Computer Science Meets Vascular Surgery: Keeping a Pulse on Artificial Intelligence

Abstract

Artificial intelligence (AI) based technologies have garnered interest across a range of disciplines in the past several years, with an even more recent interest in various healthcare fields, including Vascular Surgery. AI offers a unique ability to analyze health data more quickly and efficiently than could be done by humans alone and can be utilized for clinical applications such as diagnosis, risk stratification, and follow-up as well as patient-utilized applications to improve both patient and provider experience, mitigate healthcare disparities, and individualize treatment. As with all novel technologies, AI is not without its risks and carries with it unique ethical considerations that will need to be addressed prior to its broad integration into healthcare systems. AI has the potential to revolutionize the way care is provided to patients including those requiring vascular care.

Introduction

Artificial intelligence (AI), a discipline of Computer Science that was first popularized in the 1950s, has rapidly gained traction in business and society over the past several years and is now being applied to various aspects of healthcare. This rapid increase in AI utilization is likely due in part to significant advances in computing power and data generation, as well as growing media attention.(1) AI has been the logical choice for providers, researchers, and patients alike to help navigate the increasing amount and complexity of medical data being generated.

The term AI broadly describes a discipline of Computer Science in which computers are programmed to perform certain tasks that are typically thought to require human intelligence or features such as learning, reasoning, decision making, and sensory interpretation such as speech recognition or visual processing (Figure 1).(2) Development of AI requires a multidisciplinary approach that draws on principals of logic, mathematics, statistics and computation to create systems that mimic human intelligence and cognition.(3) Machine learning, whereby computer algorithms are trained or “learn” from the input of large dataset, has been especially popular in the healthcare field as it can be used to analyze large amounts of data, identify patterns, and make decisions without being preprogrammed to perform some sort of conventional task. Machine learning has been popularized by the advent of self-driving vehicles, which incorporate the constant input of new data to “learn” and make decisions. Deep learning, a subset of machine learning, has become increasingly investigated for use in the healthcare field due to the improved ability to classify, recognize, and detect previously recognized patterns in an enormous quantity of raw data which does not require pre-processing by a human “supervisor” as does machine learning. AI has attracted interest from nearly all fields of medicine because of the possibility to automate tasks that conventionally require specific human intervention or action.(4)

Figure 1:

Schematic of the relationships of Artificial Intelligence, Machine Learning, and Deep Learning. Created with Biorender.

There are several ways in which the field of Vascular Surgery could potentially benefit from different AI applications; to improve diagnosis, management, and follow up of vascular diseases, ignite vascular surgical research, and improve patient experience and outcomes. This review focuses on the ways in which AI has been or may be utilized in the field of Vascular Surgery with special attention paid to how AI affects both the clinician and patient experience.

AI for Clinicians

The scope of Vascular Surgery covers a wide spectrum of arterial, venous, and lymphatic diseases in an ever-evolving practice landscape that requires vascular surgeons to be on the forefront of innovation. This section focuses on practical applications of AI in commonly treated vascular pathologies such as aortic aneurysms, peripheral arterial disease, carotid stenosis, and acute and chronic venous diseases, although innumerable opportunities exist for applications throughout Vascular Surgery.(5–9)

Diagnosis

In Vascular Surgery, image analysis is a critical component in the diagnosis and management of vascular disease. The current system of image interpretation and segmentation is subject to variability due to differing acquisition techniques and individual radiologists’ skill or bias and is not only time consuming but may introduce error. AI has the potential to improve overall diagnostic efficiency, reduce mistakes, and limit discrepancies in diagnostic data interpretation. Additionally, AI may reduce the workload of healthcare personnel by reducing the human effort requirement and time needed for each imaging study.(8) For example, AI-derived methods have been used to improve aortic segmentation and allow for characterization of the geometry and morphology of aortic aneurysmal disease and predict growth, rupture, and 30-day mortality. (10–15) VASIM is an automated vascular imaging software tool developed using head and neck CTA data, to evaluate extra-cranial carotid atherosclerosis burden and evolution with an 83% success rate in detecting stenosis > 50%.(16) Imaging-based AI analyses have also been applied to peripheral arterial disease, using arterial pulse wave analysis (17) and CTA (18) to detect arterial stenosis with impressive accuracy. Titano et al. developed a deep learning algorithm to triage neurologic findings on CT head roughly 150 times faster than humans.(19) Similar techniques could be applied to vascular emergencies such as aneurysm rupture, arterial dissection, or arterial occlusion among others, which may expedite life- or limb- saving interventions.

AI techniques based on object recognition, segmentation, and classification have been applied not only to formal diagnostic imaging such as CT or ultrasound, but to patient or clinician provided photographs as well. The utility of photograph-based techniques in Vascular Surgery has been demonstrated by the development a tools to determine venous ulcer stage with an average accuracy of 99.55% (20, 21) and to identify chronic venous disease with accuracy up to 94%, (22, 23). This has been taken a step further by the use of AI to identify ulcers with characteristics that may indicate their resistance to standard therapies and allow for earlier consideration of more personalized care.(24)

In addition to improving diagnostic accuracy and efficiency, AI has the potential to improve the accessibility of specialized vascular care for those that may be homebound or live in rural or remote areas with limited access to specialists, reaching beyond the limits of conventional telehealth. Imaging studies performed at remote institutions could be interpreted using AI-based imaging software tools to effectively triage vascular pathologies and initiate an expedited referral cascade to the nearest available vascular surgeon if indicated. Similarly, physicians in remote areas could benefit from photographic imaging-based tools by providing improved diagnostic certainty and receiving expert advice on patient care in outlier type circumstances. AI-based telehealth frameworks could additionally be developed to effectively triage patients with lower acuity concerns potentially decreasing utilization of emergency resources and increasing resources and efficiency in the management of acute and life-threatening presentations.(25)

Risk-stratification

While expedient and accurate diagnosis is key to the effective delivery of vascular care, patient risk stratification remains vital to selecting the most appropriate therapeutic approach for a patient. Imaging-based AI has been used for cardiovascular specific risk-stratification, to predict aneurysm growth, rupture, complication, evolution, and the possible need for re-intervention.(12, 26, 27) Machine learning has been used to analyze electronic health records and identify patients with peripheral arterial atherosclerosis that favored detection of patients at high risk for cardiac or cerebrovascular events.(28) Similarly, deep learning approaches to electronic health record data have been used to develop predictive models for risk stratification of patient at high risk for development of deep venous thrombosis (DVT) to guide prophylactic medical therapy and to prioritize diagnostic imaging and disease management of those with suspected DVT while preventing unnecessary over-utilization of precious resources for those in which DVT is less likely.(29, 30)

With advancements in deep learning, enormous amounts of electronic health data may be processed without defined inputs or outputs, allowing the computer to identify previously unrecognized trends or patient clusters that exist under a broad diagnostic umbrella. There are many medical diagnoses and clinical syndromes that exhibit heterogeneity that present unique treatment challenges with no single drug or treatment that is universally effective. As an example to address this problem in heart failure patients, sophisticated machine learning algorithms were developed to detect novel patterns in patient data and identify “phenogroups” of patients which displayed differing pathophysiologic profiles, different outcomes, and distinct clinical trajectories. (31) Similar phenomapping type studies could be performed for heterogenous vascular conditions, such as aortic aneurysmal disease, to aid clinicians’ decision making; whether clinical observation is appropriate or more urgent surgical intervention is warranted. In this way, AI can provide suggestions for personalized medicine.

Prognosis

Similar to improved risk-stratification, AI has been used to aid in the prediction of surgical outcomes and could feasibly allow for individualized peri-operative care and follow up, as well as to improve treatment response metrics. Kordzadeh et al. created a machine learning algorithm to predict functional maturation of radiocephalic arteriovenous fistulae with >80% accuracy and identified those characteristics which placed patients at higher risk for fistula failure.(32) Others have demonstrated the utility of AI in predicting varicose vein recurrence following invasive therapies (33) and, as mentioned above, for predicting trends in healing of venous ulcers.(24) The development of these and similar tools could lead to individualized follow up for patients assessed to be at high risk for adverse outcomes and may even help to identify patients that would benefit from alternate interventions or therapies.

Conversely, AI may be used to identify patients who are at a lower-than-average risk for adverse outcomes. In a healthcare era in which resources and expenditures must be balanced against patient need, several surgical interventions have been trialed for short discharge type post operative protocols to decrease hospital length of stay without compromising patient safety. For example, patients undergoing endovascular aortic aneurysm repair (EVAR) may be safely discharged on postoperative day 0 or 1, though a degree of provider hesitation and lack of experience with short discharge following EVAR may preclude its broad application.(34–36) To provide objective measures and augment physician decision making, a prediction score based on the Vascular Quality Initiative (VQI) data for patients undergoing EVAR has been developed and used to select patients who could be safely discharged home the day of surgery or on postoperative day 1 without compromising 30 day mortality or frequency of post-operative complications.(37) AI has the potential to provide insight on how to maximize healthcare efficiency without compromising safe, equitable patient care.

Research

Clinical utilization of AI is not limited to the patient care setting, with AI being used to advance research in quite exciting ways. With the increasing computational ability and speed of commercially available computer systems, enormous quantities of data can now be processed for research purposes in only a fraction of the time it would take humans to process. With AI, multiple “views” from different datasets, including genomics, proteomics, transcriptomics, etc. can be integrated to form a “multi-omics” approach to identifying genes or polymorphisms in big data arrays.(4) Genomic and transcriptomic data has been used in conjunction with AI-based image segmentation to characterize genes associated with thoracic aortic aneurysms and develop a polygenic score for ascending aortic aneurysms. (38) Similarly, genomic data from individuals with AAA has been used to develop novel analytical frameworks to further delineate pathological variants, gene mutations, and disease associated pathways and to improve accessibility of AI-based tools.(39, 40) This improved efficiency in genomic medicine has the potential to allow for individualized clinical decisions to be made based on personal “multi-omic data.”

AI may also prove useful in the design and execution of successful clinical trials. Prior to the initiation of a trial, AI can be used by pharmaceutical companies for drug development by facilitating in situ drug development with characteristics like those currently on the market but perhaps with structural differences that would improve bioavailability, duration of action, or off-target effects. Similarly, AI could be used to analyze previously abandoned compounds quickly and efficiently or “orphan drugs” for suitability in previously unexpected treatment paradigms. AI may be helpful for patient recruitment and patient cohort selection, with the ideal situation allowing for patient-specific genome profiling to determine suitability for trial participation. (41) Assimilation of AI-based tools in patient monitoring throughout a trial may enhance patient adherence and increase the efficiency of endpoint detection. These tools could be easily and conveniently introduced into the patients’ daily activity in the form of mobile device applications or wearable technology.(41) As in other applications, AI could then be easily utilized to efficiently analyze the large quantity of data that are acquired throughout the trial period. AI has the potential to optimize all portions of a clinical trial that ultimately leads to improved trial success with reduced research and development costs.

AI for Patient Use

AI methods and models are not solely for use by clinicians, researchers, and providers but may provide improved patient care through direct patient utilization. Human behavior patterns are directly and indirectly linked to overall health status, and analysis of behavioral data using AI could insert an additional viewpoint in our holistic view on health and disease states. With the ever-increasing use of social media applications in society, AI tools that analyze patient behavior over the internet represent an untapped potential. While this has been used mainly to diagnose or predict behaviors in psychologic and psychiatric conditions, (42–46) data from patients’ Twitter posts were used to identify associations between patients’ tweets and risk of cardiovascular events; it could be possible to identify similar associations in those at risk for development of peripheral vascular disease.(47)

In addition to written behavior and general online presence for behavior analysis, use of video and conversational data should be considered as well. The Chinese company Tencent and London-based healthcare firm Medopad teamed up to develop a tool to diagnose Parkinson’s Disease based on the detection of movement characteristics from video footage of patients, with the concept that this technology could be applied using a smartphone app rather than requiring a hospital visit.(48). Similarly, data linked to mobile sensors in cellular devices may provide insight to chronic conditions such as movement disorders and chronic pain disorders among others. Correlations have been found between GPS location, phone usage, and severity of depression, as well as the severity of Parkinson’s related symptoms.(49, 50) While this has not been explicitly applied to Vascular Surgery, it could be considered for use in detecting gait abnormalities or symptom severity in those with claudication or perhaps diabetes-associated foot problems.

Consistent with contemporary mainstream use of AI, targeted advertising could be harnessed for use in Vascular Surgery. Advertisements for local vascular surgery groups or new relevant medications could be suggested for patients based on their search engine use, behavior on social media, or purchasing history on online shopping sites. For example, a patient who searches “foot ulcer treatment” or “leg cramping” could then be returned an advertisement for vascular surgery screening or follow up that then prompts them to seek additional care.

The most direct way in which patients could benefit from the use of AI is through easily accessible mobile applications. eHealth, a term for the use of technology, including AI, for the prevention, treatment, and maintenance of health as well as mHealth, or the use of mobile and wireless applications for the same purposes, represent ever-growing mechanisms for the delivery of healthcare.(51) Thousands of applications have been developed in the realm of health and fitness, from the tracking of routine diet and exercise to patient-reported symptoms, to goal setting, access to health information, and more recently EKG, oxygen saturation and heart rate monitoring by way of wearable devices. Vascular-specific applications for patients provide support and motivation for exercise programs to improve upon symptoms of claudication, track symptoms, and provide medical advice and health information (52, 53). Mobile applications offer a convenient and tailored experience for patients, and by developing an interface between mobile applications and the electronic health record could lead to more comprehensive and overall improved patient care. While eHealth and mHealth applications do not always use AI technologies to deliver patient care, they represent a well-established avenue for the incorporation of AI into everyday use.

A Word on Chatbots

Chatbots have recently been in the spotlight both in mass media and in the realm of health and scientific discovery with the introduction of the high-powered and refined chatbots ChatGPT and its successor GPT4 that use language learning models to generate a text-based output in response to a written prompt.(54) ChatGPT and GPT4 have been utilized in healthcare and research, from generating patient notes and radiology reports (55, 56) to generating full length scholarly manuscripts. Several publications have been submitted with ChatGPT listed as an independent author. This has raised concerns over the validity of using the program for scholarly publications, as the risk of introducing bias, error, and plagiarism is evident, prompting the World Association of Medical Editors to release a statement guiding use of chatbots and specifically ChatGPT in scholarly works. (57)

While the integration of ChatGPT and GPT4 into healthcare continues to be explored, less complex chatbots have already been used in several ways to improve healthcare delivery. For clinical care, chatbots can be used to gather patient data, help arrange follow up, and remind patients of upcoming appointments or other events such as medications that are soon to be refilled. Chatbots may be useful to providers for clinical decision making and care algorithms. For patients, chatbots, like other types of mobile apps, may provide an avenue for symptom tracking or reporting, or to have clinical questions answered or triaged. For example, the chatbot Insomnobot-3000 was developed to aid patients with insomnia whereas the Endurance chatbot was developed as a companion for patients with dementia. (58) For Vascular Surgery chatbots, the opportunities are endless and could be used to motivate claudicants to participate in walking regimens, answer, or triage medical questions, and help track symptoms of vascular disease, among other possibilities. It is important, however, that clinical use of chatbots be closely supervised by licensed healthcare practitioners to avoid patient harm by way of false or misleading information, and that AI based technologies do not completely replace patient-practitioner relationships.(54)

Pitfalls and Considerations

AI-based technologies offer exciting new opportunities to optimize patient care, however the incorporation of new technology into medical research and clinical invariably introduces several challenges that need to be addressed. The development of machine learning and deep learning algorithms requires large amounts of relatively homogenous, high-quality data for optimal performance. Electronic health records and the amount, type, and quality of patient data that are collected are widely variable from institution to institution, which makes the development of larger multicenter databases challenging. Training samples derived from patient samples collected from single-institution databases are typically too small to capture every possible variation that exists, which leads to data bias and may ultimately lead to worsening of healthcare bias and disparities. For example, photographic image-based algorithms that are trained using predominately light-skinned patients may lead to errors when applied to darker-skinned patients; diagnostic or treatment algorithms trained using data from predominately male patients may misdiagnose or misrepresent life threatening conditions in women.(59) As such, AI will inherently perform better for majority populations. Collection of larger, more diverse datasets across many institutions can help mitigate such data bias. Several projects, such as OHDSI or PCORnet, have been initiated in an attempt to collect longitudinal patient data from a diverse range of institutions across the country. (4, 60)

With the increasing number of AI-based models being introduced in the healthcare field comes a growing concern regarding patient safety and security. HIPAA regulations must still be applied to protect sensitive patient information, with specific attention paid to the sharing of health data among the aforementioned multicenter datasets. As with any new medical device or technology, tools based on AI should undergo rigorous testing to ensure patient safety and efficacy, and to mitigate harmful unintended effects.(61) Lastly, a security risk posed by AI and other computer based technologies is the risk of adversarial attack, meaning the creation of confusing or incorrect datasets which lead to suboptimal or incorrect decision making.(4, 62) The ever present threat to effective use of AI is its misuse by hackers to develop algorithms which put patient safety and security at risk.(63) Adversarial methods may lead to medical fraud or patient harm, among other things, and remain a significant challenge faced by developers of AI based medical technologies.(62)

Another challenge underlying machine learning, and especially deep learning, models is model interpretability. While well-trained algorithms are able to achieve high levels of precision and accuracy, the methods by which an input generates a given output may be obscured, giving more of a black-box model. Though it has been previously demonstrated that users’ trust in an AI model is similar between black-box and transparent models, healthcare may be one field in which transparency is necessary to improve decision making and integration into the clinical workflow.(64) Additionally, model transparency may help to justify model generalizability and to achieve better model performance.(4)

At its 2019 symposium on the Future of Digital Health Systems, the World Health Organization urged the importance of building and maintaining trust of AI based tools in the healthcare domain as well as the necessity of a strong ethical framework when developing AI-based technologies.(65) While ethics and values may be variable between different groups, the four principles of medical ethics, autonomy, beneficence, non-maleficence, and justice are ubiquitous and largely agreed upon regardless of background.(66) These principles, however, govern how humans are responsible for other human beings and therefore are less clear when applied to a non-human system.(67) AI carries with it the potential to inadvertently interfere with these principles as well as the fundamental rights and freedoms of mankind, and the developers, regulators, and users of AI technologies must be sure they are compatible with the ethical considerations and moral values of human beings. (67)

Conclusion

Artificial Intelligence offers novel and exciting tools to transform healthcare, and specifically Vascular Surgery, in ways that may have been previously considered science fiction. With the development of AI comes the promise of expedient, efficient, and overall improved patient care as well as the facilitation of significant advances in healthcare research. As with all new technologies, there are a number of potential pitfalls and other issues that must be considered by the developers, users, and regulators of AI based technology in order to ensure the safe and effective integration of these tools into practice.