Comparison of Different Machine Learning Methods in Prediction of Long-Term Survival After Endovascular Aneurysm Repair

Abstract

Purpose

Long-term survival after endovascular aneurysm repair (EVAR) for abdominal aortic aneurysm remains a clinical concern, particularly in elderly patients with comorbidities. This study aimed to compare different machine learning (ML) models that capture complex, nonlinear relationships among clinical variables to predict 5-year all-cause mortality following EVAR.

Methods

We retrospectively analyzed 142 patients who underwent elective EVAR between 2013 and 2018. Predictive models for 5-year mortality were developed using 3 supervised ML algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Classification (SVC). Each model was trained on the entire dataset and internally validated through 5-fold cross-validation. Model performance was evaluated using accuracy, sensitivity, specificity, precision, F1 score, and area under the curve (AUC) based on the training set and 5-fold cross-validation. Feature importance was assessed for RF and XGBoost.

Results

The RF demonstrated the most consistent performance (training AUC 0.80; cross-validation AUC 0.77 ± 0.07). XGBoost achieved the highest training accuracy (0.85) but had lower cross-validation AUC (0.68 ± 0.05). SVC showed stable but modest performance. Key predictors identified by RF and XGBoost included poor nutritional status, octogenarian status, compromised immunity, and active cancer.

Conclusions

Tree-based ML models, especially RF, may effectively predict long-term survival after EVAR. Incorporating key clinical predictors into preoperative assessment may enhance risk stratification. Future studies should explore external validation and integration with time-to-event models such as Cox proportional hazards, to enhance prognostic accuracy.