Medical students are using a popular research tool to pump out misleading studies

Every morning, Joshua Wang sits down at his computer with a pastry and a can of cold, black coffee to look for the latest papers based on data from a popular research platform called TriNetX. Studies based on the platform—which provides access to anonymized electronic health records for more than 300 million patients in the United States and abroad—have skyrocketed in recent years. Wang, a neuroscientist at Taipei Tzu Chi Hospital who trains researchers there to use TriNetX, has noticed another trend, too. Some results, he says, look “a bit dodgy.” He and others say the easy-to-use platform may be allowing inexperienced researchers—potentially aided by artificial intelligence (AI)—to churn out unreliable and bias-ridden studies with unrivaled speed. “We’ve seen a lot of these TriNetX studies, and they all seem to have very similar flaws,” says Samy Suissa, a pharmacoepidemiologist at McGill University. “They seem to always find these spectacular effects, remarkable benefits for drugs on all kinds of outcomes.” In 2025, nearly 2700 publications mentioned TriNetX in the title or abstract, up from just 33 only 5 years prior, according to the Dimensions database, which tracks abstracts and citations. Less than halfway through this year, the number already exceeds 2100. The rise mirrors recently reported trends seen in papers using publicly available health data sets, which are authored primarily by researchers in China. But TriNetX is only open to users at participating health care organizations, and most TriNetX papers come from authors at U.S. medical schools, often with a physician-in-training as lead author. Medical schools use TriNetX as a research training ground, and the resulting papers are a relatively easy way for medical students to boost their CVs before applying for residencies. “There is no substitute for learning this process than by doing it,” says Lisa Howley of the Association of American Medical Colleges (AAMC). But the combination of inexperienced users and TriNetX’s push-button analysis tools can lead to shoddy publications, which often do not correct for potential biases that can make treatments appear more effective than they are. And because the data can be analyzed so quickly, users can easily cherry-pick positive results for publication, a practice known as p-hacking. “The flow of false discoveries is hugely greater,” says Matt Spick, a health-data scientist at the University of Surrey. “My biggest concern,” Wang says, “is that doctors in 10 years’ time want to look into a particular concept and they go into the literature and everything is just associated with everything.” TriNetX Chief Scientific Officer Jeffrey Brown agrees users need epidemiologic and statistical expertise and that papers should undergo robust peer review. But, he adds, “There’s more research happening, and I think that that’s good.” Wang and many other researchers disagree. As one example, they point to a TriNetX paper published in the MDPI journal Cancers that made the news for finding what the authors described as “compelling evidence” that popular GLP-1 weight-loss drugs lower the risk for a long list of cancers in obese people. The paper failed to mention, let alone correct for, two key biases that can skew results in favor of the treatment being studied, called collider bias and immortal-time bias. Collider bias can arise when both an exposure—for example, to a weight-loss drug—and an outcome such as cancer drive health care use, the so-called collider. The bias can create a spurious negative correlation between the exposure and outcome. Immortal-time bias can occur when researchers compare outcomes between patients who receive a certain treatment after a health event—say, a heart attack—and those who do not get the treatment, because any patient who dies before treatment automatically becomes part of the untreated group. That group then appears to have higher mortality. It’s “just an awful paper,” Suissa says. Spick notes the drug being “miraculously protective” across several unrelated organ systems is “implausible,” given that “cancers are wildly different and have wildly different causes.” Neither of the paper’s two corresponding authors—one of whom had undisclosed ties to a manufacturer of weight-loss drugs—responded to emailed questions from Science. In other cases, papers claim to have used TriNetX to do analyses the platform doesn’t in fact offer. Wang came across a paper published last year in Angiology suggesting diabetes drugs called gliflozins could reduce the risk of death after a heart attack. The authors, physicians at three top-tier U.S. medical schools, wrote they had conducted a key step to correct for immortal-time bias within TriNetX. Wang knew TriNetX offers no such tool. “It really got me going,” says Wang, who has written dozens of letters to the editor pointing out problematic methods in published research using the software. “Either they have falsified their methods or they have uncritically copied a method sentence from a different article or from an AI output. … I think both are pretty scary.” In response to questions from Science, the study’s first author, Rochell Issa, a final-year internal medicine resident at the Cleveland Clinic, defended the work, reiterating the methods as written in the paper. She stopped responding to follow-up questions about exactly how they had executed the analysis on the platform. Another author of the Angiology paper, David Kaelber, chief health informatics officer at the MetroHealth System in Cleveland and associate professor at Case Western Reserve University, denied using AI “to generate the methods or approach used in this or any of our studies.” But Wang and colleagues asked seven large language models (LLMs) how to use TriNetX to complete a key step in correcting for immortal-time bias. Six suggested methods that were impossible to implement on the platform, they reported in the European Journal of Epidemiology. The researchers then searched TriNetX papers for the impossible approaches suggested by the LLMs and found eight papers, including the Angiology study. In five cases, medical students or residents in the U.S. appeared on the papers’ author lists, typically as first authors. Wang has since found five more papers suffering from the same problem. TriNetX argued in a published response to Wang’s findings that the eight studies represent “a tiny fraction” of the work done with its software. In addition, the TriNetX authors write, descriptions of methods that are impossible to execute “can plausibly arise from misunderstanding, ambiguous terminology, incomplete reporting, or analysis performed outside the platform.” Kaelber, who has amassed 125 TriNetX publications, according to the Dimensions database—more than anyone else in the world—told Science that concerns about low-quality, unnecessary, and bias-ridden studies done on the platform are “totally valid.” One key “is transparency as to all of the design and TriNetX platform configuration decisions.” But neither he nor any other author contacted for this story agreed to share their TriNetX query parameters. The problematic studies can influence patient care, says Brian VanderBeek, an ophthalmologist at the University of Pennsylvania who recently highlighted potential biases in a pair of TriNetX studies that suggested the food supplements turmeric and melatonin could drastically cut the risk of serious eye disease. “There may be some danger in that a physician could be falsely led to believe that there’s a protective effect,” VanderBeek says. For its part, AAMC, which governs the residency application process in the U.S., is trying to address the problem of quick-and-dirty papers, Howley says. For the coming cycle, it will ask applicants to shift the focus of their publications list “from quantity to quality, emphasizing meaningful contributions, depth of involvement, and the impact of applicants’ work.” Meanwhile, Wang continues his daily vigil, and he is working to promote best practices. At his own hospital, researchers seeking access to TriNetX must first complete a 1-hour training session with him. A lot of that time, he says, is spent illustrating how easy it is to get “beautiful-looking” but meaningless results. The hope, he says, is to “try and instill a little bit of fear so that they don’t run off and churn it out.”