A Bayesian approach for the categorization of radiology reports

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Data

Background:

• Radiology reports contain crucial diagnostic information but often require manual review.
• Automating the analysis of unstructured text in radiology reports can improve efficiency.
• Distinguishing between positive and negative findings for specific conditions like appendicitis is a key challenge.

Purpose of the Study:

• To develop and evaluate a Bayesian filter for classifying radiology computed tomography (CT) reports.
• To accurately differentiate between positive appendicitis findings and negative reports.

Main Methods:

• A Java-based text search engine retrieved 500 unstructured radiology reports mentioning appendicitis.
• Reports were manually categorized into positive (250) and negative (250) appendicitis findings.
• A Bayesian classifier (dbacl 1.9) was trained on these categorized reports and tested on 100 new cases.

Main Results:

• The Bayesian filter achieved rapid training in approximately 2 seconds.
• The system demonstrated high accuracy, correctly categorizing all 50 positive and 50 negative appendicitis reports in under 10 seconds.

Conclusions:

• A Bayesian filter system can effectively and rapidly categorize radiology report findings.
• This automated approach achieves high accuracy in determining specific diagnoses from text.
• The Bayesian filter has broad potential applicability to various radiologic report findings and categories.