Unsupervised Topic Modeling in a Large Free Text Radiology Report Repository

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Radiology report narratives contain critical patient health information primarily in free text.
• Variations in language and terminology hinder information extraction for research and decision support.
• Large multi-institutional repositories of free text radiology reports require effective organization and topic identification.

Purpose of the Study:

• To design and develop an unsupervised machine learning approach for capturing main concepts in radiology reports.
• To partition a large radiology report repository based on identified main foci and topics.
• To establish a framework for information extraction and retrieval from unstructured radiology data.

Main Methods:

• Modeling radiology reports in a vector space.
• Utilizing cosine similarity to compare and cluster reports.
• Applying an unsupervised machine learning approach to a repository of 1,899,482 reports from three healthcare organizations.

Main Results:

• Identification of 19 major radiology report topics within the repository.
• Successful clustering of reports according to their identified topics.
• Verification of results by a domain expert radiologist.

Conclusions:

• The developed unsupervised machine learning approach effectively organizes and identifies topics in large radiology report repositories.
• The system provides a target-based corpus and framework for advanced information extraction and retrieval systems.
• This method facilitates improved decision support, quality improvement, and biomedical research using radiology data.