A natural language processing algorithm to extract characteristics of subdural hematoma from head CT reports

Area of Science:

• Radiology
• Natural Language Processing
• Medical Informatics

Background:

• Subdural hematoma (SDH) is a common traumatic intracranial hemorrhage.
• Radiographic features of SDH predict patient complications and outcomes.
• Extracting structured data from clinical reports is challenging.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) algorithm.
• To extract structured data from cranial computed tomography (CT) scan reports for patients with SDH.
• To assess the accuracy of the NLP algorithm in identifying key SDH characteristics.

Main Methods:

• Collected CT scan reports for patients with SDH from a single center.
• Physicians manually coded reports for SDH number, midline shift, SDH thickness, and SDH side.
• Developed an NLP pipeline using a pattern-matching approach and Apache Unstructured Information Management Architecture.
• Measured algorithm performance against physician-coded data.

Main Results:

• The NLP algorithm achieved high accuracy in extracting data.
• Accuracy for side of largest SDH was 0.84.
• Accuracy for thickness of largest SDH was 0.88.
• Accuracy for size of midline shift was 0.92.

Conclusions:

• A NLP algorithm can accurately structure key data from head CT reports.
• This NLP tool can identify important radiographic findings from electronic health records.
• The algorithm has the potential to support clinical decision-making.