Drug allergy labels in medical records power searches for gene-drug associationsJun. 17, 2021, 10:16 AM
by Paul Govern
A meta-analysis published in 1998 in the Journal of the American Medical Association estimated that, leaving aside overdoses and cases of drug abuse, 2.2 million patients admitted to U.S. hospitals in 1994 (6.7% of the total) were either admitted due to a serious adverse drug reaction or had a serious ADR while in the hospital. ADRs were estimated to be the fifth leading cause of death among hospitalized patients.
When a drug or combination of drugs causes different responses in different people, genetic variation may be at play. Post-market drug safety surveillance aims to identify these variable responses, and pharmacogenomics, through discovery of genetic risk and the application of clinical genotyping, aims to reduce trial and error in drug prescribing.
Genome-wide association studies (GWAS) scan for telltale associations between living traits and bits of DNA called single-nucleotide polymorphisms, or SNPs. Vanderbilt’s BioVU is a DNA biobank linked to de-identified electronic health records (EHRs).
With a resource such as BioVU, a retrospective scan for SNPs that confer risk of an ADR involves a search for exposures to the drug of interest and natural language processing (NLP) of clinical notes to sort out normal and abnormal drug responses. Validation of any resulting SNP drug associations means re-running the scan in another genotyped EHR population.
Developing and validating a bespoke NLP algorithm for a given ADR takes work. A hitch in this scenario is that clinical documentation habits tend to vary somewhat from one institution to the next, limiting the pace of discovery: an algorithm calibrated to identify ADRs at one institution isn’t apt to work at a second institution.
Teams are working to develop portable all-purpose NLP for identifying drug responses wherever they occur. Meanwhile, in PLOS Genetics, Neil Zheng, Wei-Qi Wei, MD, PhD, and colleagues report the utility of foregoing NLP and instead scanning for SNP-drug associations based simply on occurrences of drug names in the allergy section of the EHR.
“Ours is an alternative, high-throughput approach that offers the ability to scan for risk variants for multiple drugs essentially in a single go, with very few preliminaries,” said Wei, assistant professor of Biomedical Informatics at Vanderbilt University Medical Center. “This approach has the added benefit of being fully portable to other institutions.”
In this proof-of-concept study, the researchers use EHRs and genotype data of 81,739 individuals in BioVU, including 67,323 of European ancestry and 14,416 of African ancestry. The study takes in 14 common drugs and drug classes. ADR cases are defined in the study as any occurrence of the drug name in the allergy section of the EHR (including known abbreviations and misspellings); controls are defined as the presence of the label “no known drug allergy” in the allergy section, or, in the case of individuals with a documented drug exposure, absence of the drug name in the allergy section.
In the European American cohort, the scan found SNPs associated with seven of the 14 drugs and drug classes.
“Several of these SNPs turn out to be located near genes already implicated in drug responses, and several have documented association with expression levels for these genes,” Wei said.
“While allergy labels in the EHR may in many cases not hold up to clinical scrutiny, our results nevertheless show that this high-throughput method, in large part because it’s so readily scalable and portable, has clear potential to aid the search for ADR risk variants.”
The study’s African American cohort was itself too small to include in the primary scan and none of the SNP-drug associations found in the European American cohort replicated in the African American cohort.
Zheng, the report’s lead author, is a medical student at Yale University and a former staffer in Wei’s lab. Others from VUMC on the study include Cosby Stone, MD, MPH, Lan Jiang, MS, Eric Kerchberger, MD, Cecilia Chung, MD, MPH, QiPing Feng, PhD, Nancy Cox, PhD, Michael Stein, MD, Dan Roden, MD, Joshua Denny, MD, MS, and Elizabeth Phillips, MD.