Learning Semantic Tags from Big Data for Clinical Text Representation.

Area of Science:

• Clinical informatics
• Natural Language Processing
• Medical data analysis

Background:

• Representing medical terms and n-grams in sparse clinical reports is a significant challenge in clinical text mining.
• Existing supervised and unsupervised methods struggle with effective semantic representation.

Purpose of the Study:

• To propose a novel method for semantic-level representation of words and n-grams in clinical text.
• To address the limitations of current methods in capturing the semantic meaning of medical terms.

Main Methods:

• A novel Reference Distance Estimator (RDE) is used to calculate word distances from reference features.
• Discretization, random sampling, and merging generate new binary features interpreted as semantic tags.
• These semantic tags are derived from both words and n-grams.

Main Results:

• The proposed semantic tags significantly outperform classical bag-of-words and n-grams for heart disease risk factor extraction.
• The method achieved superior performance in the i2b2 2014 challenge.
• Semantic tags can entirely replace original text with improved prediction accuracy.

Conclusions:

• The novel semantic representation method effectively captures the meaning of medical terms and n-grams.
• Semantic tags offer a powerful alternative to traditional text representations in clinical text mining.
• This approach enables the derivation of new rules beyond the lexical level, advancing medical data analysis.