"Note Bloat" impacts deep learning-based NLP models for clinical prediction tasks.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic Health Records (EHR) implementation has led to increased use of copy-paste functionality.
• This practice results in "bloated" clinical notes with significant textual redundancy.
• The impact of this note bloat on Natural Language Processing (NLP) models is not well understood.

Purpose of the Study:

• To investigate the effect of redundancy in EHR notes on deep learning-based NLP models.
• To assess how deduplication methods influence the performance of clinical prediction tasks.
• To determine the vulnerability of NLP models to adversarial attacks using duplicated text.

Main Methods:

• Utilized a publicly available EHR database for four distinct clinical prediction tasks.
• Applied two deduplication techniques to identify and quantify redundancy in hospital notes.
• Evaluated the performance of deep learning NLP models before and after redundancy removal.

Main Results:

• Identified substantial amounts of redundancy within EHR hospital notes.
• Removing redundancy generally had a minimal negative impact on downstream model performance.
• Deduplication sometimes led to significantly improved model performance on clinical prediction tasks.
• Demonstrated that adding duplicated notes can negatively impact model predictions, turning correct predictions into incorrect ones.

Conclusions:

• Textual redundancy in EHR notes significantly affects NLP models used for clinical prediction.
• Awareness of clinical context and the development of robust modeling methodologies are crucial.
• Effective and reliable NLP systems in healthcare require careful consideration of data quality and redundancy.