Supervised machine learning and active learning in classification of radiology reports.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Cancer Informatics

Background:

• Radiology reports contain crucial information for cancer registries.
• Manual review of large volumes of imaging reports is time-consuming and costly.
• Automated systems are needed to efficiently process these reports.

Purpose of the Study:

• To develop and evaluate an automated system for classifying radiology reports.
• To differentiate between reportable and non-reportable cancer cases.
• To integrate this system into a processing pipeline for the Victorian Cancer Registry.

Main Methods:

• Employed supervised learning methods like conditional random fields and support vector machines.
• Investigated active learning (AL) strategies to optimize training data selection.
• Pilot studies conducted at multiple healthcare sites in Australia.

Main Results:

• The reportability classifier achieved high performance: 98.25% sensitivity and 96.14% specificity.
• Active learning (AL) reduced the required training data for supervised machine learning by up to 92%.
• Demonstrated significant reduction in manual classification costs.

Conclusions:

• Active learning (AL) is a promising approach for optimizing supervised training in radiology report classification.
• The developed classifier can dramatically reduce human effort in identifying relevant cancer reports.
• The system effectively filters reports for cancer registries using a large, real-world dataset.