Exploring methods of sampling training data for sepsis prediction

This study compares the effectiveness of data oversampling and undersampling methods for the task of early sepsis prediction using two machine learning models. The analysis includes widely used oversampling methods (SMOTE, Borderline SMOTE, ADASYN) and undersampling methods (random undersampling, the Class Centroids methods, and the Near Miss method). Experiments are conducted on the Prediction of Sepsis dataset, which includes various patient health indicators. The results of the experiments showed that, in most cases, the considered sampling methods improve the performance metrics for binary classification evaluation. The highest values of F1-score, precision, and recall were achieved using undersampling methods.