{"title":"INCIDENTAL CORONARY ARTERY CALCIUM REPORTING ON CHEST CT SCANS USING NATURAL LANGUAGE PROCESSING: INSIGHTS FROM VETERANS HEALTH ADMINISTRATION","authors":"","doi":"10.1016/j.ajpc.2024.100763","DOIUrl":null,"url":null,"abstract":"<div><h3>Therapeutic Area</h3><div>ASCVD/CVD Risk Assessment</div></div><div><h3>Background</h3><div>Coronary artery calcium (CAC) is the strongest predictor of cardiovascular events. CAC can be identified on non-cardiac chest CTs, but reporting is inconsistent. We developed a natural language processing (NLP) algorithm to identify incidental CAC reporting on non-gated chest CT reports and described patterns in CAC reporting across the Veterans Health Administration (VHA).</div></div><div><h3>Methods</h3><div>We identified non-cardiac CT scan reports across the VHA between 2006-2024. We developed an NLP algorithm by creating Regex rules to detect mentions of CAC (none, mild, moderate, severe, or unclassified). We manually annotated 1,060 reports as the gold standard for algorithm development. We iteratively refined an NLP algorithm based on accuracy with the development scans. We validated the algorithm's performance on an independent sample of 1,000 scans and applied the algorithm to all non-cardiac chest CTs in the VHA over the study period. We described the frequency of CAC reporting over time in addition to facility-level variation.</div></div><div><h3>Results</h3><div>Across 1,000 validation reports, the algorithm had a sensitivity of 99% and a positive predictive value (PPV) of 94% for CAC being mentioned in the CT report. Among reports in which CAC was mentioned, the algorithm had a 99% sensitivity and 97% PPV for correctly noting the presence of CAC. The algorithm had a 96% accuracy for correctly detecting the reported CAC severity.</div><div>There were 6,825,889 non-cardiac chest CTs between January 2006 and March 2024 in the VHA. The presence or absence of CAC was described in 2,519,296 reports (37%). CAC reporting was highest among lung cancer screening CTs (49%). CAC reporting increased over time (Table). In 2023, reporting ranged from 0% to 63% across 128 VA facilities.</div><div>Among CTs that reported CAC presence or absence, CAC was described as present on 2,425,416 reports (96%). Among CTs that reported CAC presence, CAC severity was unclassified in 56%, mild in 16%, moderate in 13%, and severe in 15% of scans.</div></div><div><h3>Conclusions</h3><div>CAC is not reported on a majority of non-cardiac chest CTs in a large national cohort, but reporting is increasing over time. Strategies to improve CAC reporting or leverage emerging automated CAC detection algorithms are needed.</div></div>","PeriodicalId":72173,"journal":{"name":"American journal of preventive cardiology","volume":null,"pages":null},"PeriodicalIF":4.3000,"publicationDate":"2024-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"American journal of preventive cardiology","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2666667724001314","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"CARDIAC & CARDIOVASCULAR SYSTEMS","Score":null,"Total":0}
引用次数: 0
Abstract
Therapeutic Area
ASCVD/CVD Risk Assessment
Background
Coronary artery calcium (CAC) is the strongest predictor of cardiovascular events. CAC can be identified on non-cardiac chest CTs, but reporting is inconsistent. We developed a natural language processing (NLP) algorithm to identify incidental CAC reporting on non-gated chest CT reports and described patterns in CAC reporting across the Veterans Health Administration (VHA).
Methods
We identified non-cardiac CT scan reports across the VHA between 2006-2024. We developed an NLP algorithm by creating Regex rules to detect mentions of CAC (none, mild, moderate, severe, or unclassified). We manually annotated 1,060 reports as the gold standard for algorithm development. We iteratively refined an NLP algorithm based on accuracy with the development scans. We validated the algorithm's performance on an independent sample of 1,000 scans and applied the algorithm to all non-cardiac chest CTs in the VHA over the study period. We described the frequency of CAC reporting over time in addition to facility-level variation.
Results
Across 1,000 validation reports, the algorithm had a sensitivity of 99% and a positive predictive value (PPV) of 94% for CAC being mentioned in the CT report. Among reports in which CAC was mentioned, the algorithm had a 99% sensitivity and 97% PPV for correctly noting the presence of CAC. The algorithm had a 96% accuracy for correctly detecting the reported CAC severity.
There were 6,825,889 non-cardiac chest CTs between January 2006 and March 2024 in the VHA. The presence or absence of CAC was described in 2,519,296 reports (37%). CAC reporting was highest among lung cancer screening CTs (49%). CAC reporting increased over time (Table). In 2023, reporting ranged from 0% to 63% across 128 VA facilities.
Among CTs that reported CAC presence or absence, CAC was described as present on 2,425,416 reports (96%). Among CTs that reported CAC presence, CAC severity was unclassified in 56%, mild in 16%, moderate in 13%, and severe in 15% of scans.
Conclusions
CAC is not reported on a majority of non-cardiac chest CTs in a large national cohort, but reporting is increasing over time. Strategies to improve CAC reporting or leverage emerging automated CAC detection algorithms are needed.