Extraction of recommendation features in radiology with natural language processing: exploratory study

Area of Science:

• Radiology and Medical Imaging
• Natural Language Processing
• Health Informatics

Background:

• Electronic radiology reports contain valuable information on follow-up imaging recommendations.
• Manual extraction of these recommendations is time-consuming and prone to error.
• Automated methods are needed to efficiently analyze large volumes of radiology data.

Purpose of the Study:

• To validate a natural language processing (NLP) program for extracting imaging recommendation features from electronic radiology reports.
• To assess patterns of recommendation features within a large-scale radiology report database.

Main Methods:

• A natural language processing program was developed and validated against manual classifications by two radiologists.
• The program analyzed 120 randomized reports for recommendation features (imaging technique, time frame).
• The validated NLP program was applied to a database of 4,211,503 radiology reports.

Main Results:

• The NLP program achieved high accuracy: 93.2% for imaging technique and 94.3% for time frame.
• Computed tomography (CT) and magnetic resonance imaging (MRI) were the most frequently recommended imaging techniques.
• A significant majority of reports (85.4%) did not specify a recommended time frame for follow-up.

Conclusions:

• Natural language processing offers an accurate method for extracting recommended imaging techniques and time frames from radiology reports.
• The study highlights a trend towards recommending advanced, higher-cost imaging modalities.
• Automated analysis of radiology reports can reveal significant patterns in clinical practice.