Abstract Details

Title Natural Language Processing Model to Extract Acute Abnormalities from CT Head Reports

Topic General Neurology

Presentation(s) General Neurology Posters (7:00 AM-5:00 PM)

Poster/Presentation
Number 005

Objective

To automatically extract acute brain pathology from CTH reports to relate injury features, such as injury size and location, to outcomes of interest.

Background

Natural language processing (NLP) can be used to extract information from Electronic Health Records. Yet, there is limited exploration of NLP applications to extracting information about neurologic injury from radiology reports. We propose using spaCy NeuroNER, a novel approach combining deep learning and Named Entity Recognition (NER), to evaluate non-contrast head computerized tomography (CTH) reports of acute brain injury.

Design/Methods We adapted an open source NLP model, spaCy NER, into our custom spaCy NeuroNER model by training it to extract abnormalities from CTH reports. The training set consists of 3,361 adult CTH reports from 3330 patients in the Yale Acute Brain Injury Biorepository with traumatic, ischemic, or hemorrhagic injuries. We trained the spaCy NeuroNER model to recognize and classify acute pathologies into 61 pre-defined Named Entity categories outlined by a custom dictionary of words/phrases (entities), including terms describing hemorrhage, ischemia, size and location. We performed a 9-fold cross-validation, using manually labelled folds (gold standard) to train and test the model’s performance. Overall performance was evaluated by averaging the ten models to find: precision (positive predictive value), recall (sensitivity), and F1 scores. We then created a classifier for automated exportation of identified abnormalities for future analysis.

Results The average (and range) performance metrics across 9-folds are: Precision= 97.62(93.6-99), Recall 97.64(93.4-99); F1 score= 97.62(94.8-99). Our best spaCy NeuroNER model identified 16920 true positive entities, 521 false positives and 358 false negatives. A separate classification tool was developed to export identified abnormalities for further analysis.

Conclusions This model reliably extracts and compiles entities representing acute abnormalities in CTH reports. Our next steps include: cross-validation with external datasets and multi-step output classification.

Authors/Disclosures
Victor M. Torres-Lopez, MA (Yale University) PRESENTER	Mr. Torres-Lopez has nothing to disclose.
	No disclosure on file
Gabriella Garcia, MD	Dr. Garcia has nothing to disclose.
	No disclosure on file
Alison Herman	Ms. Herman has nothing to disclose.
Alexandria Soto	Ms. Soto has nothing to disclose.
	No disclosure on file
Sam Payabvash	Sam Payabvash has nothing to disclose.
Guido J. Falcone, MD (Yale School of Medicine)	The institution of Dr. Falcone has received research support from NIH. The institution of Dr. Falcone has received research support from AHA.
Richa Sharma, MD (Massachusetts General Hospital, Brigham, Harvard)	Dr. Sharma has received research support from NIH. Dr. Sharma has received intellectual property interests from a discovery or technology relating to health care.
Lauren H. Sansing, MD	Dr. Sansing has nothing to disclose.
Kevin N. Sheth, MD, FAAN (Yale UniversityDivision of Neuro and Critical Care)	Dr. Sheth has received personal compensation in the range of $500-$4,999 for serving as a Consultant for Ceribell. Dr. Sheth has received personal compensation in the range of $500-$4,999 for serving on a Scientific Advisory or Data Safety Monitoring board for Zoll. Dr. Sheth has received personal compensation in the range of $10,000-$49,999 for serving on a Scientific Advisory or Data Safety Monitoring board for NControl. Dr. Sheth has received stock or an ownership interest from Astrocyte. Dr. Sheth has received stock or an ownership interest from Alva. The institution of Dr. Sheth has received research support from Biogen. The institution of Dr. Sheth has received research support from Novartis. The institution of Dr. Sheth has received research support from Bard. The institution of Dr. Sheth has received research support from Hyperfine. Dr. Sheth has received intellectual property interests from a discovery or technology relating to health care.
Jennifer A. Kim, MD (Yale University School of Medicine)	Dr. Kim has nothing to disclose.

��ɫ��

��ɫ��