An Evaluation of Patient Safety Event Report Categories Using Unsupervised Topic Modeling.

Area of Science:

• Health Informatics
• Natural Language Processing
• Patient Safety

Background:

• Patient safety event data repositories contain structured and unstructured data.
• Manual review of free text narratives in safety reports is resource-intensive.
• Analyzing large volumes of safety event data presents significant challenges.

Purpose of the Study:

• To demonstrate the effectiveness of unsupervised natural language processing for analyzing patient safety event free text.
• To overcome the limitations of manual review processes.
• To identify latent topics and themes within safety event narratives.

Main Methods:

• Applied unsupervised natural language processing (topic modeling) to a large repository of patient safety event data.
• Identified topics and themes from free text descriptions using topic modeling.
• Utilized entropy measures to evaluate and compare generated topics against existing event type categories.

Main Results:

• Some generated topics aligned with pre-assigned clinical event type categories.
• Several new, previously unidentified latent topics emerged from the data.
• The unsupervised approach revealed insights not easily detectable through manual review.

Conclusions:

• Topic modeling offers a method to identify non-apparent themes in patient safety data.
• This approach has the potential to automatically reclassify ambiguously categorized events.
• Unsupervised NLP enhances the analysis and understanding of patient safety event narratives.