Merlijn Sevenster PhD, Presenter: Employee, Koninklijke Philips NV

Thusitha Mabotuwana, Abstract Co-Author: Nothing to Disclose

Piotr Roman Obara MD, Abstract Co-Author: Nothing to Disclose

Yuechen Qian, Abstract Co-Author: Nothing to Disclose

Paul J. Chang MD, Abstract Co-Author: Co-founder, Stentor/Koninklijke Philips Electronics NV Technical Advisory Board, Amirsys, Inc Research Contracts, Koninklijke Philips NV Medical Advisory Board, lifeIMAGE Inc Medical Advisory Board, Merge Healthcare Incorporated

Referring physicians can provide limited clinical history information when ordering imaging studies, especially with the use of computerized physician order entry (CPOE), which may adversely affect the diagnostic accuracy of the radiological interpretation and study value. To compensate for sparse clinical information, radiologists must consult different IT systems. This is time inefficient and workflow disruptive. We evaluate an algorithm that automatically synthesizes key phrases from prior reports and presents them in a “problem list” format.  

A natural language processing algorithm was developed that parses out sections from prior radiology reports, extracts key noun phrases from the reports’ clinical history sections and filters out phrases that are duplicates or do not assert a condition (“r/o PE”). Independent of algorithm development, 7 chronic conditions were selected (such as cirrhosis and HIV) that are potentially significant. For each condition, 20 patients were selected per IRB 11-0193-E. For each patient, 1 year’s worth of radiology reports were obtained and deidentified. For each of the 140 reports, the algorithm’s key phrases were obtained and scrutinized to contain direct reference to the known condition or indirect reference.  In addition, it was checked if the patients' problem list from the EMR (Epic) contained the known condition.  

In 59% of patients, the condition was mentioned in the clinical history section of a prior radiology report. For these patients, a phrase describing the condition was correctly extracted by the algorithm in 94% of cases. The condition was mentioned in the CPOE order in 77% of cases. On all patients, the EMR problem list directly referenced the condition in 74% of patients.

When  present, key noun phrases synthesized from prior radiology reports by NLP can be more reliable than the EMR problem list. Such technology can be leveraged to fill the clinical information gap radiologists experience in their routine workflow, which is partly but not entirely addressed by the EMR. However, since relevant conditions were not always mentioned in prior radiology reports, other electronic tools to automatically extract information from the EMR also need to be developed.

Extracting clinical context from prior reports by NLP can help to fill the clinical information gap radiologists experience in their routine workflow

Sevenster, M, Mabotuwana, T, Obara, P, Qian, Y, Chang, P, Automatically Synthesizing a Patient Problem List through Key Phrase Extraction from prior Reports.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14018962.html