Labeling Noncontrast Head CT Reports for Common Findings Using Natural Language Processing.

Area of Science:

• Radiology
• Medical Informatics
• Natural Language Processing

Background:

• Timely diagnosis of acute neuroradiologic findings is crucial for patient care.
• Automated triage systems for noncontrast head CTs can potentially improve patient outcomes.
• Current methods may not efficiently prioritize emergent findings.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) approach for labeling findings in noncontrast head CT reports.
• To create a large, labeled dataset of head CT images for algorithm development.
• To enable the development of algorithms for emergent finding detection and reading prioritization.

Main Methods:

• Retrospective analysis of 1002 noncontrast head CT reports (2008-2013).
• Manual labeling of reports across 12 common neuroradiologic finding categories.
• Encoding reports using an n-gram model (unigrams, bigrams, trigrams) and training a logistic regression model.
• Model assessment using L2 regularization and 5-fold cross-validation.

Main Results:

• High model performance (AUC > 0.95) for fracture, hemorrhage, herniation, mass effect, pneumocephalus, postoperative status, and volume loss.
• Good performance (AUC > 0.85) for edema, hydrocephalus, infarct, tumor, and white-matter disease.
• Identification of key finding-specific words through coefficient analysis.
• Class output probabilities indicated predictive error in higher-performing models.

Conclusions:

• A robust approach combining logistic regression with n-gram encoding effectively labels common findings in noncontrast head CT reports.
• This method facilitates the creation of labeled datasets for developing advanced diagnostic algorithms.
• The findings support the potential for automated systems to improve the efficiency of neuroradiologic interpretation.