Language algorithm could keep follow up ED imaging from falling through cracks

Natural language processing can offer a useful tool to automatically detect recommendations for additional imaging of incidental findings on the radiology reports of discharged emergency department (ED) patients, according to a study published in the August issue of Annals of Emergency Medicine.

Specifically, Sayon Dutta, MD, MPH, of Massachusetts General Hospital (MGH) in Boston, and colleagues found that an automated detection system was 97 percent sensitive to discharge-relevant imaging recommendations. If implemented, the system could help improve the way important follow-up recommendations are relayed, as a review of current discharge instructions at MGH showed that half failed to note such additional imaging recommendations.

Recommendations on how to handle incidental findings at ED discharge can sometimes get lost in the busy clinical environment or get overshadowed by recommendations related to the primary reason for imaging, noted the authors. “Fortunately, most incidental findings never result in the new diagnosis of malignancy, but for those that do, failure to act on the early finding can result in delayed treatment, worse patient outcomes, and significant legal liability,” they wrote.

Dutta and colleagues suggested an ideal system for managing additional imaging recommendations would include automated insertion of recommendations into discharge instructions and redundancies that notified both the patient and primary care physician to arrange follow-up imaging.

The authors started down the path to an ideal solution by leveraging a natural language processing tool with an algorithm to identify discharge-relevant recommendations for further imaging related to incidental findings, but the process was challenging. It was difficult for the tool to parse which recommendations in a report related to incidental findings and which related to the primary reason for testing, so they elected to have the system alert the treating clinician of all imaging recommendations to allow him or her to decide which needed communication on discharge. Once testing began, it took three iterative cycles of training and validation to optimize the system.

“Our experience illustrates something intrinsic to natural language processing: algorithms must be developed and improved in an iterative fashion to account for the full range of linguistic variation that exists among medical practitioners,” wrote Dutta and colleagues.

The third time appeared to be the charm, as the final natural language processing algorithm, trained and tested using 3,235 ED reports, had 89 percent sensitivity and 98 percent specificity for detecting recommendations for additional imaging overall, and 97 percent sensitivity for detecting discharge-relevant recommendations.

“In the near future, computerized tools should allow physicians to take advantage of electronic medical records through continuous surveillance of radiology reports to provide physicians and patients information on incidental findings,” wrote Dutta and colleagues. “Whether such activities will improve outcomes remains to be seen.”

Evan Godt
Evan Godt, Writer

Evan joined TriMed in 2011, writing primarily for Health Imaging. Prior to diving into medical journalism, Evan worked for the Nine Network of Public Media in St. Louis. He also has worked in public relations and education. Evan studied journalism at the University of Missouri, with an emphasis on broadcast media.

Trimed Popup
Trimed Popup