framework utilizing manifold subspace.
• Re-embed word vectors within a learned manifold subspace.
• Develop an efficient optimization algorithm based on manifold optimization and neighborhood preserving embedding.

Main Results:

• The proposed manifold subspace embedding framework enhances the representation of EMR data.
• Experimental evaluations demonstrate the framework's superiority over baseline methods in intrinsic and external tasks.
• The model achieves more accurate word similarity consistent with human judgment.

Conclusions:

• Manifold learning subspace embedding significantly improves distributed word representations for EMR texts.
• This approach reduces the complexity of processing unstructured EMR data for researchers.
• The framework holds considerable value for biomedical research and data mining.