The clinical stakes are considerable. Peripheral artery disease alone affected an estimated 202 million people in 2010, a 23.51 percent increase from 2000, driven largely by low- and middle-income countries, and a more recent estimate cited in the review puts the figure at approximately 316.5 million people aged 25 and older in 2023, corresponding to a global prevalence of 6.58 percent. The condition has been associated with a threefold increase in the risk of cardiovascular mortality. Minimally invasive approaches, including endovascular interventions, hybrid procedures, and robot-assisted surgery, have transformed the management of aneurysms, dissections, embolism, and ischemia by shrinking incisions, reducing procedural trauma, shortening recovery, and, in selected cases such as elective endovascular abdominal aortic aneurysm repair, allowing treatment under local anesthesia. But these techniques carry their own burdens: high cognitive workload for operators, higher rates of technical failure than open surgery, risks of endoleaks requiring reintervention, and high implementation costs in specialized facilities.

Machine learning emerges in the review as the most mature predictive tool. A series of studies by Ben Li and colleagues using the Vascular Quality Initiative registry developed models to predict major adverse limb events or death one year after endovascular surgery for peripheral artery disease, as well as stroke or mortality after transcarotid artery revascularization and outcomes after endovenous vein ablation. In each case the eXtreme Gradient Boosting algorithm, a scalable tree-boosting framework, achieved the strongest performance, with an area under the receiver operating characteristic curve of 0.94 and accuracy around 0.85 to 0.87. A related analysis of the American College of Surgeons National Surgical Quality Improvement Program database extended this to 30-day outcomes after endovascular aortoiliac revascularization, with XGBoost again leading, achieving AUROC values of 0.86 to 0.97 for secondary outcomes including reintervention, amputation, bleeding, and readmission. The authors caution, however, that because these registry studies came from largely the same research group using similar models, they should be read as a related body of evidence rather than independent confirmation of generalizability.

Other machine learning work reviewed spans risk stratification and skill assessment. Ross and colleagues found that Elastic Net penalized regression outperformed logistic regression for identifying peripheral artery disease with an AUC of 0.87, while Random Forest performed best for detecting undiagnosed disease and predicting eight-year mortality. Huang and colleagues built a logistic regression model for patients with acute vertebrobasilar artery occlusion, identifying the NIH Stroke Scale, symptomatic intracranial hemorrhage, and occlusion-to-groin puncture time as the most influential predictors, and notably validated it on an external cohort of 237 patients from 16 European stroke centers. Nishibe and colleagues used decision tree analysis in 169 patients to predict three-year mortality after endovascular aortic repair, achieving 68.7 percent accuracy. Guo and colleagues trained Support Vector Machine and Random Forest models on anatomical features and operative motion data, showing that adding a novel Vascular Difficulty index improved surgeon skill classification accuracy by 10 percent, a reminder that objective performance assessment must account for the complexity of each patient’s vasculature.

Deep learning, built on multilayered artificial neural networks, dominates the imaging side of the field. The U-Net architecture, with its U-shaped encoder-decoder structure and skip connections that preserve fine spatial detail, underpins much of this work. Li and colleagues applied semantic U-Net segmentation to fluoroscopic images to flag “No-Go” landing zones during endovascular aneurysm repair, achieving pixel-wise accuracy of 0.97, though overlap metrics were more modest. Saitta and colleagues developed a fully automated 3D U-Net pipeline trained on 395 CT scans and tested on 70 independent scans, reaching a mean Dice score of 0.95 while quantifying landing-zone diameter, angulation, tortuosity, and aortic arch curvature for thoracic aneurysm repair planning. Smorenburg and colleagues trained a multitask convolutional network on 660 digital subtraction angiography perfusion maps from 220 patients to detect and localize endoleaks during repair, achieving an AUROC of 0.85. Other models segment retinal and peripheral veins, plan guidewire paths in real time using a generative adversarial network trained on 350,000 samples from ten surgeons, and fuse physiological and motion sensor data to recognize operator behaviors with up to 0.99 accuracy.

Natural language processing occupies a narrower but growing niche, focused mainly on extracting clinical information from records. A rule-based algorithm developed at the Mayo Clinic diagnosed abdominal aortic aneurysms from radiology reports with a positive predictive value of 0.98 and accuracy of 0.97, and was validated prospectively. A commercial system evaluated on more than 230,000 imaging reports from the Kaiser Permanente Northern California registry detected aneurysms and extracted maximal diameters with high fidelity. At Duke University Health System, a label-embedding attentive model called PAD-ML diagnosed peripheral artery disease from electronic health records with an AUROC of 0.89, outperforming structured-data approaches. Perhaps most eye-catching is a voice-controlled interface integrating a large language model with a robotic C-arm X-ray system, tested on 212 prompts in a phantom setting, which executed satisfactory actions 97 percent of the time and could reduce radiation exposure and anesthesia duration by automating intraoperative imaging tasks.

Robotics and reinforcement learning represent the field’s most ambitious frontier. The review traces the lineage from the PUMA 200 robot’s stereotactic brain biopsy in 1985 and Hopkins’ endoscopy advances, through the first endovascular robotic platform CorPath 200 in 2010, to the R-One+ system launched in 2023. In a prospective human trial, robot-assisted femoropopliteal angioplasty with CorPath 200 in 20 patients achieved 100 percent technical and clinical success. Haptic-feedback catheter systems using magnetorheological fluids reduced collision forces and task times among novice operators in vitro. Autonomous systems, however, remain largely preclinical: VascularPilot3D achieved perfect success across phantom navigation tasks, deep Q-network guidewire navigation succeeded in 95 percent of trials in a cerebral neovascular model, and a proximal policy optimization system reached success rates of 96 and 94 percent in simulated vascular branching models. Untethered magnetic microrobots navigated ex vivo porcine aortas with hemocompatible coatings that reduced bacterial adhesion. None of this yet establishes clinical safety or readiness.

The translational pathway the authors describe is sobering. Regulatory clearance does not guarantee clinical effectiveness: of 521 FDA-authorized medical AI devices examined in one analysis, 226 lacked publicly available clinical validation data, and only 17 of 130 FDA-approved device evaluations reported performance across demographic subgroups. External validation is rare in vascular surgery AI; a systematic review cited in the paper found that only 9 of 212 machine learning studies conducted it, with 94.8 percent showing high overall risk of bias. The “black-box” problem, cohort bias, and the risk that models reproduce racial disparities even when race is not an input, as illustrated by a widely deployed commercial risk algorithm trained on healthcare costs rather than illness severity, all compound the concern. The authors advocate reporting frameworks such as TRIPOD+AI and PROBAST+AI, prospective “silent-mode” deployment in which model outputs are hidden from clinicians before live use, and health-economic evaluation, noting that only 28 percent of 117 medical AI studies assessed implementation costs.

The review’s cautionary tales and its limitations both deserve emphasis. The IBM Watson-based Oncology Expert Advisor project at MD Anderson Cancer Center was suspended after roughly US$62 million in expenditure, following occasionally unsafe recommendations and a software bug, a reminder that hype can outpace governance. Surgical robots priced between $1 and 2.5 million, plus training and per-procedure costs, challenge cost-effectiveness and equitable access, potentially concentrating advanced care in well-resourced centers. Liability law, which currently protects clinicians who adhere to the accepted standard of care, leaves AI-assisted deviations in a gray zone, which is why today’s systems are positioned mainly as confirmatory decision-support tools. The review itself was a narrative rather than systematic review, relying primarily on Google Scholar, and the included studies varied widely in design and validation, limiting direct comparisons. Its bottom line, however, is clear: AI techniques each hold distinct roles across the vascular surgical pathway, but converting impressive AUROC scores and phantom successes into safe, effective, and affordable clinical practice will require prospective multicenter trials, standardized reporting, and honest accounting of costs, equity, and accountability.

Subject of Research: Applications of artificial intelligence, including machine learning, deep learning, natural language processing, and robotics, in minimally invasive vascular surgery

Article Title: Artificial intelligence for vascular minimally invasive surgery: techniques, clinical integration, and challenges

Article References: Jiang, C., Bortnovskii, B., Gul, I., Özkan, M., Zhang, L., Yu, D., Pandey, V., Wang, R., Wu, J., & Qin, P. (2026). Artificial intelligence for vascular minimally invasive surgery: techniques, clinical integration, and challenges. Frontiers in Cardiovascular Medicine, 13, Article 1966224. https://doi.org/10.3389/fcvm.2026.1966224

Image Credits: AI Generated

DOI: 10.3389/fcvm.2026.1966224

Keywords: artificial intelligence, vascular surgery, minimally invasive surgery, machine learning, deep learning, natural language processing, surgical robotics, endovascular intervention, peripheral artery disease, clinical validation, medical regulation, reinforcement learning