A platform designed to accelerate clinical research has become a tool for producing what some epidemiologists call an alarming volume of misleading studies. An investigation published June 24 in Science by Frederik Joelving of Retraction Watch documents how medical students and resident physicians are using TriNetX, a platform providing anonymized electronic health records from over 300 million patients, to rapidly generate publications, often with serious methodological flaws.
How it works
TriNetX allows users to query a vast database of de-identified patient records with near-instant results. A researcher can define a cohort, run an analysis, and have publishable-looking output in minutes. The platform makes it trivially easy to run multiple analyses and report only those that produce statistically significant or clinically “spectacular” results, classic p-hacking and cherry-picking.
The incentive structure is clear: medical students and residents need publications to strengthen their residency and fellowship applications. The Association of American Medical Colleges has recognized the problem, starting next application cycle, the focus will shift “from quantity to quality, emphasizing meaningful contributions, depth of involvement, and the impact of applicants’ work”, but the change has not yet taken effect.
The methodological problems
Epidemiologist Samy Suissa of McGill University, who has reviewed many TriNetX studies, describes them as having “very similar flaws” that “always seem to find these spectacular effects, remarkable benefits for drugs on all kinds of outcomes.” Two major categories of bias are common:
Collider bias and immortal-time bias: Patients who die before receiving treatment are automatically assigned to the untreated group, making the treated group appear healthier. Collider bias creates spurious correlations that standard statistical adjustments do not correct. Suissa called a Cancers (MDPI) paper claiming GLP-1 drugs reduce the risk of many cancers “an awful paper”, an example of these biases producing entirely unreliable conclusions.
Fabricated or impossible methods: Researcher Joshua Wang and colleagues identified eight studies that claimed to use TriNetX features that do not exist. For example, an Angiology paper claimed the platform could correct for immortal-time bias, TriNetX offers no such tool. Five of the eight studies had medical students or residents as first authors. Wang later found five more such papers, bringing the total to at least 13. When Wang asked seven large language models to describe how to perform TriNetX analyses, six of them generated impossible methods, suggesting that some authors may have used AI to write their methods sections without verifying whether the described functionality exists.
The scale
Publications mentioning TriNetX in the title or abstract have skyrocketed: from approximately 33 in 2019 to roughly 2,700 in 2025. In the first half of 2026 alone, over 2,100 papers have already appeared. Most authors are at U.S. medical schools. David Kaelber of MetroHealth and Case Western Reserve University has 125 TriNetX publications, the most of any single researcher. Kaelber and others refused to share their TriNetX query parameters when asked, making replication or independent verification impossible.
The consequences
The practical concern is patient care. Ophthalmologist Brian VanderBeek, who reviewed two TriNetX studies claiming turmeric and melatonin dramatically reduce the risk of serious eye disease, described the results as “likely biased”, but a physician reading the abstract might not know that. False-positive results could lead to inappropriate prescribing, unnecessary worry, or missed diagnoses.
TriNetX chief scientific officer Jeffrey Brown defended the platform, saying the eight identified problematic studies represent “a tiny fraction” of TriNetX research and that impossible methodological descriptions could arise from “misunderstanding, ambiguous terminology, incomplete reporting, or analysis performed outside the platform.”
Wang’s assessment is more blunt: either the authors falsified their methods or uncritically copied from another article or AI output. “Both are pretty scary,” he said.
The AAMC’s shift toward evaluating quality over quantity in applications is one response. But the investigation raises a deeper question: when a platform makes it this easy to produce biased results, and when the incentive system rewards publication count over rigor, how many unreliable studies are entering the medical literature, and what will it take to pull them out?
Source: Joelving, F. Medical students are using a popular research tool to pump out misleading studies. Science (2026). Link