Data captured in electronic medical records (EMR) for routine clinical care as currently implemented is sufficient to conduct genome-wide association studies, according to the first study to look at EMR-based research.The study published today in Science Translational Medicineprovides the first proof of the concept that one of themeaningful uses of EMRs is that they might be used to conduct a variety of health research studies more efficiently and cost-effectively than is currently possible.
Led by Dr. Abel Kho, assistant professor of medicine at Northwestern University’s Feinberg School of Medicine, the consortium of researchers, used the EMRs at five institutions to identify cases and controls for four types of disease and one biologic measure of cardiac disease.
“The goal is to try to use data contained in a patient’s medical record and repurpose that information to identify whether that patient has a disease or not,” Kho told The Hub in a phone interview. “This is important because no one knew if EMRs could be used in this way. The concept is great. With all this medical data available, would we be able to use it to do research?
“Many people thought there would be too many holes in the records to be able to accurately identify enough information to be useful for research purposes. This study shows that, yes EMRs are good enough,” he added.
The study was conducted by the eMERGE Network, a national consortium formed to develop, disseminate, and apply approaches to research that combine biological repositories of DNA with electronic medical record (EMR) systems for large-scale, genome-wide association studies.
The five institutions included Group Health, Seattle; Marshfield Clinic Research Foundation, Marshfield, WI; Mayo Clinic, Rochester, MN; Northwestern University, Chicago; and Vanderbilt University, Nashville. All five had different comprehensive EMR/EHR systems using either structured data storage in predefined formats, or free-text records with natural language processing, or a mix of both. All five also had access to biological repositories of tissue or blood samples that could be linked to the EMRs.
Each institution used their EMRs and biological repositories to extract data to identify cases and controls for a specific disease, or in Vanderbilt’s case a biological measure of cardiac conduction. Group Health was tasked with mining data for dementia, including Alzheimer’s disease. Marshfield looked at cataracts, Mayo Clinic did peripheral artery disease, and Northwestern mined its EMR for Type 2 diabetes data. All sites obtained permission from patients to use both their biological samples and their medical records for this study.
To identify the diseases, Kho and colleagues searched the records for a series of criteria such as diagnosis, medications, and laboratory tests, using a complex set of algorithms to extract the data from both structured data fields and free-text data.
To determine the accuracy of defining the disease phenotypes using EMR data alone, each site reviewed 100 clinical charts from the EMR. Three sites used clinician chart review as the standard to confirm the primary phenotype from the records. One site used the clinical gold standard for their primary phenotype. The remaining site used trained EMR chart abstractors to confirm the primary phenotype.
What they found was that despite variations in categories and completeness of data capture across sites, four of the five study sites achieved positive predictive values of close to 100 percent for use of EMR data alone to identify their primary disease phenotype. The Group Health site achieved a lower positive predictive value of 73 percent using EMR data to identify cases with dementia.
Dr. Eric Larson, principal investigator for the Group Health portion of the study said that one of the reasons dementia was included in this test of EMR-based research is because it is such a difficult diagnosis.
“A Hawaiian study, similar to the Group Health study, several years ago showed that in a group of dementia patients, only 20 percent to 30 percent were known by their physicians to have dementia,” Larson said by phone. “Another 20 percent to 30 percent were known by family members to have symptoms of the disease, and the remaining 40 percent were unrecognized by either their physicians or family members.”
Larson says it was surprising that they were able to identify most of the dementia patients in the EMRbased on four or five mentions of and or use of a particular medication. He noted that the complexity of diagnosing dementia involves questionnaires and memory tests that don’t easily fit into a patient’s EMR.
In addition to looking at how accurately the EMR databases could be used to describe patients with or without a particular disease or condition, the researchers also assessed the effect on research that variations in database design would have doing this type of research.
To assess the additional benefit of natural language processing, they compared the number of cases identified using structured data alone to that using both structured data and natural language processing at the Vanderbilt site. They found that using natural language processing identified 2,950 cases of the heart conduction measure, QRS duration, compared to 1,288 cases identified using only structured data and search string matching, a 129 percent increase using natural language processing.
They also found some “holes” that were common to all the EMRs. While criteria such as diagnosis codes, medications, and laboratory tests were readily extracted to identify disease types for genome wide association studies, other criteria, such as race/ethnicity, family history, smoking, and environmental exposures were often not uniformly documented across all EMRs. When these were present, it was often captured in free-text form in clinicians’ notes, for example, and without consistent or standard nomenclature.
The researchers concluded that widespread adoption of EMRs creates the potential for a quantum shift forward in the availability of longitudinal, real-world clinical data for genetics research. Improvements in standardization and interoperability will further improve the ability to use EMRs for genome wide association studies.
“This is the first study done on a broad scale to show that yes EMRs could be used this way for research purposes, and yes some of what we found will be used to help institutions to develop more standardized EMR/EHR implementation,” Kho said.
“It is hoped that being able to more efficiently identify people with and without a specific disease, we will be able to more efficiently recruit and conduct clinical trials.”
By Michael O’Leary, contributing writing, Health Imaging Hub
No comments:
Post a Comment