Abstract
Background/Aims Information in breast MRI reports is valuable for breast cancer research, but these data are only available in free-text reports and require resource-intensive manual abstraction. We developed and tested a Natural Language Processing (NLP) algorithm to extract information and pre-fill abstraction form from free-text breast MRI reports.
Methods We identified 465 reports for women receiving breast MRI at Group Health between 2010–2012. We developed an NLP algorithm in SAS v9.2. The algorithm extracts information of reading radiologist, laterality, parenchymal enhancement, whether computer-aided technique is used, comparison exams, clinical indications and assessment from breast MRI reports. The NLP results are compared with manual abstraction from an experienced abstractor.
Results The algorithm correctly extracts reading radiologist, laterality and whether computer-aided technique for all 465 breast MRI reports, except 1 report with inconsistent information on laterality itself. It correctly extracts 83% of 465 reports for assessment for right breast and 92% for assessment for left breast. Unstable gold standard impedes performance of the NLP algorithm for extracting parenchymal enhancement and clinical indications. There is no gold standard to show NLP performance for comparison exams yet.
Conclusions This NLP algorithm holds promise for rapid, accurate extraction of information from free-text breast MRI reports. Manual review will be faster and more accurate due to the pre-filling of the abstraction form.
