Data Augmentation Using Novel Generative Adversarial Network for Improving Sepsis Mortality Prediction

Sepsis is a major cause of mortality in intensive care units (ICUs), and accurate prediction of patient outcomes is essential for improving clinical decisions and reducing deaths. However, current approaches often fail due to severe class imbalance and a lack of diverse, high-quality data, limiting their generalizability and sensitivity to high-risk cases. To overcome these challenges, we propose a novel augmentation framework, the Hybrid Data Augmentation Method (HDAM), which integrates three generative strategies&mdash standard GAN, conditional GAN1, and conditional GAN2 &mdashin a unified architecture to produce realistic and balanced synthetic samples.  The augmented dataset, generated from the MIMIC-IV database, is used to train six Machine Learning Classifiers (MLCs), including Random Forest (RF) and XGBoost. Among the combinations tested, the HDAM-RF pairing demonstrated superior performance, significantly outperforming traditional augmentation techniques such as SMOTE and single-mode GANs. Notably, HDAM-RF achieved 98.75% accuracy and 0.9981 AUROC, with a substantial improvement in recall and false negative reduction, indicating that HDAM effectively strengthens predictive performance for sepsis mortality and offers promising potential for real-world clinical deployment in ICU settings.