Pharmacovigilance using clinical notes
Physicians prescribe medications based on their diagnosis of a patient’s condition to help them get better,eventually recover from their condition if not chronic or help them manage the condition if it is chronic. According to the article, 50%  of adverse events during hospital stays are a result of drug related events. These events increase a patient’s time in the hospital, which costs both the patient, the hospital and sometime clinician who are responsible for their care, dearly. Some of these safety issues aren’t usually detected until after market approval when they’ve been in use by the general public. Coded discharge and insurance claims data have been used to try to detect drug safety issue, but experts believe more than 90% of the data needed to detect safety issues could be missing from the data . The writers of this article propose an approach that uses free text clinical notes to detect safety issues involving drugs. Their method uses annotation and medical terminologies to transform free text clinical notes into a de-identified patient feature matrix. The generated matrices serve as input for a high-throughput process that detects drug–adverse event associations and adverse events associated with drug–drug interactions, which in most cases is before official alerts are issued.
They pulled data from Stanford Translational Research Integrated Database which contains 1.8 million patients with 19 million documented encounters and more than 11 million unstructured clinical notes. A references standard was created using known drug-adverse event associations consisting of two sets (single-drug adverse events and two-drug case), to test the performance of their methods .
They used a two step approach to test their methods. The first step helped them flag putative signals while the second step helped them flag for false positives. For the first step they computed a raw association in the form of an unadjusted OR, followed by adjustment for potential confounders, using the patient-feature matrix. In the second step they adjusted for confounding using specific patient factors like; age,gender, ethnicity, comorbidity and coprescription frequency to calculate the propensity score .The propensity score indicates the probability of a patient being exposed to a drug.
They were able to detect drug-adverse event associations by reproducing the association between rofecoxib and myocardial infarction . In reproducing the association, they obtained an odds ratio of 1.31 (95% confidence interval (CI): 1.16–1.45) which correlates with previously reported associations.They were also able to detect adverse drug-drug interactions with an AUC performance of 81.5%
Applying data mining techniques to clinical notes for the purpose of pharmacovigilance is not only feasible but a smarter and faster way to detect adverse events related to drugs.
Write something here!
- Pharmacovigilance using clinical notes: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2815419/