Individualized strategy of treatment for ruptured abdominal aortic aneurysm using causal inference model: a retrospective observational study
10.4174/astr.2026.110.4.259
- Author:
Youngki SOHN
1
;
Youngje WOO
;
Sangkyun MOK
;
Eunju JANG
;
Ki-Yoon MOON
;
Sun Cheol PARK
;
Sang Seob YUN
;
Jang Yong KIM
Author Information
1. Department of Healthcare and Artificial Intelligence, The Catholic University of Korea, Seoul, Korea
- Publication Type:ORIGINAL ARTICLE
- From:Annals of Surgical Treatment and Research
2026;110(4):259-272
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:This study was performed to predict individualized treatment strategies in ruptured abdominal aortic aneurysm (rAAA) by estimating the survival benefit of endovascular aneurysm repair (EVAR) and open surgical repair (OSR) based on anatomical and physiological features using a causal inference model.
Methods:This retrospective study included 45 patients with de novo rAAA who underwent EVAR or OSR between 2012 and 2024. Thirty-three variables were analyzed. The model estimated individualized treatment effects (ITE) for 30-day survival.Model interpretability was assessed using Shapley Additive Explanations (SHAP) analysis. Five-fold cross-validation, receiver operating characteristic (ROC) analysis, and calibration plots were used for model evaluation. A clinical decision tree was developed to derive simplified decision rules.
Results:The mean ITE was 0.22 ± 0.42, with 33% of patients classified as OSR-benefit candidates. SHAP analysis revealed that suprarenal angle, infrarenal angle, iliac anatomy, and proximal neck characteristics strongly influenced treatment effects. However, some predictors, such as low hemoglobin and systolic blood pressure favoring OSR, conflicted with clinical intuition. ROC analysis showed an area under the curve of 1.00, but calibration suggested overfitting due to a small sample size. Treatment-matched patients had a higher 30-day mortality rate than mismatched patients, suggesting potential bias or unmeasured confounding. The decision tree identified clinically relevant features but displayed structural inconsistencies and impractical cutoff values due to the limited sample size.
Conclusion:The X-learner model demonstrated the feasibility of individualized treatment prediction in rAAA but suffered from overfitting and limited generalizability. Validation with larger multicenter cohorts is necessary to confirm clinical applicability.
