Relevant Word Order Vectorization for Improved Natural Language Processing in Electronic Health Records

Area of Science:

• Computational linguistics
• Biomedical informatics
• Machine learning

Background:

• Electronic health records (EHR) contain valuable unstructured text data.
• Existing natural language processing (NLP) algorithms are not optimized for EHR's unique characteristics.
• Automated data extraction from EHR is crucial for research and clinical applications.

Purpose of the Study:

• To introduce Relevant Word Order Vectorization (RWOV), a novel NLP technique for EHR data.
• To evaluate RWOV's effectiveness in structuring EHR text for machine learning.
• To compare RWOV's performance against existing methods for a specific clinical task.

Main Methods:

• Developed Relevant Word Order Vectorization (RWOV) focusing on word position and relevance.
• Applied RWOV to classify hormone receptor status in breast cancer patients using EHR data.
• Compared RWOV's performance using F1 score and Area Under the Curve (AUC) against other NLP methods.

Main Results:

• RWOV demonstrated comparable or superior performance to other methods in classifying breast cancer patient data.
• RWOV showed a distinct advantage in F1 scores across most classification tasks.
• RWOV significantly outperformed other methods for HER2 status prediction.

Conclusions:

• RWOV is a promising technique for structuring and analyzing unstructured EHR data.
• The method effectively leverages word order and relevance for improved NLP in clinical contexts.
• Further development of RWOV for EHR-related NLP tasks is warranted.