Pragya Ahuja Dang MBBS, Presenter: Nothing to Disclose

Mannudeep Karanvir Singh Kalra MD, Abstract Co-Author: Research grant, General Electric Company

Matthew David Gilman MD, Abstract Co-Author: Nothing to Disclose

Subba Rao Digumarthy MD, Abstract Co-Author: Nothing to Disclose

Thomas Schultz, Abstract Co-Author: Nothing to Disclose

Keith J. Dreyer DO, PhD, Abstract Co-Author: Employee, Perceptics, LLC Medical Advisor, Perceptics, LLC Medical Advisor, Agfa-Gevaert Group Medical Advisor, FUJIFILM Holdings Corporation Medical Advisor, General Electric Company Medical Advisor, McKesson Corporation Medical Advisor, AuntMinnie.com Medical Advisor, AMICAS, Inc Medical Advisor, Dynamic Imaging, LLC Medical Advisor, Ascom Holding AG Medical Advisor, Bracco Group Medical Advisor, Merge Healthcare Medical Advisor, Emageon, Inc Medical Advisor, RCG HealthCare Consulting Medical Advisor, Valley Radiology Medical Associates, Inc Medical Advisor, The Elizabeth Wende Breast Clinic Medical Advisor, ISCI Medical Advisor, Siemens AG Medical Advisor, Barco nv Medical Advisor, Hue AS Medical Advisor, Planar Systems, Inc Medical Advisor, Vital Images, Inc Medical Advisor, Commissure, Inc Medical Advisor, TeraRecon, Inc Medical Advisor, Mercury Computer Systems, Inc Medical Advisor, IBM Corporation Medical Advisor, Hewlett-Packard Company Medical Advisor, EMC Corp Medical Advisor, Phase Forward Incorporated Medical Advisor, Winchester Systems, Inc Medical Advisor, Dell Inc Medical Advisor, Eastman Kodak Company Medical Advisor, Amirsys, Inc Medical Advisor, Reed Elsevier Committee member, Diagnostic Imaging Committee member, AuntMinnie.com Committee member, Imaging Economics Author, Springer Science+Business Media Deutschland GmbH Shareholder, Microsoft Corporation Shareholder, Intel Corporation Shareholder, IBM Corporation Shareholder, Hewlett-Packard Company Shareholder, Dell Inc Shareholder, General Electric Company Shareholder, Siemens AG Shareholder, Google Inc

Natural Language Processing (NLP) accurately identifies presence or absence of findings in  radiology reports without taking into account clinical information. The purpose of our study was to determine the clinical relevance of radiology reports classified as positive (Fpos) or negative (Fneg) for radiology findings in a large radiology report database.

Radiology reports from 1995-2007 were exported from the RIS to create a comprehensive radiology report database which was analyzed by NLP. Consecutive chest CT reports (M:F=77:73, mean age=54.6 years) categorized as Fpos by NLP and 50 consecutive chest CT reports (M:F=26:24, mean age=45.6 years) categorized as Fneg by NLP were selected from this database irrespective of clinical indications. These reports were randomized and two radiologists blinded to NLP classification, independently graded the reports using clinical indication from HIS on a 5-point scale for clinically significant findings (Cpos) (1=no Cpos, 3=indeterminate Cpos, 5=definitely Cpos). Reports with probable or definite Cpos were further classified into 4 finding subcategories (1=none, 2=stable, 3=change from prior, 4=new). Data were analyzed to determine Cpos in reports categorized as Fpos and Fneg. The resulting proportions were projected on a large radiology report database (from 1996-2007 chest CT, n=209,033) categorized by NLP for Fpos and Fneg.

There was excellent agreement between radiologists (176/200, 88.0%). Of 150 Fpos, radiologists opined that 19 (12.7%) reports were probably and 115 (76.7%) were definitely Cpos. Of the 50 Fneg, radiologists found that 12% reports had no, probably no or indeterminate Cpos. In reports with probable or definite Cpos, most findings (98.3%) were attributable to the clinical indication and 57.4% were a new or a change in findings. Readjustment of NLP classification of a large chest CT reports database, based on the radiologists interpretation, increased yield to 89.1% from 81.6 % (classified as Fpos by NLP).

Analysis of 12 year data using NLP and manual interpretation by radiologists shows that yield of chest CT exams is substantially high (89.1%) for clinically significant findings. Most significant findings represented either new or a change from prior and attributable to clinical indications.  

Contrary to popular belief the yield of chest CT reports is very high for clinically significant findings.

Dang, P, Kalra, M, Gilman, M, Digumarthy, S, Schultz, T, Dreyer, K, Do Radiology Findings Identified by NLP Imply Clinical Significance? An Exploratory Study.  Radiological Society of North America 2008 Scientific Assembly and Annual Meeting, February 18 - February 20, 2008 ,Chicago IL. http://archive.rsna.org/2008/6010546.html