Natural language processing to identify ureteric stones in radiology reports

Area of Science:

• Radiology
• Medical Informatics
• Natural Language Processing

Background:

• Natural Language Processing (NLP) automates data extraction from unstructured text.
• NLP is valuable for cohort building in radiological epidemiological studies.
• Assessing NLP accuracy for ureteric stone detection in CT KUB reports is crucial.

Purpose of the Study:

• To evaluate the accuracy of NLP in identifying patients with ureteric stones from CT KUB reports.
• To determine the sensitivity and specificity of the NLP tool for ureteric stone detection.

Main Methods:

• Retrospective analysis of 1874 CT KUB reports from one year.
• Automated classification of reports for ureteric stones using a local NLP tool.
• Manual review and refinement of NLP classification to enhance accuracy.

Main Results:

• NLP identified ureteric stones in 27% of reports, compared to 36% manually.
• The NLP tool achieved 85% accuracy, with 66% sensitivity and 95% specificity.
• Errors stemmed from misspellings, syntax variations, terminology, pluralization, and inclusion of clinical request details.

Conclusions:

• The NLP tool exhibits high specificity but low sensitivity for detecting ureteric stones.
• Limitations include a lack of tailored feature extraction, an incomplete medical lexicon, and report heterogeneity.
• Enhancements in NLP tools are needed to improve data extraction accuracy in radiology.