De-identification of clinical narratives through writing complexity measures

Area of Science:

• Natural Language Processing
• Machine Learning
• Health Informatics

Background:

• Electronic health records (EHRs) contain valuable clinical narratives for research.
• De-identification is crucial for privacy when reusing EHR data.
• Current de-identification models often train on random document sets or specific types.

Purpose of the Study:

• Investigate if writing complexity can identify document subsets to enhance de-identification.
• Determine if clustering by complexity improves machine learning model performance.
• Compare complexity-based clustering to random grouping and document type designation.

Main Methods:

• Applied unsupervised clustering based on writing complexity measures.
• Grouped two corpora: Vanderbilt (4500+ varied documents) and i2b2 (889 discharge summaries).
• Compared de-identification model performance (F-measure) trained on clusters versus random/type-based groups.

Main Results:

• Vanderbilt: Complexity clusters (F=0.917) outperformed random (F=0.881).
• Increasing training subset size within clusters improved performance.
• i2b2: Complexity clusters (F=0.966) showed no significant advantage over random (F=0.965).

Conclusions:

• Clustering by writing complexity enhances de-identification models in varied clinical documentation environments.
• This approach is superior to random grouping and often document type designation.
• Further research may optimize training subset selection based on complexity.