Life sciences drug-drug interactions

Learnings from the FDA: use of text mining to review drug-drug interactions

January 24 2018

Understanding drug-drug interactions can improve drug safety

A considerable proportion of adverse drug events are caused by interactions between drugs. With an ageing population, and associated increasing multiplicity of age-related illnesses, there is an increase in the potential for increased risk of drug-drug interactions (DDIs). One way of alleviating some DDIs is by ensuring that potentially interacting drugs are taken at suitable time intervals apart. But, what is the best interval to recommend?

In a recent seminar, Keith Burkhardt of the FDA described a project using text mining to survey the landscape of information on DDIs from FDA Drug Labels. And, in particular, the FDA review division wanted to find labelling for drugs where the time separation was stated, in order to prevent potential drug safety events.

Mining Data from FDA Drug Labels: dosing regimens and time separation

The drug classes of interest included bile acid sequestrants and exchange resins (such as cholestyramine, colestipol, colesevelam, all LDL cholesterol lowering drugs), phosphate binders (e.g. sevelamer; used for patients with chronic kidney failure), and chelators (used to treat excessively high levels of lead, iron or copper in the blood; e.g. deferasirox, deferiprone). These drug classes can all alter the bioavailability of other drugs, particularly for those with a narrow therapeutic range such as warfarin or antiepileptic drugs.

Using I2E, the FDA team were able to search the DailyMed drug labels for the four classes of drugs. Using the text mining approach, the appropriate section in the labels could be searched for any mention of terms such as concomitant, concurrent, or co-administration, and other lexical forms of these.

Text mining results in quicker and more comprehensive analysis

The FDA team found more co-administration labels with specific language relating to the dose-timing than using manual review. Text mining also enabled the extraction of the timing description into a structured table, that allowed the researchers to analyze the variability in the label language. They found that most labels recommended 1-2 hrs before and 4-6 hours after, but some variability was noted even within class. Using text mining, this analysis was rapid, performed in a single day.

Understanding the optimal time separation for drug regimens is important to avoid potential DDIs and hence improve drug safety. This project shows how unstructured content, such as the detailed DailyMed Drug labels, can be mined to extract key information to help inform clinical and prescribing decisions. 


In the same talk, Keith described a project to accurately capture adverse event information from FDA labels and translate to a standard medical vocabulary (MedDRA Preferred Terms). They compared the I2E output to a set of manually curated labels and found I2E gave results with excellent metrics, including an F-score of 0.95.  (Other metrics:  Precision: 0.92, Recall: 0.96, Specificity 0.94, Accuracy 0.94).