Adversarial Ensemble Learning for Mortality Prediction in Intensive Care Units

Predicting patient mortality risk in intensive care units (ICUs) is one of the tasks that has strategic significance in improving clinical decisions and health care outcomes. Disease mortality monitoring methods based on machine learning models have shown efficacy; however, their susceptibility to adversarial attacks in the input data presents reliability and robustness challenges. The present work addresses these challenges by introducing an effective ensemble model enriched with adversarial training to increase the performance of mortality prediction models in the ICU context. The developed methodology combines a variety of ensemble methods, such as random forest, extreme gradient boosting, bagging, AdaBoost, extra trees, and the light gradient boosting machine. These approaches work by combining several algorithms and employing adversarial training strategies that put the stakeholder’s data in their correct order and bar data tampering at all points of the model development ecosystem. The set of experiments performed with the help of real ICU datasets proved that this approach provides better accuracy, robustness, and reliability of predictions than standard models do. The extra trees algorithm achieved the best accuracy among the tested models.