Artificial Intelligence-Driven Prediction of Acute Kidney Injury Following Acute Type A Aortic Dissection Surgery in a Chinese Population.

OBJECTIVES

To develop and validate machine learning (ML) models for predicting acute kidney injury (AKI) following acute type A aortic dissection (ATAAD) surgery.

DESIGN

A retrospective single-center study.

SETTING

Beijing Anzhen Hospital (November 2018 to October 2023).

PARTICIPANTS

1350 patients with ATAAD.

INTERVENTIONS: Predictive models have been developed using various ML algorithms to estimate the risk of postoperative AKI.

MEASUREMENTS AND MAIN

RESULTS

Patients were randomly divided into training (85%) and testing (15%) sets. Seven ML algorithms-Gradient Boosting Machine (GBM), LightGBM, Random Forest (RF), K-Nearest Neighbors (KNN), Multi-Layer Perceptron Neural Network (MLP-NN), Naive Bayes (NB), and Logistic Regression (LR)-were evaluated. Model performance was assessed using SHapley Additive exPlanations (SHAP) analysis. A web-based application was developed using the optimal model. Postoperative AKI occurred in 586 patients (43.4%). The constructed models-GBM, LightGBM, RF, KNN, MLP-NN, NB, and LR-achieved areas under the receiver operating characteristic curves of 0.849 (95% CI 0.798-0.902), 0.874 (95% CI 0.831-0.918), 0.800 (95% CI 0.737-0.855), 0.672 (95% CI 0.598-0.739), 0.529 (95% CI 0.486-0.574), 0.833 (95% CI 0.775-0.886), and 0.866 (95% CI 0.810-0.912), respectively, in the testing set. Among these, the LightGBM model surpassed others in predictive accuracy, calibration, and clinical utility.

CONCLUSION

ML models, particularly LightGBM, accurately predict postoperative AKI risk in ATAAD patients, offering a promising tool to enhance perioperative management and patient outcomes.