Semi-automated topic identification for radiation oncology safety event reports using natural language processing and...

Area of Science:

• Medical Physics
• Health Informatics
• Radiation Oncology

Background:

• Incident learning in healthcare generates valuable safety and quality data.
• Analyzing large volumes of unstructured text data from incident reports is challenging.

Purpose of the Study:

• To develop and validate a semi-automated method for categorizing safety-related event reports.
• To apply this method to Radiation Oncology safety reports.

Main Methods:

• A correlated topic model (CTM) with 50 topics was fitted to 7174 safety-related event reports from a Radiation Oncology department (2012-2021).
• Expert validation involved assigning consensus topic labels (CTLs) using a "top-10 tokens" approach and a "case-reports" approach (reports ≥90% about a topic).

Main Results:

• The model's validity was supported, with 74% (37/50) of topics achieving majority expert agreement on CTL assignment.
• The "top-10 tokens" approach assigned CTLs to 36 topics, while the "case-reports" approach assigned CTLs to 18 of 20 applicable topics.
• Seven novel safety topics were identified that were not previously discovered through manual review.

Conclusions:

• A semi-automated method can effectively categorize Radiation Oncology safety-related event reports by topic.
• This method offers a faster alternative to manual review of numerous reports.
• The approach facilitates the discovery of previously unidentified safety topics.