Abstract
Objectives
The study evaluates the plausibility and applicability of prediction, pattern recognition and modelling of complications post-endovascular aneurysm repair (EVAR) by artificial intelligence for more accurate surveillance in practice.
Methods
A single-centre prospective data collection on (n = 250) EVAR cases with n = 26 preoperative attributes (factors) on endpoint of endoleak (types I–VI), occlusion, migration and mortality over a 13-year period was conducted. In addition to the traditional statistical analysis, data was subjected to machine learning algorithm through artificial neural network. The predictive accuracy (specificity and –1 sensitivity) on each endpoint is presented with percentage and receiver operative curve. The pattern recognition and model classification were conducted using discriminate analysis, decision tree, logistic regression, naive Bayes and support vector machines, and the best fit model was deployed for pattern recognition and modelling.
Results
The accuracy of the training, validation and predictive ability of artificial neural network in detection of endoleak type I was 95, 96 and 94%, type II (94, 83, 90 and 82%) and type III was 96, 94 and 96%, respectively. Endpoints are associated with increase in weights through predictive modeling that were not detected through statistical analytics. The overall accuracy of the model was >86%.
Conclusion
The study highlights the applicability, accuracy and reliability of artificial intelligence in the detection of adverse outcomes post-EVAR for an accurate surveillance stratification.
Keywords
Introduction
Evidence suggests that one in every five individuals post-endovascular aneurysm repair (EVAR) could face endoleak, sac expansion, device migration, limb occlusion and stenosis over the first five years of surveillance.1–4 EVAR is subjected to lifelong surveillance by different modalities such as duplex sonography, contrast enhanced ultrasound (CEUS), computed tomography angiogram (CTA) and magnetic resonance imaging (MRI). 5 , 6 To date, various surveillance protocols tailored to the local population and expertise have been formulated. Despite this, 50–90% of complications remain undetected, prevail beyond surveillance, present between surveillance time frames and attribute to symptoms or outcomes that demand acute intervention associated with significant mortality and morbidity. 7 , 8 Thus, subjecting every individual to repeated radiation and contrast, besides the financial burden might not be advocated. However, the detection of complications remains vital in longevity of EVAR and limitations of adversities. 9 , 10
This notion has escalated the demand for design of predictive tools for detection of complications or identification of high risk cohort. However, the majority of the predictive models are heterogeneous and are based on local demographics, comorbidities, risk stratification, aneurysm morphology and EVAR devices. 11 , 12 These data have been subjected to traditional statistical analytics and do not replicate or represent a wider population and possess a poor predictive ability of adversities. In other words, there is no one fit-for-all accurate model in the clinical practice. 11 , 12 Thus, the question of alternative approach in design of predictive model/s based on pre-operative data through an artificial intelligence (AI) has become more plausible. AI has the ability to examine complex interplay of attributes even with minor impact and assign weights to each factor that was once considered to be insignificant through traditional statistical analytics. 13 Therefore, the primary objective of this study for the first time in the literature is to evaluate the application of AI in the detection of adverse outcomes (endoleak, occlusion, migration). In secondary aim, the study examines the potential of the pattern recognition and modelling for each endpoint with a working model for the clinical practice for an accurate surveillance stratification.
Methods
A single-centre prospective data collection over a 13-year period on n = 250 patients undergoing EVAR was conducted. This included only those cases that had EVAR for non-ruptured infra-renal abdominal aneurysm with no adjuvant surgery for any other disease (iliac aneurysm/saccular aneurysm/mycotic aneurysm).
The dataset (factors) includes: patient demographics (age, gender, weight and height), comorbidities (diabetes mellitus (DM), renal failure (RF), hypertension (HTN), ischaemic heart disease (IHD), chronic obstructive pulmonary disease (COPD), cholesterol status and smoking history), American Society of Anaesthesiologist (ASA), preoperative haematological markers (sodium, potassium, creatinine, white cell count, albumen and haemoglobin), length of stay (LOS), aneurysm morphology (neck angle (>60° or <60°), neck diameter, neck length, common iliac size (left and right), endoleak (presence, type, sac expansion and sac reduction), limb occlusion, rupture and mortality. All comorbidities were defined in accordance to WHO definitions and The Royal colleague of Surgeons Charlson Score.
The primary end point of study was measured on the presence of endoleak from the onset of completion angiography immediately following EVAR or thereafter on surveillance. The path for every endoleak (sac expansion or reduction) was carefully monitored, and the outcome of each procedure was recorded (e.g. angioplasty, extension, aortic wrap, palmaz, endoleak cessation). This was also applicable to limb occlusion or stenosis either on presentation or its detection on surveillance examination (angioplasty, stenting, cross-over grafting). Mortality was defined as any death related to EVAR or its rupture and not from any concomitant disease as an established cause. Access to the data was granted by the research and clinical audit department of the Mid Essex Hospital Services NHS Trust through Clinical Audit Number CA19-024. The study does not report any experiments, alteration in practice and/or subjecting any individual to any new practice. The study was conducted according to the Helsinki Code of Ethical Principles, and data was anonymised prior subjecting it to AI and statistical analysis.
Surveillance policy
In our unit, we conduct on table completion angiogram at the end of every EVAR. This is followed up by a CTA at 1 and 12 months follow-up. Patient is assessed by duplex sonography at six months along with abdominal X-ray and yearly duplex sonography thereafter if no endoleak is detected. Upon detection of endoleak, patient is subjected to CEUS, and with evidence of sac expansion, a CTA is conducted. The minimum CTA slice thickness at our unit is 1 mm along with complete three-dimensional (3D) construction for detailed examination. The presence or lack of endoleak along with sac expansion, limb occlusion or stenosis was evaluated through a multidisciplinary team approach with two interventional vascular radiologists, vascular scientist, vascular nurse specialist and vascular surgeons.
Statistical analysis
Data on all variables were collected and subjected to descriptive analysis using statistical package for the social sciences (SPSS) version 23, IBM. All continuous variables were reported as median with their interquartile ranges (IQR) and percentages. The endpoints were recorded in binary format (presence of endoleak versus none), occlusion versus none, and for sub-group analysis, types of endoleaks were classified according to their types (types I–V). The data set was subjected to Pearson chi-squared test (χ 2 test) of statistical analysis to evaluate the differences in the expected and frequencies of the two categories with a contingency table. The test of probability (p value) was considered to be statistically significant if p value was <0.05. The initial analysis was regarding the presence of any type of endoleak versus none (Table 1). This was followed by difference in sub category of type II endoleak versus none with sac expansion versus none (Tables 2 and 3). Due to small number of occlusions and type III endoleak incidents, no meaningful statistical analysis was plausible due to type I error. However, the details are tabulated in Table 4.
Comparative analysis of endoleak versus no endoleak.
NS: no statistical significance.
Comparative analysis of endoleak type II versus endoleak type II with no sac expansion.
NS: no statistical significance.
Details of endoleak type I.
Details of occlusion group.
Artificial intelligence
Machine learning algorithm through artificial neural network (ANN) (classifier, modelling and predictive analytics) based on Bayesian theorem was designed using MATLAB and Statistics Toolbox (Release 2018 b, The MathWorks, Inc., Natick, MA, USA). ANN was comprised of an input of 26 preoperative factors (inputs) through a three layers’ (26 input, 10-hidden, 2-output) on binary outcome. The matrix defining 26 attributes of 241 individuals (26 × 241) was set on the target of two outputs (binary: yes/no) characteristics of 241 (2 × 241) matrix for each end point (endoleak types I, II and III and occlusion). Each column indicated a correct category with a one in either element. The data is assessed through backpropagation which is the method of calculating the “gradient” required for weights of each variable for ANN. The data is eventually iterated over one epoch (reference point on a scale from which time is measured), and this is validated at a range of 0–6 (6 being the highest level of validation) on a scale. The probability distribution of the ANN performance on training and validation data is assessed through cross-entropy along with error measurement (histogram). The training and the network are adjusted according to its error, and the validation sample was used to measure network generalisation and to halt training when generalisation stops improving. Finally, an individual matrix for training, validation, testing and overall outcome for ANN is conducted. The finding of each matrix is validated through the corresponding area under the curve (AUC) for best sensitivity and 1-specificity. The predictive accuracy (sensitivity and –1 specificity) of the model on each endpoint through ANN is presented with percentage and receiver operative curve (ROC) (Supplementary Figures 1 and 2).
The data is analysed in different ways: so it could represent the outcome, and if tested again or if there is misrepresentation, this is fed backwards (backpropagation) till an ideal representation (pattern recognition) is achieved before final output layer. The learning process continues as long as the network improves its performance on endpoints of validation and accuracy by crossing the input layers’ multiple times to train the ANN. For optimisation of parameters and avoid overfitting, the data is automatically divided into three sections of 70% for training, 15% for evaluation and the F-measure (combination of precision and recall) and 15% for validation. Upon completion of training on 70% of the data, the predictive model is evaluated and measured for its independent precision and recall with final validation on the remaining 15%. This is conducted automatically without human interference, and the testing data set does not affect training and provides an independent measure of network performance during and after training. Finally, the percentage of training, testing and validation is cross-examined for the evaluation of overfitting (similar percentage across all paths of training, testing validation) along with using principle component analysis (PCA) and five-fold cross-validation to reduce the dimensionability at 95% variance (Supplementary Figures 3 to 5).
The pattern recognition and model classification were conducted using discriminate analysis, decision tree, logistic regression, Naive Bayes and support vector machines (Classifier application) using MATLAB and Statistics Toolbox (Release 2018 b, The MathWorks, Inc., Natick, MA, USA). However, the best fit (the highest accuracy) was used for pattern recognition and model creation. Of all methods, Bayesian theorem demonstrated the highest accuracy. Further description is available in the Supplementary section of this manuscript.
Statistical analysis results
Complete data was available on n = 241 individuals. A total of n = 171 individuals showed no evidence of endoleak, whereas n = 70 individuals demonstrated endoleak of some kind. Type I endoleak was noted in 4.1%, with a time to event of seven years (IQR, 1–12 years). Type II endoleak was noted in 24% of the entire cohort of which 16.1% (n = 39) demonstrated sac reduction of 5 mm (IQR, 1–26 mm) over a five-year period (IQR, 1–12 years). The remaining type II (n = 19) 7.9% demonstrated sac expansion of 5 mm (IQR, 1–16 mm) over a seven-year period (IQR, 1–11 years). Type III endoleak occurred in 0.8% (n = 2) of the cohort, and there was no report of types IV or V endoleak. Ipsilateral limb occlusion occurred in 3.7% of the cohort over a time period of 10 years (IQR, 3–12 years) and stent migration in 0.4% (n = 1). The overall mortality from sac expansion and rupture was 1.3% over a time line of two years (IQR, 1–3 years). There was no statistically significant attribute (p > 0.05) in any comparative analysis. The overall outcome of each endpoint and their comparative outcome through traditional statistical means are presented in Tables 1 to 4.
Prediction and performance of AI
The predictive ability (accuracy) of the AI through ANN was highest in the Naive Bayes model on all end points (types I–V endoleak and occlusion). The overall accuracy of training, testing, validation and overall performance of ANN on detection of type I endoleak with n = 26 attributes was 95.3%, 88.9%, 96.3% and 94.5% (Figure 1), respectively. For type II endoleak, this was 83.2%, 71.9%, 90.3% and 82% (training, testing, validation and overall performance) (Figure 2), and finally, for type III and occlusion, this was 94.3%, 100%, 92.6% and 94.6%, respectively (Figure 3). However, this prediction was not plausible on the end point of mortality and stent migration as the numbers are too limited for an accurate prediction. The outcome demonstrates that each endpoint is predicted on the basis of different attributes with their respective weights. This simply indicates that each endpoint (endoleak types and occlusion) is influenced differently.

The confusion matrix of training, validation, testing and overall accuracy of the AI (ANN) in detection of endoleak type I.

The confusion matrix of training, validation, testing and overall accuracy of the AI (ANN) in detection of endoleak type II.

The confusion matrix of training, validation, testing and overall accuracy of the AI (ANN) in detection of type III and occlusion.
The overall accuracy of the model on detection of any adverse outcome was >86%. The effectiveness of the model was also evaluated by a life chart that examines the ratio between the results through the model (AI) and those obtained without the model in percentage with confidence (Figure 4).

The life chart demonstrating the effectiveness of the model (AI) examining the ratio between the results through AI and those obtained without the model.
Pattern recognition and modelling
The respective pattern recognition for each endpoint was different. The weight of each attribute (preoperative factor) was evaluated and presented in Figures 5 and 6. Each chart demonstrating how their respective increase or decrease could influence the outcome. Overall, the higher the value of the weight, more significant its impact on the outcome prediction and model.

Predictive model (radar chart) demonstrating the weights of each attribute (n = 26) on the endpoint of endoleak type I. Increase in each weight is associated with increase in occurrence of each endpoint. This predictive model could be fed with ongoing data and deep learning.

Predictive model (radar chart) demonstrating the weights of each attribute (n = 26) on the endpoint of endoleak type II. Increase in each weight is associated with increase in occurrence of each endpoint. This predictive model could be fed with ongoing data and deep learning.
Discussion
This study demonstrates that with only n = 26 preoperative attributes, AI has the proficiency of detecting adverse outcomes with an overall accuracy of 86% despite traditional statistical analytics, suggesting otherwise (Tables 1 to 4). The current model has the potential of achieving higher sensitivity and 1-specificity if increased attributes is available for ongoing machine learning and validation. In contrast, a good example of higher performance (90% accuracy) was noted in the detection of surgical site infection with n = 29 attributes. Similar outcomes were noted in early prostate cancer diagnosis (>95%) and detection of high versus low risk EVAR (90%).14–16 In recent years, AI application in the field of vascular surgery has been rapidly expanding. Current examples include: the detection and diagnosis of abdominal aortic aneurysms (AAA) (imaging), their ongoing sac expansion (imagine), rupture and EVAR planning. All these studies have shown significant precision and accuracy in their pattern recognition and modelling.17–19
The current predictive model demonstrates that initial lack of type-I endoleak in the first seven years (Table 3) despite adherence to instruction for use (IFU) is later influenced by ongoing changes in the infra renal neck length and diameter (Figure 5). This is associated with ongoing morphological and atherosclerotic changes apparent from the weight (influence) of hypercholesterolemia, white cell count, RF and DM. This pattern is different for type-II endoleaks. It appears that the time to event of five years is mostly influenced by other preoperative factors such as left common iliac artery diameter (limb deployment access, main body through the right common iliac artery), haemoglobin and electrolytes (sodium, potassium) (Figure 6). This outcome for occlusion (Figure 7) appears to suggest that body mass index (weight and height) has far more impact on occlusion events.

Predictive model (radar chart) demonstrating the weights of each attribute (n = 26) on the endpoint of occlusion. Increase in each weight is associated with increase in occurrence of each endpoint. This predictive model could be fed with ongoing data and deep learning.
The aim of every centre is to achieve a long-lasting, durable and complication-free EVAR. Despite complying with device-specific IFU in neck diameter, neck length, neck angulation, calcification, thrombus, configuration and precise planning in our unit, we still witness EVAR complications. Furthermore, an increase in EVAR knowledge coupled with advent of new devices has resulted in reduction of EVAR migrations. 20 Nonetheless, other complications such as endoleak or occlusion still persist even after few years of null outcome in surveillance. Despite strict adherence to IFU, increase surgeon and centre experience and direct relation between sac reduction and EVAR success, we still notice sac rupture and occlusion. 21 , 22
AAA is a systemic disease and interplay of attributes ranging from demographics, comorbidities, aneurysm morphology, sero-haematological and inflammatory markers over a long period could still contribute to adverse events despite all strict compliance with IFU. 23 , 24 The traditional statistical analytics through randomised controlled trials, cohort and experimental studies have evaluated the individual and/or collective impacts of these attributes in practice. However, the influence of these factors is either conflicting and, if not, is not applicable across centres and, if so, not replicable. This was also noticed in this study when evaluation of factors by traditional statistical was not dependently or independently contributing to any type of outcome. This is why the role of AI becomes valuable in such circumstances.
The predictive ability (accuracy and precision) of AI owes its performance to computational tools that are inspired by human nervous system (neurons and synapses). They are composed of highly integrated and interconnected networks of computational process referred to as “neurons” or “artificial neural network” (ANN). 25 , 26 ANN has the capability of performing parallel analytics, learn from historical data (supervised or unsupervised deep learning), utilise important links within the dataset (that could not be apparent) and perform non-linear calculations. Further advantage of AI is integrated in its ability to automatically learn and continuously improve on imprecise input (data) by self-regulating its performance to the highest level of precision and detection. 25 , 26 In addition, AI is able to conduct multiple predictive modelling at the same time within the given input and produce the best fit model according to those attributes. Upon detection of the best fit model, such is assessed further by independent steps of validation and testing which avoids traditional statistical pitfalls and interpretations.
Overall, this study highlights that once an EVAR has been conducted with strict adherence to IFU, the interplay of attributes could still influence the outcome over a long-term period. This information in a preoperative planning could serve vital for the process of informed consent, stratification and optimisation of preoperative factors for improved outcomes and reduction in complications. The current model could be refined further if such information is combined with imaging modalities (preoperative and postoperative surveillance) through AI for an individual modelling rather than one fit-for-all model. Another potential of AI is embedded in its ability to be unit or surgeon-based.
Strengths and limitations
This article for the first time compares the long-term outcome of EVAR through traditional statistical analysis to that of AI in the literature. The study benefits from longitudinal dataset; however, a higher number of cases would have increased the robustness of this study. The current predictive model could be subjected to ongoing feed for better accuracy. External validation of the study rather than internal (unsupervised learning) could be the next step with higher number of variables. Another limitation of this study relates to lack of detailed information on level and degree of atherosclerosis, intramural thrombus, lumber arteries and presence or lack of mesenteric (inferior) artery in the initial attributes.
This model is not in any way the perfect predictive model; however, every surgeon/unit could create a predictive infinite simulation on their preoperative planning, demographics, sero-haematological, inflammatory markers, comorbidities, graft type, imaging modality and evaluate how in few years each EVAR with regards to aforementioned endpoints will present. This eventually permits direct stratification prior to deployment. Furthermore, the application of AI could expand to vascular access provision, open and endovascular surgery with creation of a predictive classifier for risk stratification and surveillance coupled with imaging modalities. The current model at its present format is not a replacement for surveillance protocol and follow-up. However, it could be used to highlight those cases that in early years of surveillance demonstrated no endoleak/occlusion but could present with one five to seven years later.
AI in combination to availability of big data (quality and quantity) could simply augment, advance and refine the clinical decision making process in practice. However, this remains dependent on meaningful harvesting and gathering of data which still relies on the human factors and possible errors.
There is no doubt that that there will some degree of scepticism regarding AI and its broader aspect in medicine, but their role in detection, prediction, modelling and decision-making process could not be denied. Overall, a thorough understanding of AI and its components remains crucial for clinicians as unfit data could result in incorrect decision-making. Currently, clinicians are in a position to integrate AI (not replace) to modernise their practice for provision of highest quality care but such requires a strong interdisciplinary collaboration and fostering of education.
Conclusion
This study demonstrates the plausibility, applicability and accuracy of AI in the detection of post EVAR complications that was not detected through traditional statistics. It appears that the interplay of attributes could still impact the outcome of EVAR in long term, and a higher number of attributes could potentiate a stronger model with higher precision for clinical practice.
Supplemental Material
sj-pdf-1-vas-10.1177_1708538120949658 - Supplemental material for Prediction, pattern recognition and modelling of complications post-endovascular infra renal aneurysm repair by artificial intelligence
Supplemental material, sj-pdf-1-vas-10.1177_1708538120949658 for Prediction, pattern recognition and modelling of complications post-endovascular infra renal aneurysm repair by artificial intelligence by Ali Kordzadeh, Mohammad A Hanif, Manfred J Ramirez, Nicholas Railton, Ioannis Prionidis and Thomas Browne in Vascular
Footnotes
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
