RadText

RadText: Radiology report NLP pipeline with OMOP CDM output

RadText processes radiology reports to de-identify text, segment and tokenize content, extract biomedical entities, detect negation, and generate structured outputs standardized to the OMOP Common Data Model (CDM) using BioC.

Key Features:

• De-identification: Removes personal identifiers from reports.
• Section segmentation and sentence splitting: Splits reports into sections and sentences for downstream analysis.
• Word tokenization: Tokenizes text into words/tokens.
• Named entity recognition (NER): Identifies and classifies entities including diseases, symptoms, and anatomical terms.
• Parsing and negation detection: Parses sentence structure and detects negations affecting clinical meaning.
• Hybrid text processing: Supports raw text analysis and local processing to support data privacy and processing flexibility.
• BioC interface: Uses BioC for compatibility with text processing tools.
• OMOP CDM standardized output: Produces structured outputs compatible with the Observational Medical Outcomes Partnership (OMOP) CDM for observational research.

Scientific Applications:

• Radiology report information extraction: Automates extraction of clinically relevant information from radiology reports.
• Observational research enablement: Generates structured data compatible with OMOP CDM for large-scale observational studies.
• Disease label classification: Classifies disease labels from radiology report text.

Methodology:

Evaluated on the MIMIC-CXR dataset with five newly annotated disease labels; achieved average precision 0.91, recall 0.94, and F-1 score 0.92 on classification tasks, and produced an annotated test set for the new disease labels to support future research.