Automatic Fully-Contextualized Recommendation Extraction from Radiology Reports.

Area of Science:

• Radiology Informatics
• Natural Language Processing
• Machine Learning

Background:

• Radiology reports contain crucial recommendations for patient care.
• Current automated systems often miss contextual details or specific recommendation types.

Purpose of the Study:

• To develop a neural network for comprehensive, contextualized recommendation extraction from radiology reports.
• To improve automated analysis of radiology reports beyond simple follow-up imaging recommendations.

Main Methods:

• A neural network architecture was designed to extract six key contextual elements: recommendation, time, reason, conditionality, strength, and negation.
• A unified task representation was created for named entity recognition (NER) and classification.
• A long short-term memory (LSTM) model was trained on 2272 annotated radiology reports.

Main Results:

• The LSTM model achieved 89.2% token-level performance for recommendation extraction.
• Token-level F1 scores between 85-95% were achieved for extracting contextual features.
• The system successfully extracts diverse recommendation types (imaging, biopsies, clinical correlation) in real-time.

Conclusions:

• A feasible and effective method for extracting fully contextualized recommendations from any radiology report has been developed.
• This approach demonstrates potential for generalization to other clinical entities within radiology reports.