This study compares the impact of undersampling and oversampling techniques on the performance of machine learning models predicting ICU patient survival using the MIMIC-III dataset. By evaluating Support Vector Machine, Random Forest, and XGBoost models through confusion matrices, classification reports, and ROC-AUC curves, the study finds that undersampling generally yields better balance, higher precision, and improved recall for certain models. The findings highlight the importance of data balancing techniques in mitigating class imbalance and suggest that undersampling offers advantages in computational efficiency and predictive performance. Overall, the study reinforces the critical role of careful data handling in clinical machine learning applications.
