Artificial Intelligence in Clinical Workflow Processes in Vascular Surgery and Beyond

Abstract

In the past decade, artificial intelligence (AI)-based applications have exploded in healthcare. In cardiovascular disease and vascular surgery specifically, AI tools such as machine learning, natural language processing, and deep neural networks have been applied to automatically detect still underdiagnosed diseases, such as peripheral artery disease, abdominal aortic aneurysms, and atherosclerotic cardiovascular disease. In addition to disease detection and risk stratification, AI has been used to identify guideline-concordant statin therapy use and reasons for nonuse, which has important implications for population-based cardiovascular disease health. While many studies highlight the potential applications of AI, few address true clinical workflow implementation of available AI-based tools. Specific examples such as determination of optimal statin treatment based on individual patient risk factors and enhancement of intraoperative fluoroscopy and ultrasound imaging demonstrate the potential promise of AI integration into clinical workflow. Many challenges to AI implementation in healthcare remain, including data interoperability, model bias and generalizability, prospective evaluation, privacy and security, and regulation. Multidisciplinary and multi-institutional collaboration as well as adopting a framework for integration will be critical for the successful implementation of AI tools into clinical practice.

1. Introduction

The integration of artificial intelligence (AI) into clinical workflow requires operationalization, performance monitoring, and quality control.¹ In this narrative review, we consider the use and implications of AI in clinical workflow processes within vascular surgery and cardiovascular disease more broadly. Artificial Intelligence is a broad term that encompasses many subfields, such as machine learning (ML), which includes deep learning, natural language processing (NLP), and computer vision. Machine learning refers to the capability of machines to learn and make predictions based on pattern recognition and is most useful for recognizing subtle trends, patterns, or associations in large datasets that a human analyst may not perceive. Deep learning is a type of ML that utilizes artificial neural networks (inspired by human biology) to process many layers of data and extract higher-level, more complex patterns. NLP refers to a computer’s ability to process human (i.e., natural) language in the form of text and is critical for analyzing large swaths of content such as electronic medical records and specifically provider narrative text notes, or the radiologist’s interpretation report on an image. Computer vision is the ability of a machine to process images and videos. A review of current applications of AI in vascular surgery was previously published in Seminars in 2021 and reported on its potential in diagnosis, perioperative risk stratification, and predicting outcomes.² Herein, we focus on literature illustrating pragmatic examples of AI’s translation into clinical workflows. We discuss the use of AI in the detection of peripheral artery disease (PAD), abdominal aortic aneurysm (AAA) disease, and atherosclerotic cardiovascular disease (ASCVD), including statin non-use. In addition, the usability, advantages, potential challenges, and disadvantages of AI in clinical workflows are reviewed. Finally, practical aspects of implementation of AI integration in clinical workflow processes are considered.

2. Methods: Literature search strategy

We conducted a PubMed literature search using the following keywords to identify potential articles on the topic of artificial intelligence in clinical workflow processes in vascular surgery and cardiovascular disease:

• Vascular clinical workflow: “artificial intelligence AND clinical workflow and vascular surgery”

• Peripheral artery disease: “artificial intelligence AND clinical workflow AND peripheral artery disease”

• Abdominal aortic aneurysm: “artificial intelligence AND clinical workflow AND abdominal aortic aneurysm”

• Atherosclerotic cardiovascular disease: “artificial intelligence AND clinical workflow AND atherosclerotic cardiovascular disease”

The articles were then reviewed and screened based on relevance. Of the relevant selected publications, their references were reviewed and screened for other potentially relevant articles to include. There was no restriction on publication year. Only articles in English were included in this review. The last search date was March 6, 2023. Expert input was also solicited for relevant articles. Summaries of selected references are provided in Table I and Table II.

Table I.

Summary of references investigating the use of artificial intelligence in the study of peripheral artery disease (PAD), abdominal aortic aneurysms (AAA), and/or atherosclerotic cardiovascular disease (ASCVD).

Disease	Reference	Study Aim	Data Source	Methods	Results
PAD	Ghanzouri et al (2022)	Detecting undiagnosed PAD	Institutional electronic medical record	Compared LASSO, random-forest, recurrent neural networks with long short-term memory, and logistic regression-based models	Deep learning models outperformed machine learning and logistic regression-based models.
PAD	Ross et al (2019)	Detecting major cardiac or cerebrovascular events in PAD patients	Multi-institutional electronic medical records	Compared penalized linear regression and random forest-based models	Machine learning models can detect which PAD patients will develop major cardiac or cerebrovascular events with AUC 0.81.
PAD	Ross et al (2016)	Detecting PAD and all-cause mortality (separate endpoints)	Genetic Determinants of Peripheral Arterial Disease (GenePAD) Study data	Compared elastic net, random forest, and stepwise linear regression	Machine learning models outperformed stepwise logistic regression in both PAD detection and all-cause mortality risk estimation.
AAA	Lareyre et al (2022)	Measuring AAA maximal diameter from CT scans	Institutional imaging database	Developed fully automated deep learning algorithm for segmentation and compared with manual segmentation	Deep learning algorithm for AAA max diameter correlated with the human measurements, had improved accuracy, and faster computational times.
AAA	Golla et al (2021)	Detecting AAA in CT scans	Institutional imaging database	Compared deep-learning networks (3D ResNet, VGG-16, AlexNet)	3D ResNet outperformed VGG-16 and AlexNet in classifying AAA from CT scans.
AAA	Caradu et al (2021)	Detecting AAA characteristics from CT scans	Institutional imaging database	Compared proprietary PREAVAorta software which utilizes convolutional neural networks to manual segmentation by vascular surgeons	AAA metrics generated by deep-learning based PREAVAorta software demonstrated over 90% similarity to manual measurements by vascular surgeons.
AAA	Hahn et al (2020)	Identify regions of AAA, AAA characteristics, and diagnose endoleak after EVAR	Institutional imaging database	Developed hybridized deep learning model based on RetinaNet, ResNet-50, and 3D-UNet convolutional neural networks	Deep learning model accurately identified an area containing AAA in 99% of cases. The best model of endoleak detection performed with an AUC 0.94.
AAA	Talebi et al (2020)	Diagnose endoleak after EVAR	Institutional imaging database	Compared a U-Net based convolutional neural network to blinded radiologists	Deep learning model demonstrated similar performance in endoleak diagnosis compared to radiologists.
AAA	Maurel et al (2018)	Determine whether image overlay in complex EVAR impacts radiation dose	Prospective single-institution study	Compared proprietary Cydar software for automated intraoperative imaging overlay to procedures without Cydar	Use of Cydar automated overlay was associated with reduced dose area product, air kerma, and number of digital subtraction angiography runs during complex EVAR.
AAA	Rolls et al (2016)	Determining accuracy of image overlay using hardware versus Cydar-based methods	Prospective single-institution study	Compared proprietary Cydar software for automated intraoperative imaging overlay to overlay based on bony landmarks and hardware	Cydar-based imaging overlay had a lower median error (higher accuracy) compared to landmark-based overlay.
ASCVD	Sarraju et al (2022)	Identify optimal statin intensity for primary ASCVD prevention	Institutional electronic medical record	Compared weighted-K-nearest neighbor regression to pooled cohort equations	Model-generated statin recommendations were associated with greater LDL-C reduction in matched historical patients.
ASCVD	Sarraju et al (2022)	Identifying presence and reasons for statin non-use in patients with cardiovascular disease	Institutional electronic medical record	Developed deep learning model using natural language processing (Clinical BERT) and assessed performance with ground truth from manual chart review	Clinical BERT identifies statin non-use (AUC 0.94) and reasons for non-use (AUC 0.88).
ASCVD	Gobbel et al (2022)	Identifying reasons for non-use of high-intensity statins in patients with atherosclerotic cardiovascular disease	Veterans Affairs system data warehouse	Adapted Canary NLP tool to extract statin use from unstructured data and compared performance to query methods using structured data alone	Supplementing analysis of structured and unstructured data with Canary NLP tool led to improved sensitivity and AUC for detecting reasons for statin non-use.
ASCVD	Olender et al (2021)	Integrate AI-generated vessel morphology with real vascular images	Not mentioned	Developed synthetic images using a conditional generative adversarial network to add intravascular morphology to ultrasound images	AI-generated synthetic images may be of clinical utility.
ASCVD	Eng et al (2021)	Coronary calcium screening and scoring from CT scans	Institutional medical record and cohort from Multi-Ethnic Study of Atherosclerosis (MESA)	Assessed performance of deep learning model with ground truth derived from a gated scoring model	Deep learning model had sensitivity and positive predictive value greater than 80% in detecting patients with coronary artery calcium.
ASCVD	Ward et al (2020)	Predicting ASCVD risk in multi-ethnic population	Institutional medical record	Compared random forest, gradient boosted machines, extreme gradient boosted models, logistic regression (standard and lasso-penalized) to the ASCVD Risk Calculator	Machine learning models achieved comparable or improved performance in ASCVD risk prediction in ethnic subgroups.
ASCVD	Flores et al (2020)	Identifying clinically significant coronary artery disease subgroups	Genetic Determinants of Peripheral Arterial Disease (GenePAD) Study data	Compared generalized low rank modeling with K-means clustering to ASCVD Risk Calculator	Subgroups developed by unsupervised machine learning methods were more informative for risk of myocardial infarction, stroke, and mortality.

Table II.

Summary of references reporting implementation-based outcomes or generating insights regarding clinical workflows

Disease	Reference	Study Aim	Methods	Results
PAD	Ghanzouri et al (2022)	Determine usability, acceptability, and accessibility of an EHR-based dashboard for PAD detection	Semi-structured interviews of cardiovascular and primary care physicians including think-aloud dashboard navigation	Most physicians are receptive to using machine learning for PAD detection.
AAA	Maurel et al (2018)	Determine utility of proprietary Cydar imaging overlay technology during complex EVAR	Survey of Cydar users at single institution	95% of users thought Cydar improved ease of performing complex EVAR.
AAA	Rolls et al (2016)	Determine usability of Cydar imaging overlay versus hardware-based overlay during EVAR	Binary failure rating based on whether a useful overlay map was generated	0% of Cydar-assisted cases and 33% of hardware-based overlay cases resulted in failure.
ASCVD	Sandhu et al (2023)	Determine effect of implementing a deep learning algorithm for coronary artery calcium detection on statin prescription	Prospectively randomized patients and physicians to be notified of coronary artery calcium on CT scan as detected by deep learning algorithm	Deep learning model integration was associated with increased statin prescription and coronary artery disease testing.
All	Li et al (2020)	Propose a workflow for AI delivery science	Advocate a structured, multidisciplinary process for creating, implementing, and evaluating AI-based healthcare innovations
All	Lyell et al (2017)	Identify factors associated with automation bias (overreliance on decision support)	Systematic review of 890 screened papers with 6 selected for analysis	Automation bias is associated with the degree of cognitive load in decision making.

3. Clinical use of AI in detection of PAD, AAA, statin nonuse in atherosclerotic cardiovascular disease

3.1. Detection and outcomes of peripheral arterial disease

PAD has been underdiagnosed due to inconsistent screening guidelines, low patient and provider awareness, and variable symptom presentation.^3,4 There is a need for improved detection of PAD amongst individuals with known risk factors, such as hypertension, hypercholesterolemia, diabetes, smoking, black race, low socioeconomic status, etc. to facilitate early intervention with behavior modification and medication. For instance, in a primary care screening study of patients aged >70 years or >59 years with diabetes or smoking history, 55% of patients had undiagnosed PAD based on ABI measurement.³ Previously, Flores and colleagues introduced key terms and concepts within machine learning and AI with regards to studying PAD.⁴ Here, we review the literature on how AI can be used to detect PAD and predict clinical outcomes after diagnosis.

Lareyre et al., summarized the current applications of AI in patients with PAD, including NLP, computer vision, and ML tools to improve screening, diagnosis, preoperative planning, and clinical workflows.⁵ NLP and ML have been used to improve screening, diagnosis, and disease severity classification. Computer vision was used on Doppler ultrasound and computed tomography angiography (CTA) exams to automatically detect and characterize arterial lesions. ML was used to build predictive models for patient mortality and create real-time models that could supplement clinical decision-making.

Ghanzouri et al. created an automated tool for detecting PAD from electronic health record (EHR) data by developing a deep learning model in a patient cohort of 3,168 cases and 16,863 controls.⁶ Performance and usability were tested and demonstrated that their deep learning model outperformed more traditional risk-factor based ML models such as random forest and logistic regression (average AUCs 0.96, 0.91 and 0.81, respectively, P < 0.0001). This study also assessed usability by physician interviews and will be discussed in Section 5 on implementation, below.

Ross et al. trained ML algorithms on clinical, demographic, imaging, and genomic data to build models to identify PAD and predict mortality in a cohort of 1,755 patients undergoing elective coronary angiography.⁷ These machine learning models outperformed traditional stepwise regression models for both PAD classification (AUC 0.87 versus 0.76, respectively, P=0.03) and mortality risk (AUC 0.76 versus 0.65, respectively, P=0.10). Further studies on model testing, automation, and prospective validation are needed prior to integration into clinical workflow.

In a subsequent study, Ross and colleagues utilized EHR data from 7,686 patients to build a common data predictive model for PAD patients who were most likely to develop major cardiac and cerebrovascular events.⁸ With nearly 1,000 variable inputs, their model accurately predicted adverse events with an AUC 0.81 (95% CI, 0.80–0.83), demonstrating the potential for automated EHR risk-stratification of cardiovascular disease outcomes. Their novel approach utilized a common data model, which allows for observational data to be combined from different institutions into a unified database structure, thus facilitating external cross-validation and improving model generalizability.⁹

3.2. Detection and management of abdominal aortic aneurysm disease

Many patients with abdominal aortic aneurysms (AAA) are undiagnosed prior to aortic rupture, which is associated with increased morbidity and mortality in >80% of cases.¹⁰ Moreover, inter-observer agreement for manually performed cross-sectional aortic measurements is poor. One study reported that 87% of manually performed interobserver comparisons exceeded the clinically accepted range of ±5 mm in diameter; semi-automated methods resulted in highest intra- and inter-observer agreement.¹¹ This variability has important implications for clinical practice as patients who are referred for AAA repair based on a diameter threshold can vary from 5–24% depending on the radiologist and method used.¹² Thus, AI has important applications for AAA detection, screening, measurement, as well as intraoperative management and postoperative surveillance for AAA, as demonstrated by the following studies.

Lareyre et al. developed a fully automated deep learning algorithm to measure maximal AAA diameter and compared its performance to standard manual segmentation completed by two human operators.¹³ Pre-operative CTAs from 34 patients with infrarenal AAA were included. The automatic DL method computed maximal AAA diameter in less than 30 seconds per scan with median absolute error 0.8 mm (0.5–4.1mm) and correlation coefficient 0.91 when compared with manual segmentation (P< 0.001). Though their results are promising regarding standardization and automation of AAA measurements, validation in larger datasets and testing the algorithms accuracy in patients with prior endografts are cited as important next steps prior to implementation into clinical decision-making workflows.

Golla et al. created a deep learning-based automated screening tool for small AAA using 187 heterogenous CT scans and three deep convolutional neural networks (CNN; ResNet, VGG-16 and AlexNet) adapted for 3D classification.¹⁴ Presence of aneurysmal disease was defined as an >50% focal increase in aortic diameter. The 3D ResNet algorithm was the best performer with an AUC of 0.93 on the training set and an AUC of 0.97 on a test set of 107 new scans, demonstrating the model’s robustness and high performance. The authors acknowledge the critical next steps for successfully implementing this AI into clinical processes, which include further training and validation on larger data sets, integrating into standard workflow procedures, demonstrating robustness, reliability, and positive impact of early detection of small AAA on patient outcomes.

Caradu et al. utilized a proprietary, fully-automated segmentation software (PRAEVAorta, developed by Nurea) to detect aortic lumen and infrarenal AAA characteristics such as thrombus and compared measurement results and processing times between the fully automatic software and senior vascular surgeon (Figure 1).¹⁵ Their results demonstrated high correlation for aortic volume, surface, and diameter measurements between the fully automated and manual segmentation (Pearson’s correlation coefficient >0.90) and a significant reduction in the segmentation time from 27 seconds-4 minutes per patient for the automated software vs 5 minutes to 80 minutes per patient for the surgeon/manually corrected method. While still proprietary and not yet integrated into their clinical workflow, this potential time saved in efficient and fully automated analysis of infrarenal AAAs may have significant implications for clinical practice.

Figure 1.

Example images of segmentation of the aortic lumen (red) and intraluminal thrombus (green) using PRAEVAorta (Nurea), an automatic segmentation software. Cross-sectional CTA images from a patient with an infrarenal abdominal aortic aneurysm are shown at three segmentation levels (rows): proximal aortic neck, maximum aortic diameter, and aortic bifurcation. For each level of segmentation (columns), the following are shown: original CTA image, fully automatic segmentation (generated by the software), senior and junior surgeon manually corrected segmentation. In the comparison, the results from the fully automated segmentation produced by PRAEVAorta are compared with the senior surgeon’s manual corrected segmentation. Best visualized in the 3D comparison rendering is the lumen common to manual and automatic segmentation (green), the thrombus common to manual and automatic segmentation (yellow), the false negatives (red), and the false positives (blue).

Reproduced with permission from Caradu C, Spampinato B, Vrancianu AM, Bérard X, Ducasse E. Fully automatic volume segmentation of infrarenal abdominal aortic aneurysm computed tomography images with deep learning approaches versus physician controlled manual segmentation. J Vasc Surg. 2021;74(1):246–256.e6.

CTA, Computed tomography angiography;

Intraoperatively, AI software (e.g., Cydar Medical, United Kingdom) is being utilized during endovascular aneurysm repair to generate fully automated 3D fusion map overlays, which are projected onto live 2D fluoroscopic images and are able to continuously adjust in real-time, accounting for patient movement and anatomic deformations due to wires, catheters, and other devices.⁴ These AI systems, which have been used for standard and fenestrated EVAR, are cloud based, integrated into existing clinical and operative workflows without requiring additional equipment or hardware installation, and demonstrated reduced radiation dose, number of DSA runs, and fluoroscopic times.^16,17 Since PRAEVAorta and Cydar are proprietary software, their underlying algorithms are not accessible for review, however, the above descriptions are based on each company’s claims of how AI is being utilized. Additional companies such as Viz Aortic (Viz.ai) and Aidoc have algorithms for detection of aortic dissections and aneurysms, however, no peer-reviewed studies on implementation of their software for aortic disease have yet been published.

Finally, there is an emerging role for machine learning in long-term surveillance after AAA repair such a post-EVAR endoleak detection. Hahn et al. built deep learning algorithms using the RetinaNet CNN to detect endoleak on 334 postoperative CTA images (N=191 patients) with high model performance (AUC=0.94).¹⁸ The authors highlight the need for external validation on larger datasets of CTAs and prospective study are needed before integration of this AI algorithm into clinical workflow. Similarly, Talebi et al. evaluated the utility of a deep neural network to detect endoleak in CTAs of post-EVAR patients using a novel data augmentation method (AUC=0.99) and found that the ML algorithm had similar performance to diagnose endoleaks as the cardiovascular specialist radiologist and outperformed the generalist radiologist on direct comparison.¹⁹ These studies demonstrate that AI models can improve efficiency, accuracy, and reliability between readers in AAA detection and post-EVAR surveillance, creating potential for AI to assist physicians in clinical practice. A significant challenge for AI remains in consistent and accurate determination of true aortic diameter versus mural thrombus or adherent structures, and so its interpretations must be validated by expert providers before direct applications to patient care.

3.3. Detection and management of atherosclerotic cardiovascular disease

Atherosclerotic cardiovascular disease (ASCVD) is a leading cause of death worldwide. AI-based tools and ML approaches can automate detection of high-risk subgroups, predict future cardiovascular event risk, identify rates and reasons of statin nonuse, and opportunistically calculate coronary artery calcium scores.

Using EHR data, Ward et al. developed ML models to predict the risk of developing ASCVD using a cohort of 262,923 multi-ethnic patients from Northern California.²⁰ Their findings underscore the utility of ML models for primary prevention of ASCVD in diverse, real-world populations, who may not be included in the pooled cohort equations used to estimate 10-year ASCD risk and recommended by current practice guidelines.²¹ Similarly, Sarraju et al. trained EHR-based ML models to estimate 5-year CVD event risk in a multi-ethnic cohort of 32,192 patients with an already known diagnosis of CVD (including ASCVD) yielding an AUC of 0.70.²² Nontraditional variables such as education level and number of primary care visits were predictive, highlighting the ability of ML to also improve risk stratification for secondary CVD prevention.

In addition to primary and secondary prevention of ASCVD, ML methods have been applied to identify high-risk coronary artery disease (CAD) subgroups of patients with differing clinical outcomes. Flores et al. applied unsupervised clustering techniques to 155 clinical, sociodemographic, biologic, and genetic variables from 1329 individuals in the Genetic Determinants of Peripheral Arterial Disease prospective study.²³ They identified four unique patient subgroups with differential rates of all-cause mortality and major adverse cardiovascular and cerebrovascular events (Figure 2). Compared to pooled cohort equations, which were first developed by the 2013 AHA/ACC Guidelines workgroup to estimate 10-year ASCVD risk among patients without prior cardiovascular disease aged 40 to 79 years,²⁴ subgroup membership derived from their AI model better informed risk of myocardial infarction, stroke, and mortality.

Figure 2.

Schematic of four AI-derived clusters of patients with CAD and significant features. Reproduced with permission from Flores AM, Schuler A, Eberhard AV, Olin JW, Cooke JP, Leeper NJ, Shah NH, Ross EG. Unsupervised Learning for Automated Detection of Coronary Artery Disease Subgroups. J Am Heart Assoc. 2021 Dec 7;10(23):e021976.

ABI, ankle‐brachial index; BMI, body mass index; CAD, coronary artery disease; CHF, congestive heart failure; CVA, cerebrovascular accident; LDL, low‐density lipoprotein; MACCE, major adverse cardiovascular and cerebrovascular events; MI, myocardial infarction; and PAD, peripheral artery disease.

While ML models can improve identification and risk stratification, AI can also identify trends in statin use amongst ASCVD patients. Sarraju et al. developed deep learning-based natural language processing (NLP) models on a cohort of 56,530 ASCVD patients to identify incidence and reasons for statin nonuse from unstructured EHR data.²² They found that 38% of patients did not have guideline-concordant statin prescriptions and did not have statins listed as an allergy. Of these, only 18% had a documented discussion about statin usage. Their best-performing classifier was able to identify statin nonuse with an AUC of 0.94 (95% CI 0.93–0.96) and reasons for nonuse (e.g., side effects, patient preference, or provider guideline-nonadherence) with an AUC of 0.88 (95% CI 0.86–0.91). This deep learning AI model could facilitate patient and provider education via decision support through EHR-based alerts or reminders and provide hospital and health systems a method for identifying and remedying ASCVD treatment gaps and disparities.

Using similar ML methodology, Gobbel et al. used NLP tools to extract data from unstructured EHR text in a cohort of randomly selected 1152 VA healthcare system patients, who were treated for ASCVD but not receiving high-intensity statins.²⁵ They found that 47% of notes documented at least one reason for not prescribing statins with side effects and general intolerance most frequently cited. The addition of NLP-extracted unstructured data to queries of structured data significantly increased sensitivity, negative predictive value, and AUC for detecting non-use compared to structured data only (p<.05), suggesting the utility of ML tools to customize and implement health system-wide initiates to increase appropriate statin use in ASCVD patients.

Finally, AI can be leveraged for opportunistic screening of cardiovascular disease, the leading cause of death in the United States.²⁶ Eng and Chute et al. developed deep learning models that can quantify coronary artery calcium (CAC) scoring using routine non-gated chest CT exams to detect CAD and inform intervention.²⁷ For patients with CAC ≥ 100, the recommended threshold for statin therapy, their model had sensitivity ranging from 71–94% and positive predictive values 88–100% when compared to reference gated coronary CT scans. These AI models provide a more cost-effective alternative to dedicated coronary CT scans, which would allow screening of many more patients undergoing routine chest imaging and potentially resulting in earlier identification and treatment of those at risk.

4. Usability and advantages of AI integration in clinical workflow

The literature on AI in vascular surgery and cardiovascular disease has primarily focused on improving automated disease detection and risk stratification, but few studies focus on current use cases for AI integration into clinical workflow for management of disease. In this section, we review select examples for optimal statin therapy, increasing statin use, and improved intravascular ultrasound imaging and discuss usability/advantages of AI integration in the clinical setting more broadly.

In the randomized NOTIFY-1 study, Sandhu et al. used automated ML methods to identify patients at risk for ASCVD based on CAC from chest CTs and coupled this risk stratification with a clinical intervention.²⁸ After applying a validated deep learning algorithm paired with radiologist review to screen of over 2000 patients for CAC on prior non-gated chest CTs, patients with incidental CAC (N=424, 20%) were randomized to notification of their PCP vs standard care, where notification included an image of CAC and appropriate guidelines on statin therapy. At 6 months follow-up, the rate of statin prescription was 51% in the notification group vs 7% in standard care (p<.001) as well as increased CAD testing in the notification group (15% vs 2%, p=.008). The authors acknowledge that further prospective research is required to determine the effect on ASCVD clinical outcomes, such as myocardial infarction, stroke, and death.

Sarraju et al. used ML models on aggregate EHR data from over 50,000 patients in Northern California aged 40–79 with no prior diagnosis of ASCVD or known statin use to determine a personalized, optimal statin therapy recommendation associated with highest LDL reduction after 1 year for primary cardiovascular event prevention.²⁹ Candidate patients were compared to similar historical controls identified in the EHR via Weighted-K-nearest-neighbor (wKNN) regression models. Four statin treatment options were modeled including none, low-intensity, moderate-intensity, and high-intensity. Interestingly, the ML model resulted in recommending low- or moderate-intensity statins in almost half the patient cohort, despite this being discordant with current therapy guidelines, which recommend high-intensity statins for those with highest ASCVD risk. These findings may be highlighting potential barriers to high-intensity statin therapy in some groups such as patient nonadherence or intolerance or provider preferences, which ML models allow us to identify and address at the patient, provider, and system level. This study may illustrate a potential use case more directly translatable to vascular surgery, whereby ML models could be used to determine a personalized anti-platelet and/or anti-coagulation regimen for PAD patients both before and after surgical intervention with therapy options including aspirin, clopidogrel, low-dose rivaroxaban, warfarin, full-dose DOAC, or some combination thereof.

Current use cases for AI integration into clinical workflow also exist in ultrasound imaging. In the field of interventional cardiology, Olender et al. leveraged AI to synthesize multiple images using intravascular ultrasound (IVUS) images and generative neural networks to augment the operator’s visual workflow during a procedure—in this case, producing synthetic images in the form of optical coherence tomography, a type of intravascular imaging used for the retina (Figure 3).³⁰ These AI-generated synthetic images capture increased vascular lumen heterogeneity, decrease noise, and can facilitate education and training, however, the efficacy and performance of this tool still must be validated before it can be integrated fully into clinical workflow as well as concerns about image manipulation and security addressed. Though, there would be many practical applications to vascular surgery cases that utilize IVUS already in their standard workflow to potentially incorporate these AI-augmented images.

Figure 3.

AI-generated synthetic images from standard intravascular ultrasound images utilizing conditional generative adversarial networks (cGANs) can improve cardiovascular imaging and augment the operator’s visual clinical workflow.

Reproduced with permission from Olender ML, de la Torre Hernández JM, Athanasiou LS, Nezami FR, Edelman ER. Artificial intelligence to generate medical images: augmenting the cardiologist’s visual clinical workflow. Eur Heart J Digit Health. 2021;2(3):539–544.

In general, advantages of AI integration into clinical workflows include enhanced and automated disease detection and patient risk stratification, identification of guideline-concordant and discordant therapy, recommendations for personalized treatment regimens, and improved imaging quality to increase efficiency intraoperatively. While many of the above studies demonstrate proof-of-concept, prospective validation with regards to clinically significant and implementation-oriented endpoints is needed prior to full integration.

5. Challenges and disadvantages of AI implementation in clinical workflow

There are many challenges still facing AI implementation in clinical systems. Key barriers include data sharing and interoperability, algorithm/model bias and generalizability, prospective evaluation of clinical utility, implementation, privacy and security, and regulation.^4,31

Data sharing and algorithm generalizability remain key hurdles for clinical integration as ML algorithms are typically developed on single-institution datasets from a limited number of patient records or radiographic images.¹⁴ In the studies reviewed here, a shift from using single-institution data toward multi-institution and regional EHR data is promising.^8,20,22,29

Prospective validation of AI models in real-world settings is essential. For example, in EHR-based detection of PAD, increased accuracy of diagnosis may be associated with increased guideline-concordant therapy and therefore potentially enhanced clinical outcomes.⁴ However these associations require future studies to demonstrate the robustness of these models and gain trust of providers and patients.³²

Specific to implementation, the field of AI delivery science has emerged to provide a framework to evaluate how ML models fit within a health system prior to model implementation and integration (Figure 4).³³ As outlined by Li et al., this process harnesses design thinking, process improvement, and implementation science to create a standardized workflow for AI software integration while unveiling logistical roadblocks and improving understanding of how care team members would use AI information in real-time clinical settings.³³ Additional challenges to consider include resource availability to handle increased patient data and volume of information⁴ and training to ensure appropriate interpretation and interventions based on automated AI outputs, which is critical to avoid overreliance on decision-support tools.³⁴

Figure 4.

Process framework integration of AI in the healthcare system: problem scoping, intervention design, implementation, and evaluation. Multidisciplinary expertise is needed from (1) user experience, (2) data science, (3) healthcare operations, (4) clinical informatics, (5) evaluation, and (6) ethics assessment.

Reproduced with permission from Li RC, Asch SM and Shah NH. Developing a delivery science for artificial intelligence in healthcare. NPJ Digit Med. 2020;3:107.

Finally, the ethical issues of algorithm bias, security, privacy, and regulation require all stakeholders to design AI-based tools with maximal transparency and dependability.³⁵ Understandably, there are concerns about patient privacy and security when using patient data to train AI models. These include the challenges of rare diagnoses or the inclusion of enough unique patient characteristics that may make it possible to re-identify patients. The FDA published its “Artificial Intelligence and Machine Learning (AI/ML) Software as a Medical Device Action Plan” in January 2021, which aims to provide a framework to AI and ML companies while ensuring the safety and effectiveness of software as a medical device. How this will be implemented in practice continues to evolve. Similar society-led initiatives in vascular surgery will be necessary to ensure the equitable and just use of AI for our vascular surgery patients.

6. Practical aspects of implementation and potential solutions

6.1. Common data models and cross-institution collaboration

Common data models allow for cross-institution research and model development, which offer a solution to the challenge of data interoperability and model generalizability. Interoperability in ML is commonly defined as ability for two or more systems to communicate effectively, such as through data exchange. Its feasibility is demonstrated by Ross et al. in their multi-institution study of predicting cardiovascular events in PAD patients derived from two EHR systems using the Observational Medical Outcomes Partnership (OMOP) common data model.⁸ The authors propose that cross-institution technology development can facilitate more efficient and widespread adoption of AI technologies.

6.2. Performance and usability testing

Usability testing involves direct observation of the end user of AI technology to increase provider usage and engagement. For example, Ghanzouri et al. found that 75% of physicians were open to utilizing an EHR-based automated PAD detection tool but surveyed providers also highlighted important feasibility aspects such as direct integration of risk assessment tools in the HER and optimizing displays and dashboards within the EHR to reduce the need for more clicks.⁶ These key insights can be derived from “think-aloud” techniques, whereby researchers directly observe users engaging with the AI-derived technology to assess how attention and cognitive processing are directed.³⁶

6.3. Stakeholder engagement and collaboration

Finally, early stakeholder engagement is of critical importance to successful AI deployment in clinical workflows. Collaboration between data scientists (e.g., information technology, bioinformatics), physicians and other clinical providers, and implementation scientists can help build technologies that are scalable, sustainable, and considerate of the end-user.^4,33

7. Discussion

We review the relevant literature on clinical use of AI in peripheral arterial, abdominal aortic aneurysm, and cardiovascular disease, current use cases for integration in clinical workflow, as well as remaining challenges and practical aspects of implementation. AI-based tools have been developed for automated disease detection, risk stratification, identification of patients receiving guideline-directed therapy, and enhancement of intraoperative imaging. ML and NLP models have demonstrated potential for automated detection of PAD, AAA, and ASCVD from large EHR datasets including unstructured text and CTA imaging. However, how these AI processes impact clinical outcomes remains unknown. Perhaps most promising is the result of the NOTIFY-1 study, which demonstrated how deep learning models can improve adherence to guidelines for ASCVD-related statin therapy. However, even this study used a surrogate marker of active statin prescription at 6 months and was not tied to a clinical endpoint of LDL or cardiovascular events. Thus, in our view, the next phase of AI based research in vascular surgery must address this gap and demonstrate clinical reliability and efficacy.

In addition to prospective evaluation of AI models tied to clinical outcomes, we have reviewed the many challenges and some solutions facing these new technologies prior to integration into clinical workflows including data interoperability, model generalizability and potential algorithm bias, patient privacy and security, implementation, and eventually regulation.

8. Conclusions

AI applications to vascular surgery and cardiovascular disease demonstrate potential for enhanced disease detection in previously unscreened patients or subgroups of the population, especially for PAD, AAA, and ASCVD. AI-derived models using machine learning, language processing, and deep neural networks show promise in diagnosis, personalized medical therapy regimens, preoperative planning, intraoperative assistance, surveillance, and predicting future adverse events. Future studies are needed to prospectively validate these models and demonstrate feasibility and acceptability of implementation before widespread adoption into clinical workflows. Successful implementation of AI tools into clinical practice will benefit from multi-institution collaboration, interoperable designs such as common data models, performance and end-user testing, and early stakeholder engagement harnessing diverse expertise across the health system.