Abstract

Purpose
To evaluate whether machine learning (ML) models can detect endoleak following endovascular aneurysm repair (EVAR) and quantify their diagnostic accuracy.

Methods
A systematic review and meta-analysis were conducted in accordance with PRISMA guidelines. PubMed, Web of Science, Embase, and Scopus were searched from inception to August 2025 for studies developing or validating ML models for endoleak predict after EVAR. Eligible studies included adult patients undergoing endovascular aortic repair. Data extraction was performed independently by two reviewers, and risk of bias was assessed using the PROBAST tool. Pooled diagnostic performance metrics were estimated using random-effects meta-analysis, including a bivariate Reitsma model to jointly synthesise sensitivity and specificity.

Results
Thirteen studies comprising 4051 patients, including 1641 with endoleaks, were included. Reported ML approaches encompassed logistic regression, random forests, support vector machines, and deep learning models. The pooled area under the curve (AUC) was 0.89 (95% CI 0.83–0.94). Pooled sensitivity was 0.86 (95% CI 0.75–0.92), specificity 0.87 (95% CI 0.83–0.91), and accuracy 0.85 (95% CI 0.80–0.90). Substantial heterogeneity was observed across all pooled estimates (I2 > 85%). All included studies were judged to be at high overall risk of bias, primarily due to retrospective single-centre designs, lack of external validation, and inconsistent handling of missing data.

Conclusion
ML-based models demonstrate high average diagnostic performance for predicting endoleak after EVAR; however, substantial heterogeneity and pervasive risk of bias limit their current clinical applicability. Future research should prioritise large, multicentre prospective validation studies, methodological standardisation, and rigorous assessment of clinical impact before routine implementation.