Algorithm Mines FDA Reports for Drug Interactions

Russ Altman (Stanford University)

Researchers at Stanford University’s medical school and bioengineering program have devised a computer algorithm that can query millions of adverse drug reports to the U.S. Food and Drug Administration (FDA) by patients and their physicians, and identify many more drug interactions and side effects than were previously known. Their work is described in the journal Science Translational Medicine released today (paid subscription required).

Clinical trials are often the usual means of identifying potential adverse effects, when testing drugs on humans, but the samples of patients in these studies is often small, which means many side effects don’t become known until after regulatory authorities approve the drug. FDA, as part of its post-marketing surveillance, collects patient or physician reports of drug side effects in its Adverse Event Reporting System.

However, factors, such as other medications being taken, patient demographics (gender, age, environment), medical histories, and reasons for prescribing a drug can complicate drawing conclusions from the reports to FDA. As a result, it is often difficult to tell if a reported adverse reaction is due solely to the drug in question or a collection of factors interacting with the drug.

Bioengineering grad student and lead author Nicholas Tatonetti developed a mathematical method for querying the FDA database that matched up as many of these confounding factors as possible to isolate the potential impact of a drug. If significantly more of the people on the drug reported an adverse event, such as headaches or vomiting for example, than did those who were not taking the drug, it is likely that the medication was indeed the culprit. A similar method can be used to analyze the interacting effects of pairs of drugs.

The algorithm helped identify known effects and interactions, but also factors they had not anticipated. “[W]e found that the more things you can match between the groups, like other drugs the people have in common,” says Tatonetti, “the more likely you are to also unintentionally match for variables you may not have even thought about but that may affect the result.”

Tatonetti and senior author Russ Altman (pictured at top), a professor in Stanford’s medical school and bioengineering program, then tested the model on electronic health records of patients at Stanford’s hospital and clinics. They confirmed that 47 new drug interactions out of 395 predicted drug-drug interactions identified in the queries of the FDA database held true, when analyzing the records of real patients.

In particular, patients receiving both an anti-depressant known as selective serotonin reuptake inhibitors (SSRIs) and a class of blood pressure medication called thiazides were more likely (by 9.3%) to exhibit prolonged QT intervals — time between waves in the heartbeat cycle — on an electrocardiogram than patients taking either medication alone (4.8% vs. 6.5%, respectively). Prolonged QT intervals are associated with increased incidences of spontaneous arrhythmias and sudden cardiac death.

The researchers have collected the results of these and subsequent queries in two public databases, named Offsides and Twosides. The Offsides database lists adverse events for the 1,332 drugs in the system. Each of the drugs in Offsides lists on average 329 adverse effects, as opposed to an average of 69 effects given on manufacturers’ package inserts. The Twosides database identifies 1,301 adverse events, resulting from an analysis of 59,220 pairs of drugs that cannot be clearly assigned to either drug alone.

Altman attributes the results of the research in part to the ability to process large data sets. “When you start with millions of pieces of information, you can be pretty rigorous about weeding out those that don’t match,” says Altman. “And if you can arrive at even just a few hundred well-matched cases, that can give a good statistical comparison.”

Read more: FDA Adverse Drug Event Reports Jump in Past Decade

*     *     *