Development of a Machine Learning Model for Early Prediction of Perioperative Acute Kidney Injury in Patients with Acute Type A Aortic Dissection

P0100 

Abstract Submission 

Yuan Li (1), Shuai Zhang (2), Yi Chang (3), Xiangyang Qian (4)

(1) N/A, N/A, (2) Department of Cardiovascular Surgery, Fuwai Hospital, CAMS&PUMC, Beijing, Beijing, (3) N/A, Beijing, China, (4) Fuwai hospital, Beijing, NA

Yuan Li    -  Contact Me
N/A

Shuai Zhang    -  Contact Me
Department of Cardiovascular Surgery, Fuwai Hospital, CAMS&PUMC

Yi Chang    -  Contact Me
N/A

Xiangyang Qian    -  Contact Me
Fuwai hospital

Yuan Li    -  Contact Me
N/A

Objective: The primary objective of this study was to develop an advanced predictive model for acute kidney injury (AKI) in patients diagnosed with acute type A aortic dissection (ATAAD) by novel machine learning (ML) algorithms to timely intervention and improving prognosis.
Methods: From January 2014 to December 2019, a total of 640 patients diagnosed with ATAAD was enrolled in. The study leveraged the Scikit-learn toolkit and employed one-way analysis of variance (ANOVA) to identify and select pertinent risk factors that exhibited significant associations with the occurrence of AKI. A Synthetic Minority Over-sampling Technique was subsequently employed to address data imbalances. For the construction of the predictive model, a set of ML algorithms, comprising Logistic Regression (LR), XGBoost, and LightGBM, was utilized. The performance of these models was assessed in terms of the area under the curve (AUC) and accuracy (ACC). The Shapley Additive Explanations (SHAP) interpreter was implemented to provide insights into the key risk factors contributing to AKI.
Results: Among the ATAAD patients considered in the study, 74 individuals (11.56%) developed AKI during the post-operative phase of hospitalization. Fifteen highly relevant and statistically significant variables were identified for inclusion in the predictive model. These variables encompassed factors such as the duration of cardiopulmonary bypass (CPB), pre-operative AST levels, the presence of limb syndrome, involvement of the right coronary artery, pre-operative and post-operative creatinine levels, creatine kinase-MB (CK-MB) levels, aortic clamping duration, combined CABG surgery, pre-operative and post-operative WBC counts, lactate dehydrogenase (LDH) levels, the neutrophil-lymphocyte ratio, ALT levels, and neutrophil counts. The predictive models, including LR, XGBoost, and LightGBM, exhibited AUC values of 0.843, 0.879, and 0.887, along with ACC values of 0.801, 0.845, and 0.867, respectively. Notably, the LightGBM model emerged as the most promising model with the highest predictive performance, demonstrating its clinical applicability.
Conclusions: Machine learning models have demonstrated substantial predictive capabilities for identifying postoperative AKI in patients with ATAAD. This study has highlighted a set of variables that can be considered as independent risk factors,offering opportunities for early intervention and improved patient outcomes.

Anesthesia and Perioperative Management