The applications of artificial intelligence (AI) have gained considerable traction in the medical community. Encompassing many different components ranging from machine learning, natural language processing, artificial neural networks, to computer vision, AI aims to develop systems that rival human thought and intelligence. Its applications in vascular surgery have been used to identify the presence of abdominal aortic aneurysms (AAA) or carotid stenosis, to risk stratify peripheral artery disease (PAD), to identify factors associated with development of varicose veins, among others1-4. Previous work from our group utilized convolutional neural networks, a subset of machine learning, to identify the presence of infrarenal AAAs using computed tomography and found that trained models demonstrated 99% accuracy in detection5-6. Importantly, the image-based visual tasks assigned to the model were able to function autonomously, independent of manual input. By training models to function independently, it may reduce analysis time and improve data reproducibility7. Development of deep learning computer vision models may assist in predictive modeling of vascular disease. In vascular surgery, with its focus on imaging and technology to drive operative decision making and to mitigate risk in a complex patient population, AI can be an integral component of future practice. Accordingly, the strength of AI lies in its ability to find complex and non-linear relationships that may not be appreciable to human, hypothesis-driven analysis.

However, it is important to consider that AI presents unique challenges that differ significantly from traditional analyses. As the field continues to develop and expand as evidenced by the precipitous rise in publications from 1 in 1993 to 249 by 2021, applications of AI in vascular surgery will only continue to proliferate8 and one must be cognizant of the inherent biases and limitations that these methodologies present. Black box decision making, biased datasets agnostic to patient-level disparities, wide variation of present methodologies, and lack of common reporting standards are common pitfalls in AI. The purpose of the present review article is to describe common biases of AI in vascular surgery and to discuss potential strategies to mitigate their effects, especially as the future of healthcare will involve the interplay of large data suppositories and AI to deliver patient-specific, clinical outcome prognostication.