Database mining unearthssecrets about imaging use

Having mastered the basic process of digitizing, storing, and retrieving medical images and reports, radiologists are now turning their attention to analyzing vast collections of the data to uncover clinical and practice trends.

Informatics scientific sessions at the RSNA meeting were packed with examples of how researchers have begun to probe databases to facilitate workflow and practice processes.

Massachusetts General Hospital led the effort this year. Dr. Pragya Dang, an MGH research fellow, presented studies that examined 10 years' worth of CT reports, more than 630,000 studies that cover nearly 200,000 patients. One investigation found that scans per patient increased only slightly among pediatric patients but jumped 50% in the adult population.

Another study of the same population found that cumulative dose rates decreased in children by 0.01% but increased for adults by 3.61%. A third study using the same database established that the incidence of multiregion CT scans grew 0.3% per year in children (under 18) and 1.9% per year in adults. Young adults (aged 19 to 40), received one-third of the multiregion (five or more) CT scans.

Perhaps the most significant of the MGH papers was one that examined trends in outpatient CT utilization before and after the implementation of an electronic order-entry decisionsupport system. This one covered a period from 2001 to 2007, during which quarterly CT volumes climbed from 8013 to a high of 14,293 before falling to 13,453. Dang and her fellow researchers concluded that implementation of the electronic orderentry system was associated with a drop in the growth of CT scanning from 3.1% to 0.2%.

The study illustrates the practical utility of database mining.

"We have case-mix–adjusted data that now show realtime decision support can reduce utilization of highcost imaging, and our physicians love it," said Dr. Keith Dreyer, vice chair of radiology informatics at MGH and a coresearcher on the paper, in an interview before the RSNA meeting. "This year we have added personalized feedback to the ordering process so that individual physicians can see where they stand compared to their peers as they are about to order an expensive diagnostic examination."

Dang, who received a research trainee award and who had eight presentations on database mining, said questions directed to her after the presentations reflected a broad spectrum of concerns.

Beyond the research from MGH, there were a number of other presentations showing the potential practice gains from database mining.

• Researchers at the University of California, San Francisco developed a PACS-based system that allows radiologists to tag imaging studies for follow- up information and query sources such as pathology reports or discharge summaries for a report back to the radiologist as soon as it was added to the patient's record. The system eliminated the need to keep a written note of the case and to actively seek out follow- up information later.

• Researchers at the University of Maryland, Baltimore created a radiology dashboard that mines multiple databases for performance and quality metrics of radiology operations such as report turnaround times. Department administrators now use the dashboard to monitor the operational health of the practice and to focus group efforts for improvements.

• A team from Singapore used natural language processing to discern changes in brain tumor status from unstructured (free-text) MRI reports. The system demonstrated low sensitivity but high specificity. The authors concluded that natural language processing may have a role in the automated classification of progression or response data from electronic databases.

• Another team from MGH used an automated search of electronic medical records to identify patients at risk for contrast-induced nephropathy. In a 50-case test cohort, a pair of human readers were 40% and 27% sensitive in predicting CIN, compared with 53% for the automated system. Human specificity was 97% compared with 69% for the automated system, but the automated system evaluated the cohort in just 9.5 seconds, compared with 53 and 24 minutes for the two human readers.