Event-Based Clinical Finding Extraction from Radiology Reports with Pre-trained Language Model.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence in Radiology

Background:

• Radiology reports contain valuable clinical information.
• Extracting this data semantically supports secondary applications like diagnosis and research.
• Existing methods lack comprehensive representations of radiological findings.

Purpose of the Study:

• To develop and validate a deep learning model for extracting detailed clinical findings from radiology reports.
• To create a new corpus of annotated radiology reports for training and evaluation.
• To demonstrate the model's generalizability across institutions and imaging modalities.

Main Methods:

• Developed an event-based annotation schema for clinical findings (lesions, medical problems).
• Utilized BERT-based deep learning architectures for entity and relation extraction.
• Trained and validated models on a corpus of 500 annotated computed tomography (CT) reports.
• Evaluated model performance on an external dataset from the MIMIC Chest X-ray (MIMIC-CXR) database.

Main Results:

• Achieved high F1 scores for finding trigger extraction (90.9-93.4%) and argument role prediction (72.0-85.6%) on the internal CT dataset.
• Demonstrated strong generalizability with F1 scores of 95.6% for triggers and 79.1-89.7% for argument roles on the MIMIC-CXR dataset.
• Successfully extracted finding events from the entire MIMIC-CXR database.

Conclusions:

• The developed deep learning model effectively extracts detailed clinical findings from radiology reports.
• The model shows excellent performance and generalizability across different institutions and imaging modalities.
• This work facilitates secondary use of radiology report data for clinical decision support and research.