Ben Li (Toronto, Canada) ponders the current state of artificial intelligence (AI) application in clinical practice, analysing the obstacles to scaled use.
AI has reached a level of technical maturity at which performance in many clinical and analytical tasks approaches that of experienced clinicians. In vascular care, advances now extend across a broad range of applications, including image analysis, risk prediction, perioperative guidance, and workflow optimisation. Collectively, these developments have generated much enthusiasm regarding the potential of AI to improve efficiency, consistency, and quality of care. Despite this progress, the translation of AI tools from research settings into routine practice remains limited. A persistent gap between methodological development and real-world clinical adoption continues to be a major challenge.
In an article I recently authored in npj Digital Medicine, I examined this translational gap through the lens of medical AI commercialisation. Rather than focusing on model accuracy or validation metrics, the article explored the broader ecosystem required for sustained clinical impact, including funding mechanisms, regulatory approval pathways, reimbursement structures, and integration into clinical guidelines. These factors are frequently treated as secondary concerns, yet they often determine whether an AI tool is ultimately used at the bedside.
The vascular AI literature is now rich with technically robust models, many of which demonstrate high accuracy in retrospective datasets. However, prospective validation and assessment under real-world conditions are less common. Even when such studies are performed, translation into clinical practice remains challenging.
One example discussed in my npj Digital Medicine article is an AI-guided ultrasound system developed to support abdominal aortic aneurysm (AAA) screening by non-expert operators. In a prospective clinical study, the system enabled nurses without formal ultrasound training to acquire diagnostically adequate images and identify aneurysms with high sensitivity and specificity. Importantly, the system provided real-time feedback to guide probe positioning and image acquisition, thereby reducing reliance on operator experience.
This example highlights how AI may extend diagnostic capability beyond specialist settings, potentially improving access to care and supporting population-based screening initiatives. At the same time, it underscores a recurring theme in AI research: strong model performance alone does not translate into clinical adoption. Without a viable pathway through regulatory approval, reimbursement, and professional endorsement, even well-validated technologies risk remaining confined to research settings.
A central argument of my npj Digital Medicine article is that commercialisation should be viewed as an integral component of translational research rather than as a downstream activity that follows technical development. Early-stage AI development is often supported by academic funding, which is typically well suited to exploratory research and initial validation. However, advancing a tool to a deployable medical product requires additional resources to support regulatory compliance, scaling, software maintenance, cybersecurity, and postmarket surveillance. The absence of dedicated funding mechanisms for these vital activities contributes to the attrition of many promising technologies before they reach clinical deployment.
Importantly, commercial viability is not determined solely by market size. In vascular care, AI applications that address clearly defined clinical problems, such as improving screening efficiency, supporting perioperative decision-making, or reducing administrative burden, may offer compelling value propositions. Health economic evaluation therefore becomes a critical component of AI development, enabling researchers and developers to articulate value in terms that are meaningful to health systems, payers, and policymakers.
Regulatory approval represents a major hurdle for AI-based medical devices. In most jurisdictions, clinical AI algorithms are subject to medical device regulations, requiring evidence of safety, effectiveness, and quality assurance.
In the npj Digital Medicine article, I emphasised the importance of incorporating regulatory considerations early in the development process. Establishing quality management systems, maintaining appropriate documentation, and adhering to recognised standards such as Good Machine Learning Practice are essential for successful approval. Attempting to address these requirements late in development is inefficient and often unsuccessful.
For academic teams, regulatory processes may appear opaque or burdensome. However, early engagement with regulatory expertise can reduce risk and facilitate translation, particularly as regulatory bodies increasingly clarify expectations for AI-based devices.
Even when regulatory approval is achieved, adoption of AI tools is unlikely without reimbursement pathways that align with existing health system incentives. Most current reimbursement frameworks were designed before the emergence of AI technologies and do not readily accommodate software-based interventions. As a result, uncertainty regarding payment remains a significant barrier to implementation.
As discussed in the npj Digital Medicine article, engagement with payers and policymakers should occur alongside clinical validation. Demonstrating not only algorithmic accuracy but also cost-effectiveness, efficiency gains, or improvements in patient outcomes strengthens the case for reimbursement.
In vascular care, where large-scale prevention, screening, and surveillance programmes must balance benefit against cost, these issues are particularly salient. AI tools that improve feasibility or reduce resource utilisation may offer system-level value. However, such benefits must be demonstrated through appropriately designed studies that reflect real-world practice.
Clinical guidelines remain a key determinant of practice and play a critical role in legitimising new technologies. Without inclusion in guidelines, AI tools are often perceived as optional or experimental, limiting clinical uptake. Integration into guidelines requires evidence that aligns with the priorities of professional societies, including population-level impact, safety, and alignment with established care pathways.
As highlighted in the npj Digital Medicine article, engagement with guideline committees should begin early in the development process. In areas such as aneurysm screening, where recommendations may differ between organisations, AI-enabled approaches that reduce cost or increase accessibility may influence future guidance and promote guideline concordance.
For vascular clinicians and researchers, the implications are clear. AI development cannot occur in isolation from the clinical, regulatory, and economic contexts in which the technologies are ultimately deployed. Clinician involvement early in development is essential to ensure usability, relevance, and alignment with clinical workflows, while multidisciplinary collaboration is necessary to navigate the complexities of commercialisation.
For clinician-scientists, success in AI research should increasingly be measured by translation and impact rather than technical novelty alone. Model performance metrics remain important, but they represent only one component of a broader pathway towards meaningful clinical change.
AI has demonstrated considerable promise across multiple domains of vascular care, but its clinical impact will depend on more than technical performance. As explored in my recent npj Digital Medicine article, commercialisation, encompassing funding, regulation, reimbursement, and guideline integration, represents the critical pathway through which innovation becomes routine practice.
Addressing these challenges requires early planning, sustained multidisciplinary collaboration, and a broader conception of translational success. If these conditions are met, AI has the potential to become a meaningful and sustainable component of vascular care.
Ben Li is a vascular surgery resident in the Division of Vascular Surgery at the University of Toronto in Toronto, Canada.
The author declared no relevant disclosures.
