SCVS Main Site  |  2021 Virtual Portal  |  Past & Future Symposia
Society For Clinical Vascular Surgery

Back to 2022 Abstracts


Natural Language Processing Based Auditing Of Radiology Reports To Improve Detection Of Aortic Aneurysms
James Kaan, MD, Kayley Abell-Hart, MS, Victor Garcia, MD, Nicos Labropoulos, PhD, Janos Hajagos, PhD, Mary Saltz, MD, Joel Saltz, MD, Apostolos Tassiopoulos, MD.
Stony Brook University, South Setauket, NY, USA.

OBJECTIVES: Natural language processing (NLP) has been used in conjunction with the electronic medical record (EMR) as an effective tool to reliably identify various pathologies. The aim of this quality improvement study was to develop and optimize an EMR term-based query for the automatic detection of Aortic Aneurysms (AAs) through the application of NLP methods.
METHODS: All patient encounters within our hospital system were queried on a regular schedule over 18 months (February 2020 - August 2021). The study group consisted of both smokers and non-smokers aged ≥ 50 with imaging studies of the aorta. Radiology reports were uploaded from the EMR to Elasticsearch, an open source database and analytics engine. A custom term-based query, developed from a manually-derived lexicon of phrases, was used in conjunction with a negation detection algorithm to identify reports with evidence of AA. The query was validated against a test dataset of 2040 patients, manually labeled for presence of AA. Reports captured by the pipeline were compiled into a list and forwarded to the vascular service for review. Studies were manually read and presence of AA confirmed. Maximum diameter was measured and results were logged into an AA database for longitudinal tracking of disease progression. If AA was determined to be an incidental finding, the patientsí primary care providers were notified.
RESULTS: The NLP surveillance protocol filtered through 11,936 unique patient encounters meeting the inclusion criteria. 126 unique AAA patients (3.0-3.9cm: 74 patients, 4.0-4.9cm: 45 patients, ≥5.0cm: 7 patients) and 63 tAAA patients (4.5-4.9cm: 47 patients, ≥5.0cm: 16 patients) were captured. Of those, 35 (27.7%) AAA patients and 13 (20.6%) tAAA patients did not have a previous diagnosis. Of captured AAA patients, 33 (26.2%) were female and 8 (6.3%) were patients age <65. Sensitivity and positive predictive value was 93.1% and 86.8% respectively.
CONCLUSIONS: An NLP-assisted EMR query protocol can be an effective, accurate, and time-saving tool in AA detection and surveillance. It can be applied as an ongoing screening tool in large population cohorts, including patients outside the recommended screening criteria (i.e. patients age <65, females) and provides an opportunity to improve care for patients with AA who would have otherwise been missed.


Back to 2022 Abstracts