Application of Resampling Techniques on Sepsis Data to Balance Model Performance with Accuracy, AUC, Kappa and MCC

This research focuses on assessing predictive model performance of sepsis survival using the “Sepsis Survival Minimal Clinical Records” dataset. This research work includes experiments on the “Primary Cohort” and “Study Cohort”, utilizing diverse ML algorithms and resampling techniques. In the dataset, models demonstrated high accuracy but low performance in MCC and Kappa values due to class imbalance. By incorporating resampling techniques, specifically upscaling and hybrid resampling, into our data preprocessing pipeline for both the primary cohort and study cohort, there is a substantial enhancement in the performance metrics, particularly in terms of Kappa and MCC scores while maintaining a nominal level of accuracy. The Kappa boosted from 0 to 0.32 and MCC raised from 0 to 0.33 on upscaling for the primary cohort. Coming to the study cohort Kappa improved from 0 to 0.16 and MCC increased from 0 to 0.17 on upscaling and hybrid resampling.