Predicting ICU Mortality Based on Generative Adversarial Nets and Ensemble Methods

Abstract

The intensive care unit (ICU) typically admits patients who require urgent medical intervention. Predicting ICU mortality is crucial for identifying those who are at higher risk. Traditional statistical methods, such as logistic regression, have been widely used for ICU survival prediction. However, these methods often have limitations in capturing complex nonlinear relationships between the clinical features. A prediction model based on ensemble learning was proposed for the ICU mortality prediction problem: MTX-stacking model. Firstly, the imbalanced data was processed based on the modified generative adversarial network method. This approach is more explanatory and more effective than traditional data generation methods. Secondly, XGBoost was optimized by tree-structured parzen estimator and stacking structure to prevent overfitting. The proposed MTX-stacking model was evaluated using 131,051 patients from MIT’s GOSSIS initiative. The results indicate that MTX-stacking outperforms the state-of-the-art approaches in terms of area under the receiver operator characteristic (ROC) curve (91.2% and 90.9%). These findings demonstrate the ability and efficiency of the MTX-stacking model to predict ICU mortality.

Keywords

• ICU survival prediction
• ensemble learning
• generative adversarial nets
• Bayesian optimization