Hospital-acquired blood clots are one of medicine’s quiet killers — preventable in theory, missed in practice often enough that they remain a leading cause of avoidable harm in hospitalized patients. Vanderbilt University Medical Center is now running a randomized clinical trial to find out whether an AI model embedded directly into the electronic health record can change that, in what researchers describe as one of the first randomized trials to measure AI-driven clinical decision support against standard care across both medical and surgical units, and across both urban and rural hospitals.
Built from 132,000 patient encounters
The trial, called VTE-AI, is led by Colin G. Walsh, MD, of Vanderbilt’s Department of Biomedical Informatics, along with colleagues including Yufei Long, Laurie Lovett Novak, Megan E. Salwei, Benjamin Tillman, Benjamin French, Amanda S. Mixon, Michelle E. Law, Jacob Franklin, and Peter J. Embi. The underlying prediction model, also called VTE-AI, was built from 132,330 adult patient encounters recorded at Vanderbilt between 2018 and 2020. Researchers narrowed an initial list of 82 candidate risk factors — spanning demographics, vital signs, lab values, diagnoses, and procedures — down to 25 that best predicted which patients would develop a hospital-acquired venous thromboembolism, the clinical term for a blood clot that forms in a vein, often in the leg, that can break loose and travel to the lungs. In retrospective testing, the model achieved a C-statistic of 0.891, a measure of how well it distinguishes high-risk from low-risk patients, with a working risk threshold of 3.6 percent used to trigger an alert.
Testing a nudge, not a mandate
Since enrollment began on October 1, 2025, the trial has been randomizing eligible adult patients 1:1 inside the electronic health record at Vanderbilt’s main Nashville hospital and three affiliated rural hospitals in Middle Tennessee. Patients in the intervention arm trigger a “nudge” practice alert for their care team when the model flags elevated clot risk; patients in the control arm receive standard-of-care risk assessment without the AI prompt. The study is tracking roughly 2,236 encounters — about 1,118 per arm — to reach 80 percent statistical power, with primary completion expected by December 2026 and full trial completion by March 2027. The primary outcome is simple: does the rate of hospital-acquired blood clots actually go down. Secondary outcomes include bleeding complications, 30-day readmissions, and length of stay, since over-aggressive clot prevention can itself cause harm.
Why this kind of trial is unusually rare
What makes VTE-AI notable is less the algorithm than the method. Hospitals have deployed AI-based risk scores for years, but head-to-head randomized evidence that a given AI alert changes patient outcomes — rather than just flagging risk — is still scarce. The researchers note that behavioral “nudge” interventions have been studied before in cardiology and oncology settings, but not previously in hematology or in the kind of routine, daily prophylaxis decisions this trial targets. That gap matters because clinical decision support tools have a long, mixed history: plenty of hospital alert systems have been built and deployed based on retrospective accuracy alone, with much weaker evidence that they changed what clinicians actually did at the bedside.
The researchers’ own worries
The Vanderbilt team has been unusually candid in its published trial protocol about what could go wrong. They flag “alert fatigue” directly, noting that clinicians bombarded with electronic alerts tend to start ignoring them — a well-documented failure mode in hospital IT — and have designed the nudge to fire only starting on a patient’s second hospital day, when initial admission workflows are less chaotic. They also warn of contamination risk: physicians who see the AI alert on one patient might unconsciously apply the same vigilance to a similar patient in the control arm, which would understate the alert’s true effect. And they acknowledge that prediction models drift over time as patient populations and hospital practices change — a drift this trial is not designed to detect or correct, since it is evaluating the model as currently built, not maintaining it indefinitely. The study also spans only one region of one state, a limitation the authors openly concede when it comes to generalizing results nationally.
What it means for the field
If VTE-AI shows a meaningful reduction in clot rates without an unacceptable rise in bleeding, it would offer something the broader AI clinical-decision-support field badly needs: real randomized evidence, not just retrospective accuracy statistics, that one of these tools changes what happens to patients. If it doesn’t, that will be instructive too, feeding into the larger debate — visible in the FDA’s January 2026 clinical decision support software guidance and in newly established reimbursement codes for AI-assisted imaging — over how much evidence should be required before AI tools are treated as part of routine hospital care rather than experimental add-ons. Either way, the trial’s design, including its built-in safeguards against alert fatigue and its rural-urban comparison, is likely to become a reference point for how future hospital AI trials ought to be structured.
Photo: bergy59 / PIXABAY via Pixabay
