Analysis of EHR pathology reports aids understanding of skin biopsies
UW Medicine leverages NLP software to provide new data on the frequency of melanocytic lesions.
Researchers from the University of Washington School of Medicine have successfully used natural language processing software to analyze electronic health record pathology reports of patients who underwent skin biopsies, generating population-level estimates of diagnoses.
Investigators leveraged internally-developed NLP software for the analysis of more than 80,000 pathology reports in order to identify biopsied lesions that were melanocytic and likely to develop into malignant melanoma.
According to Joann Elmore, MD, a faculty physician in the Department of Medicine and an epidemiologist with the UW School of Public Health, it’s very difficult for doctors to interpret melanocytic lesions—which is only made more difficult by the fact that pathologists use inconsistent wording in differentiating between benign and malignant diagnoses.
However, Elmore and Michael Piepkorn, MD, clinical professor in the Department of Dermatology, have found that NLP software accurately gathers details from EHRs as a text-analysis tool and in a manner that is just as proficient—if not better than—human annotators reviewing records manually.
“There’s a wealth of information in electronic medical records that has been untapped for population-level studies like this,” said Elmore. “Text-analysis software is especially helpful to decipher pathologists’ reports, because they have a wide array of terms for the same type of lesion.”
“It would be nice if everybody categorized all of their pathology reports using the same words,” she adds. “Right now, people are using 50 to 60 different terms. Everybody has their own special way they like to describe lesions. We used this NLP to categorize all of the many different terms that lots of different doctors are using when they’re interpreting the same kind of lesions.”
Also See: Deep learning excels at analysis of breast cancer biopsy slides
Their multi-site study, which included researchers from Connecticut, New Hampshire, Pennsylvania, Rhode Island, and Paris, France, was published earlier this month in JAMA Dermatology showing that melanocytic lesions accounted for 23 percent of all diagnoses while 77 percent of biopsies were diagnosed as non-melanocytic.
Elmore says she was surprised at the finding, noting that 5 or 10 percent melanocytic diagnoses were her expectation going into the study.
“Electronic medical records enable automated extraction of pathology report data to improve our epidemiologic understanding of skin biopsy outcomes, specifically those of melanocytic origin,” concludes the study. “These data provide the first population-based estimates across the spectrum of melanocytic lesions ranging from benign through dysplastic to malignant. These results may serve as a foundation for future research seeking to understand the epidemiology of melanocytic proliferations and optimization of skin biopsy utilization.”
In addition, Elmore contends that the NLP software could also be applied to radiology reports and any number of clinical situations which require sorting and categorizing the data into discrete categories.
Investigators leveraged internally-developed NLP software for the analysis of more than 80,000 pathology reports in order to identify biopsied lesions that were melanocytic and likely to develop into malignant melanoma.
According to Joann Elmore, MD, a faculty physician in the Department of Medicine and an epidemiologist with the UW School of Public Health, it’s very difficult for doctors to interpret melanocytic lesions—which is only made more difficult by the fact that pathologists use inconsistent wording in differentiating between benign and malignant diagnoses.
However, Elmore and Michael Piepkorn, MD, clinical professor in the Department of Dermatology, have found that NLP software accurately gathers details from EHRs as a text-analysis tool and in a manner that is just as proficient—if not better than—human annotators reviewing records manually.
“There’s a wealth of information in electronic medical records that has been untapped for population-level studies like this,” said Elmore. “Text-analysis software is especially helpful to decipher pathologists’ reports, because they have a wide array of terms for the same type of lesion.”
“It would be nice if everybody categorized all of their pathology reports using the same words,” she adds. “Right now, people are using 50 to 60 different terms. Everybody has their own special way they like to describe lesions. We used this NLP to categorize all of the many different terms that lots of different doctors are using when they’re interpreting the same kind of lesions.”
Also See: Deep learning excels at analysis of breast cancer biopsy slidesTheir multi-site study, which included researchers from Connecticut, New Hampshire, Pennsylvania, Rhode Island, and Paris, France, was published earlier this month in JAMA Dermatology showing that melanocytic lesions accounted for 23 percent of all diagnoses while 77 percent of biopsies were diagnosed as non-melanocytic.
Elmore says she was surprised at the finding, noting that 5 or 10 percent melanocytic diagnoses were her expectation going into the study.
“Electronic medical records enable automated extraction of pathology report data to improve our epidemiologic understanding of skin biopsy outcomes, specifically those of melanocytic origin,” concludes the study. “These data provide the first population-based estimates across the spectrum of melanocytic lesions ranging from benign through dysplastic to malignant. These results may serve as a foundation for future research seeking to understand the epidemiology of melanocytic proliferations and optimization of skin biopsy utilization.”
In addition, Elmore contends that the NLP software could also be applied to radiology reports and any number of clinical situations which require sorting and categorizing the data into discrete categories.
More for you
Loading data for hdm_tax_topic #better-outcomes...