AI model could help identify children at higher risk for blood clots
Click Here to Manage Email Alerts
Key takeaways:
- An AI model demonstrated the ability to identify children at higher risk for hospital-acquired venous thromboembolism.
- Most clinicians did not accept recommendations based on the AI data.
An AI model may help identify children at higher risk for developing hospital-acquired venous thromboembolism, according to a presenter at AIMed24 Annual Meeting.
However, research into the viability of this approach identified barriers hindering implementation of this AI approach into practice.
“The model works well. I believe that it can be embedded into the medical record system to give people information more quickly than we could come up with ourselves as physicians,” Shannon Walker, MD, assistant professor at Vanderbilt University Medical Center, told Healio. “But I think there’s still a lot of room for improvement on how we deliver that information to other medical providers, and how we can kind of convince them that AI works very well to do these things.”
Background
VTE occurs in approximately one in 100,000 children. However, VTE events increased from about five per 10,000 hospital admissions during the early 1990s to between 30 and 50 per 10,000 admissions in the 2010s, Walker said.
Treatment of VTE differs considerably between adults and children.
“In adults, they use blood thinner medications in everybody because [clinicians] know that the risk is pretty high,” Walker said. “In pediatrics, it’s pretty much the opposite. We’re not doing a very good job. There are not standard recommendations in terms of who we should be thinking about treating.”
Walker and colleagues began work on an AI model to see if they could identify patients at greater risk for VTE and, thus, intervene early.
“We wanted to design something that wasn’t specific for just one population of patients,” she said. “It’s not just a model only for the patients in the [neonatal intensive care unit] or just for patients who have cardiology problems, but something that could be applied broadly to all the patients who are in the hospital.”
Walker and colleagues built and validated the model in more than 155,000 children at Monroe Carell Jr. Children’s Hospital at Vanderbilt.
Of the 11 variables in the model, the ones that had the greatest predictive value for VTE included central line inserted (OR = 4.9; 95% CI, 4.1-5.9), thrombosis history (OR = 8.6; 95% CI, 6.6-11.3), cardiology consultation (OR = 4.2; 95% CI, 3.5-5) and blood gas measured (OR = 3.1; 95% CI, 2.5-3.7).
Model development had a c-statistic of 0.908 (95% CI, 0.896-0.918) and the validation had a c-statistic of 0.904 (95% CI, 0.894-0.913).
Methods and results
Walker and colleagues tested the model in the randomized Children’s Likelihood of Thrombosis trial.
The trial included 17,427 children (median age, 1.7 years; interquartile range, 0-11.1; 52.5% girls; 67.4% white) hospitalized at Monroe Carell Jr. between Nov. 2, 2020, and Jan. 31, 2022.
The model calculated each child’s hospital-associated VTE (HA-VTE) daily.
Researchers randomly assigned patients to one of two groups.
They assigned 8,717 to the intervention group but took no action on lower-risk patients (risk score less than 2.5). The 8,710 children assigned to the control group received standard of care regardless of score.
All high-risk cases (risk score of at least 2.5) in the intervention group were reviewed. Researchers reached out to primary teams for children they felt needed prophylaxis and suggested treatment.
HA-VTE rate served as the primary endpoint.
Results showed comparable HA-VTE rates in the intervention group and control group (0.7% vs. 0.9%).
However, primary teams only took researchers treatment recommendations about one-quarter (25.8% of the time).
The AI model classified 86.2% of patients who had HA-VTE in the control group as having high risk. The model classified 92.2% of those who had HA-VTE in the intervention group as having high risk.
The model had a c-statistic for the control group of 0.799 (95% CI, 0.725-0.856).
“We think that a large reason why the study was not successful was because the uptake of the recommendations was not very successful,” Walker said.
Next steps
Researchers identified several reasons why primary teams declined treatment recommendations. These included imminent patient discharge, central line removal likely within 24 hours, and subspecialists declining the recommendation.
Walker cited concern about bleeding to be one of the main reasons their prophylaxis recommendations had been declined.
“People overestimate the risk for bleeding when patients are on blood thinners, especially the doses we were going to recommend,” she said. “I do think we’re both underidentifying and undertreating patients who are at risk right now.”
Walker added that teams that had seen higher rates of clotting “were more accepting” of the recommendations, Walker added
“A sizeable group of providers who said something like, ‘I don’t necessarily think this patient is at high risk, so I’m not going to do blood thinners,’” Walker said. “They did not acknowledge the combination of variables make you high risk. People disagreed with the model based on their own judgement.”
AI models will not be successful unless clinicians buy into its abilities, Walker said. That requires education and determining the best way to present the information so clinicians act on it, she added.
Walker and colleagues got preliminary feedback that providers preferred to receive information from a pediatric hematologist instead of an automated alert, but the results did not bear that out.
“How do we improve that and optimize displaying the information to the teams?” Walker said. “That’s where we’re trying to go next. I think that’s where all medical AI should be going.”
References:
- Walker SC, et al. Using AI to predict and prevent pediatric blood clots. Presented at: AIMed24 Annual Meeting; Nov. 17-19, 2024; Orlando.
- Walker SC, et al. JAMA Netw Open. 2023;doi:10.1001/jamanetworkopen.2023.37789.
For more information:
Shannon Walker, MD, can be reached at shannon.walker@vumc.org.
Collapse
Walker SC, et al. Using AI to predict and prevent pediatric blood clots. Presented at: AIMed24 Annual Meeting; Nov. 17-19, 2024; Orlando.
Click Here to Manage Email Alerts