The application of artificial intelligence (AI) is becoming an increasingly prominent topic on the agendas of key vascular meetings and the contents pages of vascular surgery journals. At this juncture, Vascular News looks at some of the recently shared research and opinion on the topic, highlighting a need for “guardrails” amid significant clinical potential.
Several papers on the emerging uses of AI in vascular surgery were presented at the recent 2025 Vascular Annual Meeting (VAM; 4–7 June, New Orleans, USA). Among these were two papers from Ben Li (University of Toronto, Toronto, Canada) and colleagues—one looked at developing an AI model to predict one-year mortality after major lower extremity amputation, the other to predict one-year successful clinical use of an arteriovenous access for haemodialysis. The team found that their two models—both of which were trained on Vascular Quality Initiative (VQI) patient data—could “very accurately predict” outcomes and performed better than logistic regression. “Having that kind of information preoperatively helps the clinician and the patient decide the best treatment plan moving forward,” Li comments, considering the clinical implications of the work.
Another paper presented at VAM 2025 looked at the use of a large language model to accurately extract aortic information from abdominal imaging reports in a large, real-world, multicentre database in San Francisco, USA. Robert Chang (Kaiser Permanente Northern California, Oakland, USA) and Colleen Flanagan (University of California, San Francisco, San Francisco, USA) led the research, which found that an opensource, foundational, or “off-the-shelf” large language model was able to extract critical information about the aorta from the imaging reports. “We did not train this model ourselves at all; we crafted and refined a prompt to extract information of interest,” Chang and Flanagan detail, speaking to Vascular News about the research. The team then used this prompt on over 16,000 imaging reports in their real-world arterial aneurysm registry. The researchers found that the accuracy of the model, LLaMa 3.3, was over 90% overall. “We think this could support our ability to closely track AAAs [abdominal aortic aneurysms], including those found on screening and those found incidentally on non-aortic imaging studies, by automating the flagging of these patients for appropriate referral and follow-up,” Chang and Flanagan remark.
Elsewhere at VAM 2025, Justin Bader (Yale School of Medicine, New Haven, USA) presented a paper outlining the creation of a prediction model for safe contrast volume thresholds to prevent post-contrast acute kidney injury (PCAKI) after endovascular aneurysm repair (EVAR). The research team is now using AI to harness the model’s future potential. Bader explains that the team—led by Cassius Iyad Ochoa Chaar (Yale School of Medicine, New Haven, USA)—used VQI data to create a “calculator” that allows physicians to generate a recommended contrast volume to minimise PC-AKI risk by inputting 13 patient-specific variables. “It serves as a guideline for surgeons when they’re operating,” he says. The model, Bader tells Vascular News, is “working extremely well” and the team is presently collaborating with statistics experts at Yale on advanced AI and machine learning techniques to make the calculator “even more accurate”.
VAM 2025 also saw data from Prem Chand Gupta (CARE Hospitals, Hyderabad, India) and colleagues that looked at the correlation of imaging characteristics of carotid plaque with clinical and histopathological features, and the application of AI. “We found that routinely performed ultrasound by us was better than CT [computed tomography] angiography and MR [magnetic resonance] angiography in deciding whether the plaque was vulnerable or not,” he tells Vascular News. Gupta shares that, for asymptomatic carotid patients, it is currently unclear whether intervention should be offered or not. While guidelines suggest that patients who are asymptomatic should not be offered surgery, around one-third of the patients in Gupta and colleagues’ study who were not operated on were found to have vulnerable or soft plaque, with almost 20% of these patients going on to have a stroke within one year. “If we are able to identify soft or vulnerable plaque in asymptomatic patients, we can be more confident that operating on these patients will prevent stroke,” Gupta notes, considering his ultimate ambition for the clinical application of the team’s research.
“AI is going to have an increasingly important role in medicine” — Ben Li
As research dedicated to the applications of AI in vascular surgery proliferates, so too do calls for its safe integration into the specialty.
Bader raises the issue of liability, for example, asking: “Who’s liable for a botched surgery if AI is leading it?” This is an issue also raised in a letter to the editor of the Journal of Vascular Surgery (JVS), in which Antonio V Sterpetti (Sapienza University, Rome, Italy) and colleagues state: “Recommendations made by AI should remain simple suggestions, which cannot substitute for the opinion of surgeons in a complex clinical analysis, to prevent legal controversies and to preserve the dignity of patients and vascular surgeons.” The letter was published in response to a study by Joachim Sejr Skovbo (Odense University Hospital, Odense, Denmark) and colleagues, who in JVS had previously outlined the successful development of the SHAPFire AI tool to identify AAAs at increased risk of rupture with “significantly higher” accuracy than diameter alone. In their response to the letter, Skovbo et al write that “AI recommendations should be viewed as valuable adjuncts that enhance—not substitute for—clinical expertise”, going on to stress that vascular societies “could help facilitate the integration of AI decision-making tools while ensuring they support, rather than replace, clinical judgment”.
Amun G Hofmann (Klinik Ottakring, Vienna, Austria) also raises some of the challenges associated with integrating AI in vascular surgery in a recent letter to the editor of the European Journal of Vascular and Endovascular Surgery (EJVES). Specifically, he highlights the importance of establishing the digital infrastructure, monitoring safety and efficacy, and considering legal and ethical requirements.
Li tells Vascular News there is a need to ensure “guardrails” are in place before AI is deployed in healthcare. “AI is going to have an increasingly important role in medicine,” he says, before stressing that these are early days and that the “true impact” of AI in the field may not be apparent for some time to come. Chang and Flanagan agree that “caution is vital” when it comes to utilising AI tools, especially with regard to sensitive patient data, and suggest that “protocols would need to be in place to bring some of these technologies online on reasonable timelines while also balancing potential security concerns”.
Another concern, highlighted by Gupta, is whether AI could replace the surgeon altogether if combined with the power of robotics. At present, however, Gupta is “really not worried” about this possibility. “There are just too many variables which come into surgery,” he says. “It is going to be very, very difficult for AI to have enough data to be able to deal with different variables […] unless there is a human to guide it.” Instead, Gupta sees AI as an adjunct for now. “At least for the next few generations I don’t think AI is going to take over diagnostics or robotics in a way we should be concerned about,” he comments. “I think we should use it as a good tool.”
