Prediction of Acute Kidney Injury for Critically Ill Cardiogenic Shock Patients with Machine Learning Algorithms

Background: The aim of this study was to use five machine learning approaches and logistic regression to design and validate the acute kidney injury (AKI) prediction model for critically ill individuals with cardiogenic shock (CS).

Methods: All patients who diagnosed with CS from the MIMIC-IV database, the eICU database, and Zhongnan hospital of Wuhan university were included in this study. Clinical information, including demographics, comorbidities, vital signs, critical illness scores and laboratory tests was retrospectively collected. Five machine learning algorithms (LightGBM, decision tree, XGBoost, random forest, and ensemble model) and one conventional logistic regression were applied for the prediction of AKI in critically ill individuals with CS. ROC curves were generated via python software to assess the overall performance of machine learning algorithms and the SHAP analysis was adopted to reveal the impact of prediction for each feature.

Results: The ensemble model exhibited the best predictive ability (AUC:0.91, 95% CI, 0.88-0.94), followed by random forest (AUC:0.90, 95% CI, 0.86-0.94) and XGBoost (AUC:0.89, 95% CI, 0.84-0.92). While the logistic regression model obtained the worst predictive performance (AUC:0.62, 95% CI, 0.56-0.68). When validated the prediction models with eICU database, the ensemble model exhibited the best predictive ability (AUC:0.92, 95% CI, 0.89-0.96), while the logistic model obtained the worst predictive performance (AUC:0.61, 95% CI, 0.56-0.67). Finally, we verified the prediction models using the data from our hospital and ensemble model still exhibited the best predictive ability (AUC:0.74, 95% CI, 0.62-0.86), while the decision tree model obtained the worst predictive performance (AUC:0.52, 95% CI 0.35-0.70).

Conclusion: Machine learning algorithms could be utilized for the AKI prediction among critically ill CS patients, and exhibit superior predictive performance compared to the conventional logistic regression analysis.