NLP-Enabled Review Accurate, Less Costly Than Human Review
Researchers at the Regenstrief Institute in Indianapolis say they have created a system for tracking the quality of colonoscopies and determining the appropriate intervals between these procedures based on natural language processing that is as accurate but less expensive than human review.
Researchers at the Regenstrief Institute in Indianapolis say they have created a system for tracking the quality of colonoscopies and determining the appropriate intervals between these procedures based on natural language processing that is as accurate but less expensive than human review.
The study found that in spite of linguistic variation in the way providers at 13 medical centers across the country used words to express the same concept, 95 percent of the time computerized natural language processing was correct regarding type and location of an adenomatous polyp, compared to human expert reviewers who were assumed to be 100 percent correct. As a result, Regenstrief officials said the evidence suggests software can perform the same job in interpreting and correlating colonoscopy and pathology reports as humans, only much faster, reproducibly, and far less expensively.
"We found that rapid and inexpensive natural language processing, which utilizes free-text data that was previously unusable for efficient computer-based analysis, was extremely accurate in measuring adenoma detection rate during colonoscopy," said Timothy Imler, M.D., a Regenstrief Institute investigator and Indiana University School of Medicine assistant professor of medicine, who led the study. "The presence of adenomas in the colon is predictive of a patients risk of later developing colon cancer, and the detection rate has been identified as the critical measure of a high-quality endoscopist, the specialist who performs colonoscopy."
Of more than 90,000 colonoscopies performed at the Albany, N.Y.; Ann Arbor, Mich.; Bronx, Denver, Detroit, Houston, Indianapolis, Nashville, New York Harbor; Northport, N.Y.; San Francisco, West Haven, Conn., and White River Junction, Vt. Veterans Affairs medical centers, slightly fewer than half had findings that generated pathology reports.
Seven hundred and fifty paired colonoscopy and pathology reports from this group were randomly sampled and compared by human reviewers using 19 measurements related to quality and procedure interval determination. The remaining 41,819 paired reports were processed through natural language processing to access performance consistency using the same 19 measurements.
The study, published online in the American Journal of Gastroenterology, is available here.
The study found that in spite of linguistic variation in the way providers at 13 medical centers across the country used words to express the same concept, 95 percent of the time computerized natural language processing was correct regarding type and location of an adenomatous polyp, compared to human expert reviewers who were assumed to be 100 percent correct. As a result, Regenstrief officials said the evidence suggests software can perform the same job in interpreting and correlating colonoscopy and pathology reports as humans, only much faster, reproducibly, and far less expensively.
"We found that rapid and inexpensive natural language processing, which utilizes free-text data that was previously unusable for efficient computer-based analysis, was extremely accurate in measuring adenoma detection rate during colonoscopy," said Timothy Imler, M.D., a Regenstrief Institute investigator and Indiana University School of Medicine assistant professor of medicine, who led the study. "The presence of adenomas in the colon is predictive of a patients risk of later developing colon cancer, and the detection rate has been identified as the critical measure of a high-quality endoscopist, the specialist who performs colonoscopy."
Of more than 90,000 colonoscopies performed at the Albany, N.Y.; Ann Arbor, Mich.; Bronx, Denver, Detroit, Houston, Indianapolis, Nashville, New York Harbor; Northport, N.Y.; San Francisco, West Haven, Conn., and White River Junction, Vt. Veterans Affairs medical centers, slightly fewer than half had findings that generated pathology reports.
Seven hundred and fifty paired colonoscopy and pathology reports from this group were randomly sampled and compared by human reviewers using 19 measurements related to quality and procedure interval determination. The remaining 41,819 paired reports were processed through natural language processing to access performance consistency using the same 19 measurements.
The study, published online in the American Journal of Gastroenterology, is available here.
More for you
Loading data for hdm_tax_topic #better-outcomes...