Machine learning of the patient characteristics and of vascular features observed on pre-procedural computed tomography angiographs helps to predict endovascular leaks after thoracic endovascular aneurysm repair

doi:10.21203/rs.3.rs-2725074/v1

Download PDF

Research Article

Machine learning of the patient characteristics and of vascular features observed on pre-procedural computed tomography angiographs helps to predict endovascular leaks after thoracic endovascular aneurysm repair

https://doi.org/10.21203/rs.3.rs-2725074/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 02 May, 2024

Read the published version in Physical and Engineering Sciences in Medicine →

You are reading this latest preprint version

Purpose:

To predict endoleaks after thoracic endovascular aneurysm repair (TEVAR) we submitted patient characteristics and vessel features observed on pre- operative computed tomography angiography (CTA) to machine-learning.

Methods:

We evaluated 1-year follow-up CT scans (arterial and delayed phases) in

patients who underwent TEVAR for the presence or absence of an endoleak. We evaluated the effect of machine learning of the patient age, sex, weight, and height, plus 22 vascular features on the ability to predict post-TEVAR endoleaks. The extreme Gradient Boosting (XGBoost) for ML system was trained on 14 patients with- and 131 without endoleaks. We calculated their importance by applying XGBoost to machine learning and compared our findings between with those of conventional vessel measurement-based methods such as the 22 vascular features by using the Pearson correlation coefficients.

Results:

Pearson correlation coefficient and 95% confidence interval (CI) were r = 0.86 and 0.75 to 0.92 for the machine learning, r = -0.44 and – 0.56 to -0.29 for the vascular angle, and r = -0.19 and -0.34 to -0.02 for the diameter between the subclavian artery and the aneurysm (Fig. 3a-3c, all: p < 0.05). With machine-learning, the univariate analysis was significant higher compared with the vascular angle and in the diameter between the subclavian artery and the aneurysm such as the conventional methods (p < 0.05).

Conclusions:

To predict the risk for post-TEVAR endoleaks, machine learning was superior to the conventional vessel measurement method when factors such as patient characteristics, and vascular features (vessel length, diameter, and angle) were evaluated on pre-TEVAR thoracic CTA images.

thoracic endovascular aneurysm repair

machine learning

computed tomography

computed tomography angiography

aortic aneurysms

endoleaks

Thoracic endovascular aortic repair (TEVAR), used to treat thoracic aortic diseases [1–4], yields satisfactory short- and midterm (up to 5 years) results [5, 6]. However, its long-term efficacy remains to be confirmed and the incidence of post-TEVAR endoleaks and emboli must be reduced.

Endoleaks, i.e. blood flow outside the lumen of the stent-graft but within the aortic aneurysmal sac [7], present a serious problem. As the prognosis of patients with endoleaks is poor, they must be identified and addressed quickly.

Machine-learning algorithms have been applied to clinical data; their performance for predicting the outcome after acute stroke is high [8]. It might be determined whether machine learning is also useful for predicting post-TEVAR endoleaks.

The purpose of this study wes to investigate the ability of machine learning to predict post-TEVAR endoleaks, we subjected the type of stent-graft and patient- and vessel characteristics obtained by pre-operative computed tomography angiography (CTA), to machine learning, we also compared the predictive ability of conventional vessel measurement- and machine-learning data.

This retrospective study was approved by our institutional review board; patient informed consent was waived.

Patients

Between April 2015 and August 2019, 154 consecutive patients with suspected thoracic aortic aneurysms (TAA) underwent both contrast-enhanced CTA and TEVAR. Contrast-enhanced CTA combined with delayed-phase imaging performed one year later identified 14 patients with- (group 1) and 131 without endoleaks (group 2). Excluded from further study were 9 patients (renal dysfunction, n = 6; severe calcification, n = 3). As shown in Table 1, there was no significant inter-group difference in the age, sex, and body habitus.

Table 1

Patient characteristics
	Endoleak group	Non-endoleak group	p-value
Age (year)	79 (44–90)	76 (43–93)	0.12
Sex (M / F)	11 / 3	104 / 27	0.27
Height (cm)	162.8 (140.0–174.3)	163.7 (130.0–185.0)	0.39
Weight (kg)	51.8 (35.5–73.2)	58.5 (37.7–97.2)	0.1

Thoracic CTA- and contrast medium injection protocol

All patients were scanned on a 64-detector-row CT scanner (Lightspeed VCT; GE Healthcare, Milwaukee, WI). Thoracic CTA was from the top of the lung apex to the lower end of the diaphragm in the craniocaudal direction. The scan parameters were 0.5-sec rotation scan, 1.25-mm detector row width, 0.984 helical pitch (beam pitch), 78.75-mm table movement, and 50-cm scan field of view. The tube voltage was 100 kVp; the tube current was changed from 200 to 770 mA to maintain the image quality (noise index = 9) using automatic tube current modulation. The scan length along the z axis ranged from 30 to 40 cm depending on the patient body size. The scanning time varied from 7.0 to 13.0 sec. The contrast material (Omnipaque-300; Daiichi-Sankyo, Tokyo, Japan) was injected through a 20-gauge catheter into the antecubital vein with a power injector (Dual Shot; Nemoto-Kyorindo, Tokyo, Japan). We used smart prep to obtain the scan delay for thoracic CTA. The injection volume was the total body weight in kg x 1.8 ml, administered during 30 sec.

TEVAR

All interventions were performed under aseptic conditions. The same angiography system (Alphenix Hybrid, Canon, Japan) was used by a multidisciplinary team that included interventional radiologists and vascular surgeons, each with at least 5 years of experience in the endovascular treatment of TAAs. Based on the vascular morphology, the stent-graft s were TAG (W. L. Gore & Associates, Flagstaff, AZ, USA), Zenith TX2 (William Cook Europe, ApS, Bjaeverskov, Denmark), RELAY PLUS (ABS Bolton Medical, Barcelona, Spain), VALIANT (Medtronic, Santa Rosa, CA, USA), and ZenithAlpha (Cook Medical, Bloomington, IN, USA).

The contrast material was Iopamidol 370 (BRACCO Imaging, Constance, Germany). The stent-graft was deployed in the descending aorta via the right femoral artery. The condition of the stent-graft was checked on angiograms; the entry tear exclusion was confirmed, as was the absence of endoleaks and of backflow from the distal re-entry site.

Machine-learning

One year post-TEVAR we compared the 145 patients with- (group 1, n = 14) and without endoleaks (group 2, n = 131) with respect to their age, sex, height, and weight, and the 22 vessel measurements shown in Table 2. We used data in their electronic medical records (Fujitsu, Japan) and in the Picture Archiving and Communications System (PACS). Vascular data were obtained on thoracic CTA images.

Table 2

Pre-procedure vessel measurements obtained in 14 patients with- and 131 patients without post-TEVAR endoleaks
	Endoleak group	Non-endoleak group	p-value
Angle (º)	19.0 (15.0–25.0)	37.0 (7.0–47.0)	< 0.01
Distance between valsalva and brachiocephalic artery	84.0 (68.0–113.0)	84.0 (53.0–135.0)	0.76
Short axis of brachiocephalic artery	33.0 (23.0–41.0)	32.0 (24.0–41.0)	0.83
Short axis of vertebral artery	31.5 (23.0–42.0)	31.0 (24.0–46.0)	0.76
Short axis of left subclavian artery	31.5 (22.0–40.0)	31.0 (24.0–45.0)	0.58
Diameter of subclavian artery	11.0 (9.0–14.0)	11.0 (8.0–17.0)	0.84
Proximal short axis of landing zone 1	3.0 ( 23.0–54.0)	30.0 (18.0–42.0)	0.22
Proximal short axis of landing zone 2	3.0 ( 23.0–54.0)	30.0 (17.0–41.0)	0.21
Proximal short axis of landing zone 3	3.0 ( 23.0–54.0)	30.0 (18.0–42.0)	0.23
Proximal long axis of aneurysm	30.5 (24.5–55.0)	31.0 (19.0–42.0)	0.73
Proximal short axis of aneurysm	30.5 (22.4–54.0)	30.0 (18.0–41.0)	0.44
Long axis of inner diameter of aneurysm	47.0 (27.0–61.0)	42.0 (16.0–75.0)	0.65
Short axis of inner diameter of aneurysm	39.5 (22.0–57.0)	36.0 (11.0–65.0)	0.12
Long axis of outer diameter of aneurysm	49.0 ( 35.0–88.0)	48.0 (31.0–95.0)	0.44
Short axis of outer diameter of aneurysm	42.0 (34.7–75.0)	40.0 (29.0–94.0)	0.24
Distal long axis of aneurysm	30.0 ( 22.0–52.0)	29.0 (21.0–44.0)	0.69
Distal short axis of aneurysm	29.5 ( 21.0–50.0)	29.0 (21.0–43.0)	0.87
Distal short axis of landing zone 1	28.5 (24.0–35.0)	29.0 (22.0–37.0)	0.6
Distal short axis of landing zone 2	29.0 (23.0–35.0)	29.0 (22.0–37.0)	0.87
Distal short axis of landing zone 3	28.5 (24.0–35.0)	29.0 (22.0–37.0)	0.71
Diameter between subclavian artery and aneurysm	17.0 ( 10.0–54.0)	32.0 (9.2–189.0)	< 0.01
Long axis of aneurysm	115.0 (35.0–220.0)	121.0 (35.0–241.0)	0.65
TEVAR: Thoracic endovascular aortic repair
Unless otherwise indicated, measurements are the mean and shown in mm

We created machine-learning classifiers that applied extreme Gradient Boosting (XGBoost) (version 1.6.2; https://pypi.org/project/xgboost/) using Python (version 3.5; https://www.python.org/) and scikit-learn (version 0.18.1, http://scikit-learn.org/stable/) to discriminate between lumens with/without post-EVAR endoleaks. Scikit-learn is a python library and XGBoost is a library that can work with the same python. We divided the145 thoracic CTA patients into training data and test data in an 70/30 ratio so that the ratio of endoleak, without endoleak was as comparable as possible. As a result, we obtained 102 cases of training data (10 endoleak and 91 without endoleak) and 43 cases of test data (4 endoleak and 40 without endoleak). For parameter tuning we used the K-fold cross-validation function (K = 10), and applied oversampling for imbalanced data (synthetic minority over-sampling, SMOTE) [9], and Bayesian optimization (Hyperopt) [10]. The parameters for Bayesian optimization were: {'x_colsample_bytree': 0.8510707298078505, 'x_gamma': 0.45545493021965733, 'x_max_depth': 5.0, 'x_min_child': 2.0, 'x_reg_lambda': 0.6139857890064586, 'x_subsample': 0.9242267670478332}. Cross-validation was applied twice: the first time between training and validation data for tuning hyperparameters; the second time between training and test data after obtaining tuning data for model evaluation.

Machine-learning classifiers

For the post-TEVAR endoleak and non-endoleak classification we recorded the implanted stent-graft device (TAG, Zenith TX2, RELAY PLUS, VALIANT, ZenithAlpha), the patient age, sex, height, and weight (Table 1), and the 22 vessel standard set of measurements obtained on pre-operative CTA images (Fig. 1, Table 2). As the classification endoleak/non-endoleak with XGBoost using multiple items was unsatisfactory, we applied SMOTE for correcting imbalance data. Simple normalization was performed by first calculating the mean and standard deviation (SD) for each item. We then subtracted the mean from each item and divided the result by the SD. We trained the machine-learning classifiers on endoleaks and non-endoleaks by applying K-fold cross-validation to separate training sessions performed during testing of the discriminant using the classifiers (K = 10). The cross-validation function of scikit-learn automatically separated the population into K-fold cross-validation (K = 10). The parameters of XGBoost were selected by Bayesian optimization; to evaluate the importance of each item we applied the feature-importance algorithm.

Statistical analysis

Statistical analyses were performed with free “R” (R, version 3.2.2; The R Project for Statistical Computing; http://www.r-project.org/). Vessel measurements were performed between with the endoleak and without the endoleak group to identify the factors that affect for the endoleak. We compared the endoleak/non-endoleak groups with respect to 26 items using the Mann-Whitney U-test. The importance of individual features in the development of endoleaks was determined using XGBoost. We calculated their importance by applying XGBoost to machine learning and compared our findings between with those of conventional vessel measurement-based methods such as the 22 vascular features by using the Pearson correlation coefficients. To determine the strength of association, we took ‘r’ as the absolute value of the normal data distribution; r = 0–0.19 was recorded as very weak-, 0.2–0.39 as weak-, 0.40–0.59 as moderate, 0.6–0.79 as strong-, and 0.8–1.0 as very strong correlation. We compared the correlation coefficients using the method of Meng et al. [11].

Comparison of the age, sex, height, and weight, and of 22 vessel measurements performed on pre-TEVAR CTA images of 14 patients with (group 1) and 131 without post-TEVAR endoleaks (group 2) (26 x 145 = 3,770 total items) revealed no significant difference in the patient characteristics (Table 1). The only significant inter-group difference was observed in the vascular angle and in the diameter between the subclavian artery and the aneurysm (p < 0.01, Table 2).

Although XGBoost showed that the vessel angle (Fig. 2) played an important role in the post-TEVAR development of endoleaks.

Pearson correlation coefficient and 95% confidence interval (CI) were r = 0.86 and 0.75 to 0.92 for the machine learning, r = -0.44 and – 0.56 to -0.29 for the vascular angle, and r = -0.19 and − 0.34 to -0.02 for the diameter between the subclavian artery and the aneurysm (Fig. 3a-3c, all: p < 0.05). With machine-learning, the univariate analysis was significant higher compared with the vascular angle and in the diameter between the subclavian artery and the aneurysm such as the conventional methods (p < 0.05). The train- and eval for test logloss curves are presented in Fig. 4.

Our findings suggest that machine learning using the patient age, sex, and body habitus and specific vascular features observed on pre-TEVAR thoracic CT angiographs helps to predict post-procedural endoleaks.

XGBoost showed that the vessel angle played an important role in the development of endoleaks. As the long segments of current stent-grafts are relatively stiff, they cannot conform to highly angulated or curved arches. Therefore, the angulation and curvature at the aortic arch may present a risk factor for technical failure. The lack of apposition between the device and the aortic wall in vessels with a slight curve results in the bird-beak configuration that is radiologically recognized as a wedge-shaped gap between the undersurface of the stent-graft and the aortic wall [12–14]. Consequently, the vessel angle is an important factor for predicting post-TEVAR endoleaks and its pre-operative submission to machine learning is of great predictive value.

There was a significant inter-method difference with respect to the measured diameter between the subclavian artery and the aneurysm. To avoid stent-graft migration, collapse, or endoleak formation, TEVAR requires adequate stability at the proximal and distal landing zones [2, 15–17]. Flat, straight, long, and cylindrical landing zones are ideal for the stable deployment of stent grafts [16, 18]; landing zones near or proximal to the origin of the left subclavian artery are associated with a high risk for endoleak formation [19–21]. As machine learning incorporates patient- and vessel characteristics, it is better at predicting post-procedural endoleaks than conventional vessel measurement methods.

To avoid or address post-procedural endoleaks in patients scheduled for TEVAR, the aortic morphology must be assessed on pre-procedural images [22]. Our research allows us to advertise more about clinical applications. We therefore can present a method that may predict response to a therapy before the therapy is performed. This would be a useful tool in clinical decision on whether or not to provide a therapy. We found that the application of machine learning was highly useful for predicting pre- and post-TEVAR endoleaks.

Previous research has revealed that ML facilitates the detection of endoleaks subsequent to endovascular aortic repair [23–27]. Nonetheless, limited reports are available on the prognostication of thoracic endoleaks. Dong et al. reported on the development of models for prognosticating the requirement for reintervention following TEVAR for Stanford type B aortic dissection [28]. To the best of our knowledge, no studies have investigated the use of ML in identifying neoaneurysms in TAVER. Our investigation demonstrated the significant value of pre-TEVAR thoracic CTA documentation and machine learning in anticipating endoleaks subsequent to stent-graft implantation. Based on our findings, we propose that patient characteristics, pre-operative thoracic CTA imaging that delineates vascular features, and the type of stent-graft employed be analyzed using machine learning to enhance the ability to predict the risk for post-TEVAR endoleaks.

Some study limitations must be addressed. Ours was a small, single-center-, single CTA-protocol study and only the performance of the machine-learning algorithm was evaluated. The clinical implications of our findings remain unclear and require the integration of machine-learning models into image analysis software; also, sub-analysis-based studies are needed. We did not discriminate among endoleak types nor did we evaluate pressure differences between internal and external aneurysms.

In conclusion, to predict the risk for post-TEVAR endoleaks, machine learning was superior to the conventional vessel measurement method when factors such as patient characteristics, and vascular features (vessel length, diameter, and angle) were evaluated on pre-TEVAR thoracic CTA images.

Thoracic endovascular aneurysm repair (TEVAR)

extreme Gradient Boosting (XGBoost)

computed tomography angiography (CTA)

under the curve (AUC)

confidence interval (CI)

thoracic aortic aneurysms (TAA)

synthetic minority over-sampling (SMOTE)

standard deviation (SD)

receiver operating characteristic (ROC)

Declarations : “Not applicable”

Funding: No funding

Ethics approval: This retrospective study was approved by our institutional review board; informed

consent was waived.

Approval was granted by the institutional review board of Tsuchiya General Hospital. The requirement for individual informed consent was waived due to the retrospective nature of the study.

Chu MW, Forbes TL, Kirk Lawlor D, Harris KA, Derose G. Endovascular repair of thoracic aortic disease: early and midterm experience. Vascular and endovascular surgery 2007; 41:186-191
Svensson LG, Kouchoukos NT, Miller DC, et al. Expert consensus document on the treatment of descending thoracic aortic disease using endovascular stent-grafts. The Annals of thoracic surgery 2008; 85:S1-41
Nienaber CA, Fattori R, Lund G, et al. Nonsurgical reconstruction of thoracic aortic dissection by stent-graft placement. The New England journal of medicine 1999; 340:1539-1545
Greenberg RK, O'Neill S, Walker E, et al. Endovascular repair of thoracic aortic lesions with the Zenith TX1 and TX2 thoracic grafts: intermediate-term results. Journal of vascular surgery 2005; 41:589-596
Demers P, Miller DC, Mitchell RS, et al. Midterm results of endovascular repair of descending thoracic aortic aneurysms with first-generation stent grafts. The Journal of thoracic and cardiovascular surgery 2004; 127:664-673
Verhoye JP, Miller DC, Sze D, Dake MD, Mitchell RS. Complicated acute type B aortic dissection: midterm results of emergency endovascular stent-grafting. The Journal of thoracic and cardiovascular surgery 2008; 136:424-430
Stavropoulos SW, Carpenter JP. Postoperative imaging surveillance and endoleak management after endovascular repair of thoracic aortic aneurysms. Journal of vascular surgery 2006; 43 Suppl A:89a-93a
Heo J, Yoon JG, Park H, Kim YD, Nam HS, Heo JH. Machine Learning-Based Model for Prediction of Outcomes in Acute Stroke. Stroke 2019; 50:1263-1265
Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: Synthetic Minority Over-sampling Technique. J Artificial Intelligence Res 2002; 16: 321-357.
Agarwal R. How to use Hyperopt? https://www.kaggle.com/mlwhiz/how-to-use-hyperopt.
Meng X-L, Rosenthal R, Rubin DB. Comparing correlated correlation coefficients. Psychol Bull 1992; 111:172-175.
Fayad A. Thoracic endovascular stent graft with a bird's beak sign. Canadian journal of anaesthesia = Journal canadien d'anesthesie 2008; 55:785-786
Kashyap VS, Pavkov ML, Bishop PD, et al. Angiography underestimates peripheral atherosclerosis: lumenography revisited. Journal of endovascular therapy : an official journal of the International Society of Endovascular Specialists 2008; 15:117-125
Canaud L, Alric P, Laurent M, et al. Proximal fixation of thoracic stent-grafts as a function of oversizing and increasing aortic arch angulation in human cadaveric aortas. Journal of endovascular therapy : an official journal of the International Society of Endovascular Specialists 2008; 15:326-334
Garzón G, Fernández-Velilla M, Martí M, Acitores I, Ybáñez F, Riera L. Endovascular stent-graft treatment of thoracic aortic disease. Radiographics : a review publication of the Radiological Society of North America, Inc 2005; 25 Suppl 1:S229-244
Ueda T, Fleischmann D, Rubin GD, Dake MD, Sze DY. Imaging of the thoracic aorta before and after stent-graft repair of aneurysms and dissections. Seminars in thoracic and cardiovascular surgery 2008; 20:348-357
Sze DY, Mitchell RS, Miller DC, et al. Infolding and collapse of thoracic endoprostheses: manifestations and treatment options. The Journal of thoracic and cardiovascular surgery 2009; 138:324-333
Sze DY, van den Bosch MA, Dake MD, et al. Factors portending endoleak formation after thoracic aortic stent-graft repair of complicated aortic dissection. Circulation Cardiovascular interventions 2009; 2:105-112
Tse LW, MacKenzie KS, Montreuil B, Obrand DI, Steinmetz OK. The proximal landing zone in endovascular repair of the thoracic aorta. Annals of vascular surgery 2004; 18:178-185
Ishimaru S. Endografting of the aortic arch. Journal of endovascular therapy : an official journal of the International Society of Endovascular Specialists 2004; 11 Suppl 2:Ii62-71
Melissano G, Civilini E, Bertoglio L, et al. Results of endografting of the aortic arch in different landing zones. European journal of vascular and endovascular surgery : the official journal of the European Society for Vascular Surgery 2007; 33:561-566
van Keulen JW, van Prehn J, Prokop M, Moll FL, van Herwaarden JA. Dynamics of the aorta before and after endovascular aneurysm repair: a systematic review. European journal of vascular and endovascular surgery : the official journal of the European Society for Vascular Surgery 2009; 38:586-596
Raffort J, Adam C, Carrier M, et al. Artificial intelligence in abdominal aortic aneurysm. Journal of vascular surgery 2020; 72:321-333.e321
Karthikesalingam A, Holt PJ, Vidal-Diez A, et al. Predicting aortic complications after endovascular aneurysm repair. Br J Surg 2013; 100:1302-1311
Barnes M, Boult M, Maddern G, Fitridge R. A model to predict outcomes for endovascular aneurysm repair using preoperative variables. European journal of vascular and endovascular surgery : the official journal of the European Society for Vascular Surgery 2008; 35:571-579
Talebi S, Madani MH, Madani A, et al. Machine learning for endoleak detection after endovascular aortic repair. Sci Rep 2020; 10:18343
Masuda T, Baba Y, Nakaura T, et al. Applying patient characteristics, stent-graft selection, and pre-operative computed tomographic angiography data to a machine learning algorithm: Is endoleak prediction possible? Radiography (Lond) 2022; 28:906-911
Dong Y, Que L, Jia Q, et al. Predicting reintervention after thoracic endovascular aortic repair of Stanford type B aortic dissection using machine learning. Eur Radiol 2022; 32:355-367

Download PDF

Journal Publication

published 02 May, 2024

Read the published version in Physical and Engineering Sciences in Medicine →

Editorial decision: Minor revisions
27 Oct, 2023
Reviewers agreed at journal
09 Jul, 2023
Reviewers invited by journal
07 Jul, 2023
Editor invited by journal
26 Mar, 2023
Editor assigned by journal
24 Mar, 2023
First submitted to journal
22 Mar, 2023

You are reading this latest preprint version

Machine learning of the patient characteristics and of vascular features observed on pre-procedural computed tomography angiographs helps to predict endovascular leaks after thoracic endovascular aneurysm repair

Status:

Journal Publication

Version 1

Abstract

Figures

INTRODUCTION

MATERIALS AND METHODS

TEVAR

Statistical analysis

RESULTS

DISCUSSION

Abbreviations

Declarations

References

Status:

Journal Publication

Version 1