Explainable machine learning models for early prediction of acute kidney injury after cardiac surgery

用于早期预测心脏手术后急性肾损伤的可解释机器学习模型

阅读:1

Abstract

BACKGROUND: This study aimed to develop a data-driven prediction model for cardiac surgery-associated acute kidney injury (CSA-AKI) at the end of surgery using machine learning (ML) algorithms. METHODS: We retrospectively collected clinical data from patients undergoing cardiac surgery at Nanjing First Hospital between June 2016 and January 2023. Feature selection was performed using Lasso regression to identify variables most strongly associated with CSA-AKI. Six ML algorithms-logistic regression (LR), random forest (RF), gradient boosting decision tree (GBDT), eXtreme Gradient Boosting Trees (XGBoost), support vector machine (SVM), and an ensemble model (ENS, integrating GBDT, XGBT, and RF)-were developed and evaluated. External validation was performed on 1,174 patients from the MIMIC-IV database. Model interpretability was assessed via SHAP analysis. RESULTS: In this study of 2,277 cardiac surgery patients, the incidence of CSA-AKI was 22.6%. Patients were divided into training (n = 1,593) and internal validation (n = 684) sets, with comparable AKI rates (22.6% vs. 22.5%). An external validation cohort (n = 1,174) showed a higher AKI incidence (86.9%) and differing baseline characteristics. Feature selection via Lasso regression identified 11 key predictors, with intraoperative red blood cell (RBC) transfusion and cardiopulmonary bypass (CPB) time consistently ranking among the top contributors. Six ML models were developed, with the ensemble model achieving the highest AUC (0.856). External validation confirmed model robustness, with SVM performing best (AUC = 0.847). Confusion matrix analysis demonstrated high accuracy across models, and SHAP analysis reinforced the importance of transfusion and CPB time in AKI prediction. CONCLUSIONS: ML models, particularly the ensemble approach, provide reliable prediction of CSA-AKI, with generalizability validated in an external cohort. These models facilitate early identification of high-risk patients, enabling timely perioperative interventions to improve clinical outcomes.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。