Prediction of Sepsis Mortality Risk Based on Ensemble Learning Algorithm FBTV

Sepsis is one of the serious complications of ICU patients and one of the most common causes of death. Research on the ICU sepsis patients’ data in MIMIC-IV database, we propose an ensemble learning algorithm FBTV with voting mechanism to achieve six-hour advance mortality risk prediction. At first, we extract 15,558 sepsis data from MIMIC-IV database and select key features with de-duplication, filtering, padding and outlier processing, etc. After that, 31 key features are extracted by using principal component analysis, and 2,281 data that meet the requirements are used as the dataset for the research. Then, an ensemble learning model are built based on voting mechanism with random Forest, xgBoost and decision Tree as FBTV. Finally, indicators such as accuracy, recall, F1 value, and AUC were used to evaluate the performance of the model. It shows that the FBTV algorithm achieves an accuracy of 0.956 compared with the single algorithm decision Tree which has the highest accuracy 0.927. This work is confirmed to assist clinicians in decision making and adjusting treatment plans to reduce the occurrence of adverse outcomes.