Abstracts from The 38 th Annual National Neurotrauma Symposium July 11–14, 2021 Virtual Conference

Abstract

Introduction: Artificial intelligence (AI) has become a powerful tool in augmenting clinical research. However, there are certain types of clinical data such as free-text in physician notes which make it difficult to use for data analysis. Natural language processing (NLP) is a form of AI that looks at discrete words as well as syntax to assist in categorizing bodies of text. Cranial hematomas such as epidural (EDH), subarachnoid (SAH), subdural (SDH), intraparenchymal (IPH), or a combination of these each have unique descriptors in radiologist's notes as well as labeling in neurosurgical consult notes. For our project, we assessed the use of open-source NLP to accurately label traumatic cranial injuries across consultation notes and radiology reports. Method(s): We performed a retrospective review of 60 patients presenting to the emergency department for which 51 were cranial traumas and 9 were used as negative controls. All of these participants had a neurosurgical consultation as well as radiographic imaging as part of their workup. Ground truth was established through a single reviewer, blinded looking at respective CT imaging for patients. Assessment portions of consult notes, radiologist impression, and descriptions were de-identified and analyzed through CHARTextract, an open-source NLP software. Result(s): We found that software was able to label radiologist impressions more accurately than resident consultation notes. The sensitivity for the neurosurgery note for SDH, SAH, EDH, IPH was 87.2%, 56.5%, 80.0%, 57.9% with specificity of 95.3%, 94.6%, 100%, and 95.1% respectively. The sensitivity for the radiology report for SDH, SAH, EDH, and IPH was 94.9%, 87%, 80%, 94.7% and specificity of 100%, 97.3%, 100%, and 90.2% respectively. Of the 9 included cases without neurotrauma, 1 consult note had a false positive. Discussion(s): Open-source NLP is an effective research tool to categorize traumatic head bleeds from radiology and neurosurgical notes. In addition, we identified that radiology notes may be more useful in categorizing cranial hemorrhages.We hope to use this software in the future to create a large Neurotrauma database for patient outcome analysis.