ICU Survival Prediction Incorporating Test-Time Augmentation to Improve the Accuracy of Ensemble-Based Models

This work presents a novel method for applying test-time augmentation (TTA) to tabular data. We used TTA along with an ensemble of 42 models achieve higher performance on the MIT Global Open Source Severity Illness Score dataset consisting 131,051 ICU visits and outcomes. achieved AUC 0.915 private test set (19,669 admissions) won first place at Stanford University's WiDS Datathon 2020 challenge Kaggle, while Acute Physiology Chronic Health Evaluation (APACHE) IV model (commonly survival prediction in literature) 0.868. In addition increasing score, our also reduces “unfair” bias.