Toward Complete Structured Information Extraction from Radiology Reports Using Machine Learning.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Unstructured radiology reports contain valuable data for machine learning (ML) applications.
• Current information extraction (IE) methods in radiology often extract isolated terms, not complete facts.
• Interpretable ML systems require comprehensive data extraction for clinical trust.

Purpose of the Study:

• To develop a prototype system for extracting complete, contextualized facts from abdominopelvic radiology reports.
• To create a general-purpose system for diverse clinical informatics tasks.
• To improve the utility of radiology reports for ML-based applications.

Main Methods:

• Constructed an information schema to capture key report elements.
• Developed real-time ML models for information extraction.
• Focused on extracting complete facts, linking multiple entities and modifiers.

Main Results:

• Demonstrated the feasibility of extracting comprehensive information from radiology reports.
• Showcased the system's ability to capture findings, recommendations, clinical history, and more.
• Validated the performance of the developed ML models.

Conclusions:

• The developed system effectively extracts complete facts from radiology reports.
• This approach enhances the potential of ML in clinical informatics and radiology research.
• Interpretable ML models can be built by extracting all human-level assertions from reports.