By Deborah Borfitz

July 12, 2021 | Monroe Carell Jr. Children's Hospital at Vanderbilt has responded to the growing incidence of hospital-associated venous thromboembolism (blood clots) in kids by developing an algorithm to predict patients who are at high-risk in real time. A pragmatic clinical trial underway now is testing the effectiveness of a newly validated, general pediatric predictive model along with targeted hematology review of high-risk patients to identify patients who may benefit from prophylactic interventions. The goal of the study is to reduce the overall rate of blood clots at the hospital, according to pediatric hematologist Shannon Walker, M.D., the study’s principal investigator.

Multiple studies across the country have shown that the frequency of blood clots is increasing in pediatrics, Walker says. The proliferating number of children’s hospitals in recent decades, coupled with more support devices and surgeries keeping seriously ill children alive, are thought to be the main reasons.

Admission to a children’s hospital is associated with a 1% risk of developing a blood clot—far lower than that among hospitalized adults—but the condition can be severe (e.g., compromising a limb, causing a pulmonary embolism) and lead to lifelong complications, Walker says. Clots are treated with blood thinning medications for weeks to months post-discharge, and the hope is that the new risk-prediction model will spare children those types of complications down the road.

A recently published article in Pediatrics (DOI: 10.1542/peds.2020-042325) describes the predictive model, which includes 11 risk factors based on an initial analysis of more than 110,000 pediatric admissions and validation in over 44,000 separate admissions to the hospital. The model yielded excellent discriminatory ability in both the derivation and validation cohorts, with blood clots being most strongly associated with a history of thrombosis, presence of a central venous catheter, and having an underlying cardiovascular condition (e.g., congenital heart disease).

The predictive model has been automated to run within the Epic electronic medical record (EMR) of each patient admitted to the Children’s Hospital at Vanderbilt, says Walker, and is intended to broadly apply to everyone from tiny infants in the neonatal intensive care unit to older adolescents. The broad use of the predictive model necessitated elimination of some variables, such as body mass index, that has multiple appropriate ranges based on pediatric age.

Currently, national groups are examining why blood clots are becoming more common in children and how to identify those at risk, Walker says. But this is the first such predictive tool built into the EMR that automatically extracts risk variables and provides updated scores daily.

The intent of making the information in the model applicable to all patients regardless of admission reason or underlying clinical conditions is to broaden the clinical impact hospital-wide, Walker continues. Programmers charged with building predictive models into the EMR also needed a “point of simplicity” to accomplish this goal in a reasonable timeframe. A separate group at Vanderbilt is trying to develop a similar predictive model for adult patients, she adds, since their risk factors for blood clots are different than those in children.

A key player in the development of the model and the pragmatic clinical trial is the Advanced Vanderbilt Artificial Intelligence Laboratory (AVAIL), Walker says, which was critical to bringing together the needed expertise, facilitating the necessary approvals, and getting the model built into the EMR. Only in its second year, the program is devoted to shepherding the use of AI-powered risk prediction tools at Vanderbilt University Medical Center (of which Monroe Carell Jr. Children’s Hospital is a part) to provide more personalized care to its patients.

Dynamic Tool

Use of the model together with a targeted intervention is being investigated in the Children's Likelihood of Thrombosis (CLOT) study that is enrolling everyone admitted to the Children’s Hospital and calculating each patient’s score from the clot-prediction model, Walker says. Half of those patients are randomized so that the study physicians see the scores in real time; scores for the other half are not released to the study investigators. At all times, patients receive care as usual.

Those in the intervention arm whose scores are released and are identified as being at high risk for a blood clot receive an additional subspecialty review by a study-related hematology provider, says Walker. The hematologist makes individualized recommendations for clot risk reduction to the treatment team and the patient’s family.

Risk scores are recalculated daily, so new risk factors (e.g., infection requiring central venous catheter for antibiotics) are picked up as they happen, she notes. “As patients become higher and higher risk, we’re able to target them at any point in their hospital stay.”

If the clinical decision support tool successfully deploys at Children’s Hospital at Vanderbilt, Walker says, the plan is to broaden its use to other pediatric facilities. Epic is a commonly used EMR in many hospitals around the country, making it relatively easy to translate to other children’s hospitals.

Many Models

Based on the success of the predictive model for blood clots in hospitalized children, AVAIL created a similar model for hospitalized adults, says biostatistician Daniel Byrne, director of artificial intelligence research at the lab. That model is currently being updated and validated with a new data set, and in the fall a grant will be submitted to the National Institutes of Health (NIH) in support of a pragmatic randomized controlled trial (much like CLOT) at Vanderbilt’s adult hospital.

“AVAIL is also working to predict date of neonatal ICU discharge to improve the transition for the patient and family,” says Byrne. “For the past 10 years, Vanderbilt has been running models to predict hospital readmission and pressure injuries, or bedsores… [and is] currently updating those models and planning to assess [them] in randomized pragmatic trials.”

A model that taps the electronic health record to predict which cardiac surgery patients will need to be intubated is the subject of a paper just accepted for publication in The Annals of Thoracic Surgery, he says. “This will be assessed with randomization soon.” 

Another project has AVAIL helping to build prediction models of atrial fibrillation after cardiac surgery, Byrne says. An NIH grant was just submitted in support of research to quantify the predictive value of clinical and genetic risk factors for adverse outcomes in patients with systemic lupus erythematosus.

“While there is great interest in artificial intelligence in medicine and billions of dollars are being invested in this area, to date there is little evidence that it benefits patient outcomes,” says Byrne.  “AVAIL’s niche is to move these projects to the end of the last mile by assessing with randomized pragmatic trials if they have an impact compared with usual care.” This is additionally essential to ensuring that deployed models do not create any inequities.