• Majid Afshar, Cara Joyce, Anthony Oakey, Perry Formanek, Philip Yang, Matthew M Churpek, Richard S Cooper, Susan Zelisko, Ron Price, and Dmitriy Dligach.
• Division of Pulmonary and Critical Care Medicine, Loyola University Medical Center, Maywood, IL.
• AMIA Annu Symp Proc. 2018 Jan 1; 2018: 157-165.

AbstractAcute Respiratory Distress Syndrome (ARDS) is a syndrome of respiratory failure that may be identified using text from radiology reports. The objective of this study was to determine whether natural language processing (NLP) with machine learning performs better than a traditional keyword model for ARDS identification. Linguistic pre-processing of reports was performed and text features were inputs to machine learning classifiers tuned using 10-fold cross-validation on 80% of the sample size and tested in the remaining 20%. A cohort of 533 patients was evaluated, with a data corpus of 9,255 radiology reports. The traditional model had an accuracy of 67.3% (95% CI: 58.3-76.3) with a positive predictive value (PPV) of 41.7% (95% CI: 27.7-55.6). The best NLP model had an accuracy of 83.0% (95% CI: 75.9-90.2) with a PPV of 71.4% (95% CI: 52.1-90.8). A computable phenotype for ARDS with NLP may identify more cases than the traditional model.

22 keywords...

hide…