Abstract
Background
Postoperative continuous renal replacement therapy (CRRT) initiation for severe acute kidney injury is a clinically consequential complication after open thoracoabdominal aortic repair. We aimed to develop an interpretable machine-learning (ML) model for predicting postoperative CRRT in this high-risk population.Methods
This single-center retrospective cohort study included consecutive adult patients who underwent open thoracoabdominal aortic repair between January 2010 and December 2025. After exclusions, 372 patients were analyzed, including 66 (17.7%) who initiated postoperative CRRT. The cohort was randomly divided into a training set (75%, n = 278) and an internal test set (25%, n = 94). After multistep feature selection, six routinely available perioperative predictors were retained. Eight ML models and a baseline logistic regression model were developed. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).Results
In bootstrap internal validation, Light Gradient Boosting Machine showed the highest optimism-corrected AUC of 0.814 (95% CI, 0.773-0.856) and the lowest optimism-corrected Brier score of 0.131 (95% CI, 0.112-0.150). However, performance differences among several models were modest. LightGBM was therefore selected for interpretation in this internally validated dataset. SHAP analysis identified surgery duration, D-dimer, serum creatinine, intraoperative red blood cell transfusion, maximum intraoperative lactate, and Crawford extent II as important contributors to predicted CRRT risk.Conclusions
A parsimonious LightGBM model showed favorable internally validated performance for predicting postoperative CRRT after open thoracoabdominal aortic repair. Given the single-center design, limited event number, and lack of external validation, the model should be considered exploratory and requires external validation and potential recalibration before clinical implementation.Clinical trial number
Not applicable.