Dr. Paul Chang from the University of Chicago had challenged the Philips team to develop methods that automatically prove the downstream effect of quality radiology reports, said co-author Merlijn Sevenster, PhD, of Philips.
"Such methods will empower the radiology department with quantitative proof of its contribution to the care chain, and will help differentiate its value-add in the healthcare enterprise," Sevenster said. "With the accountable care organization [model], such analytics methods will be increasingly important."
The researchers developed natural language processing technology to extract keywords from narrative radiology reports that indicate certainty, such as "diagnostic of" and "may represent." Using a numeric scale derived from previous research, they were then able to calculate an aggregated report-specific certainty score, Sevenster said.
In testing on a set of 17,000 breast reports from the University of Chicago, the group found a strong correlation between the certainty scores of the reports and the certainty level conveyed by the BI-RADS score on breast exams (i.e., BI-RADS 1, 2, and 5: high certainty; BI-RADS 2 and 4: low certainty).
"This indicates the natural language processing technology is able to pick up the certainty 'mood' of narrative reports," Sevenster told AuntMinnie.com. "In future research, we hope to find correlations between certainty scores of general radiology reports and downstream impact parameters, such as follow-up imaging, biopsies, and lab tests."
