Abdominal Aortic Aneurysm Enlargement, Type II Endoleak
Conditions
Keywords
type II endoleak, abdominal aortic aneurysm, AI, cardiovascular surgery, aorta, radiology, RADAR, EVAR, CT, CTA, T2ELs, sac enlargement, endograft, quantitative features
Brief summary
Background and rationale: Abdominal aortic aneurysm (AAA) is a weakening and enlargement of the main artery in the abdomen. Endovascular aneurysm repair (EVAR) is a minimally invasive treatment used to repair an AAA. After EVAR, some patients develop a complication, called a type II endoleak (T2EL), in which blood continues to flow into the aneurysm sac through small blood vessels. Many T2ELs disappear on their own, but in some patients they can cause the aneurysm sac to enlarge and may require another procedure. The RADAR study aims to develop a deep-learning computer model that can use the CT scan performed before EVAR to predict which patients are more likely to develop a T2EL associated with aneurysm enlargement. The model will be developed using information from several hospitals and tested on patients from hospitals that were not involved in its development. Duration: July 2026 - December 2028 The study will include patients who have undergone EVAR between 1 January 2015 and 31 December 2025. Their available follow-up information will be collected until 31 December 2026 or until an earlier event such as the last available CT scan, a procedure related to T2EL, another defined medical event, or death. The baseline postoperative CTA, acquired 1-3 months after EVAR, will serve as the reference examination for assessment of aneurysm-sac growth during follow-up. Patients who have not developed the study outcome generally need at least 24 months of imaging follow-up to be classified reliably. Objectives: The primary objective is to develop and test a deep-learning model that can predict, before EVAR, whether a patient will develop a T2EL and whether it will be associated with significant enlargement of the aneurysm sac. The model will be tested using data from hospitals that were not involved in its development. Secondary objectives are to assess whether the model works consistently across different hospitals and CT scanning methods; compare the new model with the original NornirNet model; determine whether adding clinical and anatomical information improves the predictions; assess differences in performance between hospitals; and evaluate how well the model's predicted risks correspond to the outcomes actually observed. Study population: The study is a multicenter observational study involving patients over the age of 18 who underwent elective EVAR for an intact fusiform abdominal aortic aneurysm at participating hospitals in Switzerland, Europe, and the United States between 2015 and 2025. Patients must have suitable CT scans before and after EVAR and sufficient medical and imaging information to determine whether a T2EL occurred and how the aneurysm changed over time. Patients treated for a ruptured aneurysm, patients with certain other types of aneurysms, patients who had previous aortic procedures, and patients whose CT scans are not suitable for analysis will be excluded. The planned study population is approximately 1,250 patients, depending on the number of eligible patients available at the participating hospitals. Study procedures: This is a retrospective study, meaning that it uses information and CT scans that were already collected as part of routine medical care. No additional examinations or procedures are performed for the study, and no biological samples are collected. Participating hospitals will provide coded clinical information and anonymized CT scans. The information collected may include age, sex, other medical conditions, body weight and height, laboratory results, heart-related information, details about the aneurysm and blood vessels, and information about the EVAR procedure. Follow-up CT scans will be reviewed by specialists to determine whether a T2EL occurred and whether the aneurysm sac became larger. Patients will be classified into three groups: those with no T2EL, those with a T2EL without significant aneurysm enlargement, and those with a T2EL associated with aneurysm enlargement of at least 5 mm or a related additional procedure. The deep-learning model will be developed using data from some participating hospitals and then tested on data from other hospitals that were not involved in developing the model. Its ability to make accurate and consistent predictions will then be evaluated.
Detailed description
RADAR is a retrospective multicenter observational study designed to develop and independently evaluate a three-dimensional deep-learning model for preoperative prediction of clinically relevant type II endoleak (T2EL) after endovascular aneurysm repair (EVAR). The model will be developed using preoperative CTA data from designated development centers. Model fitting, hyperparameter tuning, and selection of operating thresholds will be performed exclusively within these centers using center-aware cross-validation. After completion of development, the model architecture, parameters, and operating thresholds will be frozen before evaluation on independent held-out test centers. Data from these centers will not be used for model development, tuning, or threshold selection. The primary analysis will assess the ability of the frozen model to distinguish three predefined clinical outcome classes based on the occurrence of T2EL and subsequent aneurysm-sac behavior. This center-level separation between development and testing is intended to provide an estimate of model performance and generalizability across institutions and heterogeneous imaging protocols. Secondary analyses will assess inter-center variability in performance, model calibration, and the incremental predictive value of clinical and anatomical variables. The newly developed model will also be compared with the original NornirNet model on a separate common dataset reserved specifically for this comparison.
Interventions
OTHERArtificial intelligence model analyzing preoperative CT angiography
This study involves the retrospective multicenter analysis of preoperative CT angiography (CTA) and clinical data from patients who underwent EVAR. A new and improved three-dimensional deep-learning model, building on the NornirNet framework, will be developed using data from designated development centers and subsequently evaluated, after model freezing, on independent held-out centers.
Sponsors
Ente Ospedaliero Cantonale, Bellinzona
Study design
Observational model
COHORT
Time perspective
RETROSPECTIVE
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
No
Inclusion criteria
Eligibility is defined by the inclusion and
Exclusion criteria
below. Conditions that depend on the study outcome (occurrence of type II endoleak, competing endoleak types, and duration of imaging follow-up) are deliberately not treated as eligibility criteria; they are specified separately as analysis-set, outcome-classification and censoring rules, in order to avoid outcome-dependent selection bias. The minimum duration of imaging follow-up specified in the section "Outcome classification and follow-up requirement" is a requirement for valid assignment of the reference outcome class and applies only to the event-free classes, for which the absence of the outcome can be established only by a sufficient period of observation. It does not apply to patients in whom the outcome has already been documented. Inclusion criteria * Age ≥ 18 years at the time of the index procedure. * Elective repair of an intact, fusiform, infrarenal abdominal aortic aneurysm (AAA), symptomatic or asymptomatic, with a maximum aneurysm diameter ≥50 mm in women and 55 mm in men, or repaired for symptoms or documented rapid growth (≥5 mm in 6 months or 10 mm in 12 months). * Treatment by standard EVAR, defined as implantation of a commercially available bifurcated infrarenal stent-graft with proximal sealing in the infrarenal neck, below the lowermost renal artery, and without any of the following: * fenestrated, branched or scallop-modified devices (F/BEVAR); * parallel-graft techniques; * adjunctive sac or side-branch embolization at or before the index procedure; * endostapling endoanchor devices at the index procedure; * endovascular aneurysm sealing systems (e.g. EVAS); * proximal or distal extension cuffs placed during or after the index procedure; * aorto-uni-iliac configurations and iliac branch devices; * Index procedure performed between 1 January 2015 and 31 December 2025 at one of the participating centers, with follow-up data censored at the administrative cut-off date of 31 December 2026. * Preoperative arterial-phase CTA available in digital (DICOM) format, acquired ≤ 6 months before the index procedure. * Baseline postoperative CTA acquired between 1 and 3 months after the index procedure. * All qualifying CTA examinations (preoperative, baseline and follow-up) acquired with a reconstructed slice thickness ≤ 2.5 mm, with anatomical coverage extending at minimum from the celiac trunk to the external iliac arteries, and including an arterial phase and a delayed (venous) phase. * Clinical and imaging documentation sufficient to determine the occurrence and clinical course of type II endoleak (primary outcome), including at least one CTA subsequent to the baseline postoperative CTA. Incompleteness of other clinical variables does not preclude inclusion; such variables are recorded as missing and handled according to the prespecified analytical procedures.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Performance of the deep-learning model in held-out test centers | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Performance of the frozen deep-learning model will be evaluated in independent held-out test centers for three-class prediction of T2EL outcome: Class 0 (no T2EL), Class 1 (T2EL without significant aneurysm sac growth), and Class 2 (T2EL with significant sac growth or T2EL-related reintervention). Discrimination and classification performance will be assessed using prespecified performance metrics, including AUC, sensitivity, and specificity. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| AUC comparison of the deep-learning models | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Comparison of the newly developed deep-learning model with the original NornirNet model on the reserved common comparison dataset using the area under the receiver operating characteristic curve (AUC). |
| Sensitivity comparison of the deep-learning models | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Comparison of the newly developed deep-learning model with the original NornirNet model on the reserved common comparison dataset using sensitivity. |
| Specificity comparison of the deep-learning models | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Comparison of the newly developed deep-learning model with the original NornirNet model on the reserved common comparison dataset using specificity. |
| Inter-center variability in AUC | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Assessment of inter-center variability in model performance across held-out test centers using the area under the receiver operating characteristic curve (AUC). |
| Inter-center variability in specificity | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Assessment of inter-center variability in model performance across held-out test centers using specificity. |
| Change in AUC with clinical and anatomical variables | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Assessment of the incremental predictive value of clinical and anatomical variables using change in the area under the receiver operating characteristic curve (AUC). |
| Net Benefit with clinical and anatomical variables | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Assessment of the incremental predictive value of clinical and anatomical variables using Decision Curve Analysis (Net Benefit). |
| Brier score for model calibration | From the baseline postoperative CTA (1-3 months after EVAR) through the last available follow-up, T2EL-related reintervention, censoring event, death, or administrative censoring on 31 December 2026. | Assessment of model calibration using the Brier score. |
Contacts
CONTACTGiorgio Prouse, MD
CONTACTMariacarla Andreozzi, PhD
STUDY_CHAIRGiorgio Prouse, MD
Ente Ospedaliero Cantonale, Bellinzona
Outcome results
None listed