Semi‐automated topic identification for radiation oncology safety event reports using natural language processing and statistical modeling
Qian Sophia Zhang, John Kang, Kevin Lybarger, Mallory C. Glenn, Patricia A. Sponseller, Kathryn Havard Blau, Eric Ford · Medical Physics · 2025
BACKGROUND: Incident learning provides valuable narrative text about safety and quality, but the large-volume unstructured data is difficult to analyze. PURPOSE: We present a semi-automated method to categorize safety-related event reports and validate it using Radiation Oncology reports. METHODS: We analyzed English text from 7174 safety-related event reports in a Radiation Oncology department, dated between 2012 and 2021. Units of text called "tokens" were preprocessed for meaningful content, resulting in 4216 tokens. 6760 reports had at least one filtered token. We fit a correlated topic model (CTM) with K = 50 topics, estimating Bayesian posterior probability distributions over tokens and proportions of each report that were about each topic. To validate the model by assessing for expert agreement on meaningful topic labels for topics, five experts independently assigned topic labels to topics by reading the top-10 most probable tokens per topic and tokens' posterior probabilities ("top-10 tokens" approach), and by separately reading any reports that were estimated to be at least 90% about a topic ("case-reports" approach). Consensus topic labels (CTLs) such as "Brachytherapy" or "Orders" were assigned to topics. RESULTS: Of 50 modeled topics, 37 (74%) had a majority agreement of experts on the CTL assignment, supporting the model's validity. 36 topics had a CTL assigned to them via the top-10 tokens approach. Of 50 topics, 20 had at least one report that was ≥ 90% about the topic such that the case-reports approach applied. 18 of the 20 had a CTL assigned via that approach. Of 55 CTLs assigned, seven were not found by prior labeling of reports' topics during review of reports over the years. CONCLUSIONS: We demonstrated a semi-automated method to categorize Radiation Oncology safety-related event reports by topic, offering an expedited alternative to a person reading many reports and enabling the identification of topics not discovered by reading individual reports.