Researchers at Vanderbilt University Medical Center have developed a new way to identify dangerous drug interactions that may have been missed during clinical trials, using artificial intelligence to analyze medical literature and verify findings in patient records.
The study, reported in Clinical Pharmacology & Therapeutics, confirmed nine severe adverse drug interactions that, while previously reported in scientific journals, had never been logged in DrugBank, a database of more than 1.3 million drug-drug interactions used by health care providers to check for medication risks. These include the finding that combining methadone with the antibiotic ciprofloxacin may cause dangerous respiratory depression, and that taking the blood thinner ibrutinib with the heart drug amiodarone may increase the risk of atrial fibrillation.
“Many serious drug reactions don’t become apparent during pre-market testing because clinical trials typically involve only a few thousand patients and exclude more vulnerable population groups,” said the paper’s senior and corresponding author, You Chen, PhD, associate professor of Biomedical Informatics at VUMC. “Our approach allows us to detect these problems by analyzing millions of real-world patient records.”
Currently, after-market drug safety monitoring relies heavily on pharmaceutical manufacturers, health care providers and patients reporting suspected adverse reactions to regulatory authorities. This system, while valuable, depends on observers recognizing and reporting problems, which can miss rare complications.
“Think of drug interactions like traffic accidents: even if each car is safe on its own, putting certain cars together on the same road can lead to dangerous crashes that weren’t seen during test drives,” Chen said. “Our approach is like a new traffic safety system — it uses artificial intelligence tools to scan medical literature and patient records for these ‘accidents’ that weren’t caught in earlier tests. By doing this, we can spot and confirm unrecognized drug safety problems more quickly and accurately, helping to protect patients before serious harm occurs.”
The research team used natural language processing — a type of artificial intelligence — to sift through 160,321 scientific articles, published between 1962 and 2023, identifying potential adverse reactions when certain drugs are taken together. Arriving at 111 drug-drug interactions tied to severe adverse drug reactions, the team verified these findings using medical records from over 3.4 million patients at VUMC and the National Institutes of Health’s All of Us research program.
“What makes our method powerful is the validation step,” said lead author Eugene Jeong, a doctoral student in the Department of Biomedical Informatics. “When we find the same drug interaction problems occurring in large patient databases, it gives us confidence that these are real effects that doctors need to know about.”
The study focused on interactions involving drugs that are processed by any of five liver enzymes, amounting to approximately 80% of all pharmaceuticals. When two such drugs are taken together, one of the drugs may slow the breakdown of the other, raising its concentration in the body and potentially increasing the risk of dangerous side effects.
Other confirmed risks include hallucinations when combining the pain medication tramadol with the antifungal drug fluconazole, and kidney damage when taking the antibiotic clarithromycin with the antifungal medication voriconazole.
The researchers found that patients taking these drug combinations during the same period had over 90% greater likelihood of experiencing these dangerous side effects compared to those taking a single drug alone.
Chen noted that the findings could help improve patient safety. “Doctors often have to prescribe multiple medications, especially for older patients or those with complex conditions. Better understanding of drug interactions helps them make safer choices.”
Jeong and Chen were joined by two researchers from The Ohio State University in Columbus. The study was supported by the National Institutes of Health (grants R01LM014199, UL1TR002243).
